2014-06-05 12:01:32 +08:00
|
|
|
/*
|
|
|
|
* Copyright © 2014 Red Hat
|
|
|
|
*
|
|
|
|
* Permission to use, copy, modify, distribute, and sell this software and its
|
|
|
|
* documentation for any purpose is hereby granted without fee, provided that
|
|
|
|
* the above copyright notice appear in all copies and that both that copyright
|
|
|
|
* notice and this permission notice appear in supporting documentation, and
|
|
|
|
* that the name of the copyright holders not be used in advertising or
|
|
|
|
* publicity pertaining to distribution of the software without specific,
|
|
|
|
* written prior permission. The copyright holders make no representations
|
|
|
|
* about the suitability of this software for any purpose. It is provided "as
|
|
|
|
* is" without express or implied warranty.
|
|
|
|
*
|
|
|
|
* THE COPYRIGHT HOLDERS DISCLAIM ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
|
|
|
|
* INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS, IN NO
|
|
|
|
* EVENT SHALL THE COPYRIGHT HOLDERS BE LIABLE FOR ANY SPECIAL, INDIRECT OR
|
|
|
|
* CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF USE,
|
|
|
|
* DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR OTHER
|
|
|
|
* TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR PERFORMANCE
|
|
|
|
* OF THIS SOFTWARE.
|
|
|
|
*/
|
|
|
|
|
|
|
|
#include <linux/delay.h>
|
|
|
|
#include <linux/errno.h>
|
2019-05-06 17:52:47 +08:00
|
|
|
#include <linux/i2c.h>
|
|
|
|
#include <linux/init.h>
|
|
|
|
#include <linux/kernel.h>
|
2014-06-05 12:01:32 +08:00
|
|
|
#include <linux/sched.h>
|
2014-07-09 09:15:09 +08:00
|
|
|
#include <linux/seq_file.h>
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
#if IS_ENABLED(CONFIG_DRM_DEBUG_DP_MST_TOPOLOGY_REFS)
|
2019-10-31 17:43:49 +08:00
|
|
|
#include <linux/stacktrace.h>
|
2019-06-21 05:59:25 +08:00
|
|
|
#include <linux/sort.h>
|
|
|
|
#include <linux/timekeeping.h>
|
|
|
|
#include <linux/math64.h>
|
|
|
|
#endif
|
|
|
|
|
2017-07-12 23:51:02 +08:00
|
|
|
#include <drm/drm_atomic.h>
|
|
|
|
#include <drm/drm_atomic_helper.h>
|
2019-05-06 17:52:47 +08:00
|
|
|
#include <drm/drm_dp_mst_helper.h>
|
|
|
|
#include <drm/drm_drv.h>
|
|
|
|
#include <drm/drm_print.h>
|
2019-01-18 05:03:34 +08:00
|
|
|
#include <drm/drm_probe_helper.h>
|
2014-06-05 12:01:32 +08:00
|
|
|
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
#include "drm_crtc_helper_internal.h"
|
2019-09-04 04:45:45 +08:00
|
|
|
#include "drm_dp_mst_topology_internal.h"
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* DOC: dp mst helper
|
|
|
|
*
|
|
|
|
* These functions contain parts of the DisplayPort 1.2a MultiStream Transport
|
|
|
|
* protocol. The helpers contain a topology manager and bandwidth manager.
|
|
|
|
* The helpers encapsulate the sending and received of sideband msgs.
|
|
|
|
*/
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
struct drm_dp_pending_up_req {
|
|
|
|
struct drm_dp_sideband_msg_hdr hdr;
|
|
|
|
struct drm_dp_sideband_msg_req_body msg;
|
|
|
|
struct list_head next;
|
|
|
|
};
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static bool dump_dp_payload_table(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
char *buf);
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
static void drm_dp_mst_topology_put_port(struct drm_dp_mst_port *port);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
static int drm_dp_dpcd_write_payload(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
int id,
|
|
|
|
struct drm_dp_payload *payload);
|
|
|
|
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
static int drm_dp_send_dpcd_read(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
|
|
|
int offset, int size, u8 *bytes);
|
2014-06-05 12:01:32 +08:00
|
|
|
static int drm_dp_send_dpcd_write(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
|
|
|
int offset, int size, u8 *bytes);
|
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
static int drm_dp_send_link_address(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb);
|
2014-06-05 12:01:32 +08:00
|
|
|
static int drm_dp_send_enum_path_resources(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_mst_port *port);
|
|
|
|
static bool drm_dp_validate_guid(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
u8 *guid);
|
|
|
|
|
|
|
|
static int drm_dp_mst_register_i2c_bus(struct drm_dp_aux *aux);
|
|
|
|
static void drm_dp_mst_unregister_i2c_bus(struct drm_dp_aux *aux);
|
|
|
|
static void drm_dp_mst_kick_tx(struct drm_dp_mst_topology_mgr *mgr);
|
2019-01-23 04:03:01 +08:00
|
|
|
|
2019-09-04 04:45:45 +08:00
|
|
|
#define DBG_PREFIX "[dp_mst]"
|
|
|
|
|
2019-01-23 04:03:01 +08:00
|
|
|
#define DP_STR(x) [DP_ ## x] = #x
|
|
|
|
|
|
|
|
static const char *drm_dp_mst_req_type_str(u8 req_type)
|
|
|
|
{
|
|
|
|
static const char * const req_type_str[] = {
|
|
|
|
DP_STR(GET_MSG_TRANSACTION_VERSION),
|
|
|
|
DP_STR(LINK_ADDRESS),
|
|
|
|
DP_STR(CONNECTION_STATUS_NOTIFY),
|
|
|
|
DP_STR(ENUM_PATH_RESOURCES),
|
|
|
|
DP_STR(ALLOCATE_PAYLOAD),
|
|
|
|
DP_STR(QUERY_PAYLOAD),
|
|
|
|
DP_STR(RESOURCE_STATUS_NOTIFY),
|
|
|
|
DP_STR(CLEAR_PAYLOAD_ID_TABLE),
|
|
|
|
DP_STR(REMOTE_DPCD_READ),
|
|
|
|
DP_STR(REMOTE_DPCD_WRITE),
|
|
|
|
DP_STR(REMOTE_I2C_READ),
|
|
|
|
DP_STR(REMOTE_I2C_WRITE),
|
|
|
|
DP_STR(POWER_UP_PHY),
|
|
|
|
DP_STR(POWER_DOWN_PHY),
|
|
|
|
DP_STR(SINK_EVENT_NOTIFY),
|
|
|
|
DP_STR(QUERY_STREAM_ENC_STATUS),
|
|
|
|
};
|
|
|
|
|
|
|
|
if (req_type >= ARRAY_SIZE(req_type_str) ||
|
|
|
|
!req_type_str[req_type])
|
|
|
|
return "unknown";
|
|
|
|
|
|
|
|
return req_type_str[req_type];
|
|
|
|
}
|
|
|
|
|
|
|
|
#undef DP_STR
|
|
|
|
#define DP_STR(x) [DP_NAK_ ## x] = #x
|
|
|
|
|
|
|
|
static const char *drm_dp_mst_nak_reason_str(u8 nak_reason)
|
|
|
|
{
|
|
|
|
static const char * const nak_reason_str[] = {
|
|
|
|
DP_STR(WRITE_FAILURE),
|
|
|
|
DP_STR(INVALID_READ),
|
|
|
|
DP_STR(CRC_FAILURE),
|
|
|
|
DP_STR(BAD_PARAM),
|
|
|
|
DP_STR(DEFER),
|
|
|
|
DP_STR(LINK_FAILURE),
|
|
|
|
DP_STR(NO_RESOURCES),
|
|
|
|
DP_STR(DPCD_FAIL),
|
|
|
|
DP_STR(I2C_NAK),
|
|
|
|
DP_STR(ALLOCATE_FAIL),
|
|
|
|
};
|
|
|
|
|
|
|
|
if (nak_reason >= ARRAY_SIZE(nak_reason_str) ||
|
|
|
|
!nak_reason_str[nak_reason])
|
|
|
|
return "unknown";
|
|
|
|
|
|
|
|
return nak_reason_str[nak_reason];
|
|
|
|
}
|
|
|
|
|
|
|
|
#undef DP_STR
|
2019-09-04 04:45:45 +08:00
|
|
|
#define DP_STR(x) [DRM_DP_SIDEBAND_TX_ ## x] = #x
|
|
|
|
|
|
|
|
static const char *drm_dp_mst_sideband_tx_state_str(int state)
|
|
|
|
{
|
|
|
|
static const char * const sideband_reason_str[] = {
|
|
|
|
DP_STR(QUEUED),
|
|
|
|
DP_STR(START_SEND),
|
|
|
|
DP_STR(SENT),
|
|
|
|
DP_STR(RX),
|
|
|
|
DP_STR(TIMEOUT),
|
|
|
|
};
|
|
|
|
|
|
|
|
if (state >= ARRAY_SIZE(sideband_reason_str) ||
|
|
|
|
!sideband_reason_str[state])
|
|
|
|
return "unknown";
|
|
|
|
|
|
|
|
return sideband_reason_str[state];
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
|
|
|
drm_dp_mst_rad_to_str(const u8 rad[8], u8 lct, char *out, size_t len)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
u8 unpacked_rad[16];
|
|
|
|
|
|
|
|
for (i = 0; i < lct; i++) {
|
|
|
|
if (i % 2)
|
|
|
|
unpacked_rad[i] = rad[i / 2] >> 4;
|
|
|
|
else
|
|
|
|
unpacked_rad[i] = rad[i / 2] & BIT_MASK(4);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* TODO: Eventually add something to printk so we can format the rad
|
|
|
|
* like this: 1.2.3
|
|
|
|
*/
|
|
|
|
return snprintf(out, len, "%*phC", lct, unpacked_rad);
|
|
|
|
}
|
2019-01-23 04:03:01 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* sideband msg handling */
|
|
|
|
static u8 drm_dp_msg_header_crc4(const uint8_t *data, size_t num_nibbles)
|
|
|
|
{
|
|
|
|
u8 bitmask = 0x80;
|
|
|
|
u8 bitshift = 7;
|
|
|
|
u8 array_index = 0;
|
|
|
|
int number_of_bits = num_nibbles * 4;
|
|
|
|
u8 remainder = 0;
|
|
|
|
|
|
|
|
while (number_of_bits != 0) {
|
|
|
|
number_of_bits--;
|
|
|
|
remainder <<= 1;
|
|
|
|
remainder |= (data[array_index] & bitmask) >> bitshift;
|
|
|
|
bitmask >>= 1;
|
|
|
|
bitshift--;
|
|
|
|
if (bitmask == 0) {
|
|
|
|
bitmask = 0x80;
|
|
|
|
bitshift = 7;
|
|
|
|
array_index++;
|
|
|
|
}
|
|
|
|
if ((remainder & 0x10) == 0x10)
|
|
|
|
remainder ^= 0x13;
|
|
|
|
}
|
|
|
|
|
|
|
|
number_of_bits = 4;
|
|
|
|
while (number_of_bits != 0) {
|
|
|
|
number_of_bits--;
|
|
|
|
remainder <<= 1;
|
|
|
|
if ((remainder & 0x10) != 0)
|
|
|
|
remainder ^= 0x13;
|
|
|
|
}
|
|
|
|
|
|
|
|
return remainder;
|
|
|
|
}
|
|
|
|
|
|
|
|
static u8 drm_dp_msg_data_crc4(const uint8_t *data, u8 number_of_bytes)
|
|
|
|
{
|
|
|
|
u8 bitmask = 0x80;
|
|
|
|
u8 bitshift = 7;
|
|
|
|
u8 array_index = 0;
|
|
|
|
int number_of_bits = number_of_bytes * 8;
|
|
|
|
u16 remainder = 0;
|
|
|
|
|
|
|
|
while (number_of_bits != 0) {
|
|
|
|
number_of_bits--;
|
|
|
|
remainder <<= 1;
|
|
|
|
remainder |= (data[array_index] & bitmask) >> bitshift;
|
|
|
|
bitmask >>= 1;
|
|
|
|
bitshift--;
|
|
|
|
if (bitmask == 0) {
|
|
|
|
bitmask = 0x80;
|
|
|
|
bitshift = 7;
|
|
|
|
array_index++;
|
|
|
|
}
|
|
|
|
if ((remainder & 0x100) == 0x100)
|
|
|
|
remainder ^= 0xd5;
|
|
|
|
}
|
|
|
|
|
|
|
|
number_of_bits = 8;
|
|
|
|
while (number_of_bits != 0) {
|
|
|
|
number_of_bits--;
|
|
|
|
remainder <<= 1;
|
|
|
|
if ((remainder & 0x100) != 0)
|
|
|
|
remainder ^= 0xd5;
|
|
|
|
}
|
|
|
|
|
|
|
|
return remainder & 0xff;
|
|
|
|
}
|
|
|
|
static inline u8 drm_dp_calc_sb_hdr_size(struct drm_dp_sideband_msg_hdr *hdr)
|
|
|
|
{
|
|
|
|
u8 size = 3;
|
|
|
|
size += (hdr->lct / 2);
|
|
|
|
return size;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_encode_sideband_msg_hdr(struct drm_dp_sideband_msg_hdr *hdr,
|
|
|
|
u8 *buf, int *len)
|
|
|
|
{
|
|
|
|
int idx = 0;
|
|
|
|
int i;
|
|
|
|
u8 crc4;
|
|
|
|
buf[idx++] = ((hdr->lct & 0xf) << 4) | (hdr->lcr & 0xf);
|
|
|
|
for (i = 0; i < (hdr->lct / 2); i++)
|
|
|
|
buf[idx++] = hdr->rad[i];
|
|
|
|
buf[idx++] = (hdr->broadcast << 7) | (hdr->path_msg << 6) |
|
|
|
|
(hdr->msg_len & 0x3f);
|
|
|
|
buf[idx++] = (hdr->somt << 7) | (hdr->eomt << 6) | (hdr->seqno << 4);
|
|
|
|
|
|
|
|
crc4 = drm_dp_msg_header_crc4(buf, (idx * 2) - 1);
|
|
|
|
buf[idx - 1] |= (crc4 & 0xf);
|
|
|
|
|
|
|
|
*len = idx;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_decode_sideband_msg_hdr(struct drm_dp_sideband_msg_hdr *hdr,
|
|
|
|
u8 *buf, int buflen, u8 *hdrlen)
|
|
|
|
{
|
|
|
|
u8 crc4;
|
|
|
|
u8 len;
|
|
|
|
int i;
|
|
|
|
u8 idx;
|
|
|
|
if (buf[0] == 0)
|
|
|
|
return false;
|
|
|
|
len = 3;
|
|
|
|
len += ((buf[0] & 0xf0) >> 4) / 2;
|
|
|
|
if (len > buflen)
|
|
|
|
return false;
|
|
|
|
crc4 = drm_dp_msg_header_crc4(buf, (len * 2) - 1);
|
|
|
|
|
|
|
|
if ((crc4 & 0xf) != (buf[len - 1] & 0xf)) {
|
|
|
|
DRM_DEBUG_KMS("crc4 mismatch 0x%x 0x%x\n", crc4, buf[len - 1]);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
hdr->lct = (buf[0] & 0xf0) >> 4;
|
|
|
|
hdr->lcr = (buf[0] & 0xf);
|
|
|
|
idx = 1;
|
|
|
|
for (i = 0; i < (hdr->lct / 2); i++)
|
|
|
|
hdr->rad[i] = buf[idx++];
|
|
|
|
hdr->broadcast = (buf[idx] >> 7) & 0x1;
|
|
|
|
hdr->path_msg = (buf[idx] >> 6) & 0x1;
|
|
|
|
hdr->msg_len = buf[idx] & 0x3f;
|
|
|
|
idx++;
|
|
|
|
hdr->somt = (buf[idx] >> 7) & 0x1;
|
|
|
|
hdr->eomt = (buf[idx] >> 6) & 0x1;
|
|
|
|
hdr->seqno = (buf[idx] >> 4) & 0x1;
|
|
|
|
idx++;
|
|
|
|
*hdrlen = idx;
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2019-09-04 04:45:45 +08:00
|
|
|
void
|
|
|
|
drm_dp_encode_sideband_req(const struct drm_dp_sideband_msg_req_body *req,
|
|
|
|
struct drm_dp_sideband_msg_tx *raw)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int idx = 0;
|
|
|
|
int i;
|
|
|
|
u8 *buf = raw->msg;
|
|
|
|
buf[idx++] = req->req_type & 0x7f;
|
|
|
|
|
|
|
|
switch (req->req_type) {
|
|
|
|
case DP_ENUM_PATH_RESOURCES:
|
2019-09-04 05:57:02 +08:00
|
|
|
case DP_POWER_DOWN_PHY:
|
|
|
|
case DP_POWER_UP_PHY:
|
2014-06-05 12:01:32 +08:00
|
|
|
buf[idx] = (req->u.port_num.port_number & 0xf) << 4;
|
|
|
|
idx++;
|
|
|
|
break;
|
|
|
|
case DP_ALLOCATE_PAYLOAD:
|
|
|
|
buf[idx] = (req->u.allocate_payload.port_number & 0xf) << 4 |
|
|
|
|
(req->u.allocate_payload.number_sdp_streams & 0xf);
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.allocate_payload.vcpi & 0x7f);
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.allocate_payload.pbn >> 8);
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.allocate_payload.pbn & 0xff);
|
|
|
|
idx++;
|
|
|
|
for (i = 0; i < req->u.allocate_payload.number_sdp_streams / 2; i++) {
|
|
|
|
buf[idx] = ((req->u.allocate_payload.sdp_stream_sink[i * 2] & 0xf) << 4) |
|
|
|
|
(req->u.allocate_payload.sdp_stream_sink[i * 2 + 1] & 0xf);
|
|
|
|
idx++;
|
|
|
|
}
|
|
|
|
if (req->u.allocate_payload.number_sdp_streams & 1) {
|
|
|
|
i = req->u.allocate_payload.number_sdp_streams - 1;
|
|
|
|
buf[idx] = (req->u.allocate_payload.sdp_stream_sink[i] & 0xf) << 4;
|
|
|
|
idx++;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case DP_QUERY_PAYLOAD:
|
|
|
|
buf[idx] = (req->u.query_payload.port_number & 0xf) << 4;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.query_payload.vcpi & 0x7f);
|
|
|
|
idx++;
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_DPCD_READ:
|
|
|
|
buf[idx] = (req->u.dpcd_read.port_number & 0xf) << 4;
|
|
|
|
buf[idx] |= ((req->u.dpcd_read.dpcd_address & 0xf0000) >> 16) & 0xf;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.dpcd_read.dpcd_address & 0xff00) >> 8;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.dpcd_read.dpcd_address & 0xff);
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.dpcd_read.num_bytes);
|
|
|
|
idx++;
|
|
|
|
break;
|
|
|
|
|
|
|
|
case DP_REMOTE_DPCD_WRITE:
|
|
|
|
buf[idx] = (req->u.dpcd_write.port_number & 0xf) << 4;
|
|
|
|
buf[idx] |= ((req->u.dpcd_write.dpcd_address & 0xf0000) >> 16) & 0xf;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.dpcd_write.dpcd_address & 0xff00) >> 8;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.dpcd_write.dpcd_address & 0xff);
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.dpcd_write.num_bytes);
|
|
|
|
idx++;
|
|
|
|
memcpy(&buf[idx], req->u.dpcd_write.bytes, req->u.dpcd_write.num_bytes);
|
|
|
|
idx += req->u.dpcd_write.num_bytes;
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_READ:
|
|
|
|
buf[idx] = (req->u.i2c_read.port_number & 0xf) << 4;
|
|
|
|
buf[idx] |= (req->u.i2c_read.num_transactions & 0x3);
|
|
|
|
idx++;
|
|
|
|
for (i = 0; i < (req->u.i2c_read.num_transactions & 0x3); i++) {
|
|
|
|
buf[idx] = req->u.i2c_read.transactions[i].i2c_dev_id & 0x7f;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = req->u.i2c_read.transactions[i].num_bytes;
|
|
|
|
idx++;
|
|
|
|
memcpy(&buf[idx], req->u.i2c_read.transactions[i].bytes, req->u.i2c_read.transactions[i].num_bytes);
|
|
|
|
idx += req->u.i2c_read.transactions[i].num_bytes;
|
|
|
|
|
2020-01-03 13:50:01 +08:00
|
|
|
buf[idx] = (req->u.i2c_read.transactions[i].no_stop_bit & 0x1) << 4;
|
2014-06-05 12:01:32 +08:00
|
|
|
buf[idx] |= (req->u.i2c_read.transactions[i].i2c_transaction_delay & 0xf);
|
|
|
|
idx++;
|
|
|
|
}
|
|
|
|
buf[idx] = (req->u.i2c_read.read_i2c_device_id) & 0x7f;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.i2c_read.num_bytes_read);
|
|
|
|
idx++;
|
|
|
|
break;
|
|
|
|
|
|
|
|
case DP_REMOTE_I2C_WRITE:
|
|
|
|
buf[idx] = (req->u.i2c_write.port_number & 0xf) << 4;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.i2c_write.write_i2c_device_id) & 0x7f;
|
|
|
|
idx++;
|
|
|
|
buf[idx] = (req->u.i2c_write.num_bytes);
|
|
|
|
idx++;
|
|
|
|
memcpy(&buf[idx], req->u.i2c_write.bytes, req->u.i2c_write.num_bytes);
|
|
|
|
idx += req->u.i2c_write.num_bytes;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
raw->cur_len = idx;
|
|
|
|
}
|
2019-09-04 04:45:45 +08:00
|
|
|
EXPORT_SYMBOL_FOR_TESTS_ONLY(drm_dp_encode_sideband_req);
|
|
|
|
|
|
|
|
/* Decode a sideband request we've encoded, mainly used for debugging */
|
|
|
|
int
|
|
|
|
drm_dp_decode_sideband_req(const struct drm_dp_sideband_msg_tx *raw,
|
|
|
|
struct drm_dp_sideband_msg_req_body *req)
|
|
|
|
{
|
|
|
|
const u8 *buf = raw->msg;
|
|
|
|
int i, idx = 0;
|
|
|
|
|
|
|
|
req->req_type = buf[idx++] & 0x7f;
|
|
|
|
switch (req->req_type) {
|
|
|
|
case DP_ENUM_PATH_RESOURCES:
|
|
|
|
case DP_POWER_DOWN_PHY:
|
|
|
|
case DP_POWER_UP_PHY:
|
|
|
|
req->u.port_num.port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
break;
|
|
|
|
case DP_ALLOCATE_PAYLOAD:
|
|
|
|
{
|
|
|
|
struct drm_dp_allocate_payload *a =
|
|
|
|
&req->u.allocate_payload;
|
|
|
|
|
|
|
|
a->number_sdp_streams = buf[idx] & 0xf;
|
|
|
|
a->port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
|
|
|
|
WARN_ON(buf[++idx] & 0x80);
|
|
|
|
a->vcpi = buf[idx] & 0x7f;
|
|
|
|
|
|
|
|
a->pbn = buf[++idx] << 8;
|
|
|
|
a->pbn |= buf[++idx];
|
|
|
|
|
|
|
|
idx++;
|
|
|
|
for (i = 0; i < a->number_sdp_streams; i++) {
|
|
|
|
a->sdp_stream_sink[i] =
|
|
|
|
(buf[idx + (i / 2)] >> ((i % 2) ? 0 : 4)) & 0xf;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case DP_QUERY_PAYLOAD:
|
|
|
|
req->u.query_payload.port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
WARN_ON(buf[++idx] & 0x80);
|
|
|
|
req->u.query_payload.vcpi = buf[idx] & 0x7f;
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_DPCD_READ:
|
|
|
|
{
|
|
|
|
struct drm_dp_remote_dpcd_read *r = &req->u.dpcd_read;
|
|
|
|
|
|
|
|
r->port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
|
|
|
|
r->dpcd_address = (buf[idx] << 16) & 0xf0000;
|
|
|
|
r->dpcd_address |= (buf[++idx] << 8) & 0xff00;
|
|
|
|
r->dpcd_address |= buf[++idx] & 0xff;
|
|
|
|
|
|
|
|
r->num_bytes = buf[++idx];
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_DPCD_WRITE:
|
|
|
|
{
|
|
|
|
struct drm_dp_remote_dpcd_write *w =
|
|
|
|
&req->u.dpcd_write;
|
|
|
|
|
|
|
|
w->port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
|
|
|
|
w->dpcd_address = (buf[idx] << 16) & 0xf0000;
|
|
|
|
w->dpcd_address |= (buf[++idx] << 8) & 0xff00;
|
|
|
|
w->dpcd_address |= buf[++idx] & 0xff;
|
|
|
|
|
|
|
|
w->num_bytes = buf[++idx];
|
|
|
|
|
|
|
|
w->bytes = kmemdup(&buf[++idx], w->num_bytes,
|
|
|
|
GFP_KERNEL);
|
|
|
|
if (!w->bytes)
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_READ:
|
|
|
|
{
|
|
|
|
struct drm_dp_remote_i2c_read *r = &req->u.i2c_read;
|
|
|
|
struct drm_dp_remote_i2c_read_tx *tx;
|
|
|
|
bool failed = false;
|
|
|
|
|
|
|
|
r->num_transactions = buf[idx] & 0x3;
|
|
|
|
r->port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
for (i = 0; i < r->num_transactions; i++) {
|
|
|
|
tx = &r->transactions[i];
|
|
|
|
|
|
|
|
tx->i2c_dev_id = buf[++idx] & 0x7f;
|
|
|
|
tx->num_bytes = buf[++idx];
|
|
|
|
tx->bytes = kmemdup(&buf[++idx],
|
|
|
|
tx->num_bytes,
|
|
|
|
GFP_KERNEL);
|
|
|
|
if (!tx->bytes) {
|
|
|
|
failed = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
idx += tx->num_bytes;
|
|
|
|
tx->no_stop_bit = (buf[idx] >> 5) & 0x1;
|
|
|
|
tx->i2c_transaction_delay = buf[idx] & 0xf;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (failed) {
|
|
|
|
for (i = 0; i < r->num_transactions; i++)
|
|
|
|
kfree(tx->bytes);
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
|
|
|
r->read_i2c_device_id = buf[++idx] & 0x7f;
|
|
|
|
r->num_bytes_read = buf[++idx];
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_WRITE:
|
|
|
|
{
|
|
|
|
struct drm_dp_remote_i2c_write *w = &req->u.i2c_write;
|
|
|
|
|
|
|
|
w->port_number = (buf[idx] >> 4) & 0xf;
|
|
|
|
w->write_i2c_device_id = buf[++idx] & 0x7f;
|
|
|
|
w->num_bytes = buf[++idx];
|
|
|
|
w->bytes = kmemdup(&buf[++idx], w->num_bytes,
|
|
|
|
GFP_KERNEL);
|
|
|
|
if (!w->bytes)
|
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_FOR_TESTS_ONLY(drm_dp_decode_sideband_req);
|
|
|
|
|
|
|
|
void
|
|
|
|
drm_dp_dump_sideband_msg_req_body(const struct drm_dp_sideband_msg_req_body *req,
|
|
|
|
int indent, struct drm_printer *printer)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
#define P(f, ...) drm_printf_indent(printer, indent, f, ##__VA_ARGS__)
|
|
|
|
if (req->req_type == DP_LINK_ADDRESS) {
|
|
|
|
/* No contents to print */
|
|
|
|
P("type=%s\n", drm_dp_mst_req_type_str(req->req_type));
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
P("type=%s contents:\n", drm_dp_mst_req_type_str(req->req_type));
|
|
|
|
indent++;
|
|
|
|
|
|
|
|
switch (req->req_type) {
|
|
|
|
case DP_ENUM_PATH_RESOURCES:
|
|
|
|
case DP_POWER_DOWN_PHY:
|
|
|
|
case DP_POWER_UP_PHY:
|
|
|
|
P("port=%d\n", req->u.port_num.port_number);
|
|
|
|
break;
|
|
|
|
case DP_ALLOCATE_PAYLOAD:
|
|
|
|
P("port=%d vcpi=%d pbn=%d sdp_streams=%d %*ph\n",
|
|
|
|
req->u.allocate_payload.port_number,
|
|
|
|
req->u.allocate_payload.vcpi, req->u.allocate_payload.pbn,
|
|
|
|
req->u.allocate_payload.number_sdp_streams,
|
|
|
|
req->u.allocate_payload.number_sdp_streams,
|
|
|
|
req->u.allocate_payload.sdp_stream_sink);
|
|
|
|
break;
|
|
|
|
case DP_QUERY_PAYLOAD:
|
|
|
|
P("port=%d vcpi=%d\n",
|
|
|
|
req->u.query_payload.port_number,
|
|
|
|
req->u.query_payload.vcpi);
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_DPCD_READ:
|
|
|
|
P("port=%d dpcd_addr=%05x len=%d\n",
|
|
|
|
req->u.dpcd_read.port_number, req->u.dpcd_read.dpcd_address,
|
|
|
|
req->u.dpcd_read.num_bytes);
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_DPCD_WRITE:
|
|
|
|
P("port=%d addr=%05x len=%d: %*ph\n",
|
|
|
|
req->u.dpcd_write.port_number,
|
|
|
|
req->u.dpcd_write.dpcd_address,
|
|
|
|
req->u.dpcd_write.num_bytes, req->u.dpcd_write.num_bytes,
|
|
|
|
req->u.dpcd_write.bytes);
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_READ:
|
|
|
|
P("port=%d num_tx=%d id=%d size=%d:\n",
|
|
|
|
req->u.i2c_read.port_number,
|
|
|
|
req->u.i2c_read.num_transactions,
|
|
|
|
req->u.i2c_read.read_i2c_device_id,
|
|
|
|
req->u.i2c_read.num_bytes_read);
|
|
|
|
|
|
|
|
indent++;
|
|
|
|
for (i = 0; i < req->u.i2c_read.num_transactions; i++) {
|
|
|
|
const struct drm_dp_remote_i2c_read_tx *rtx =
|
|
|
|
&req->u.i2c_read.transactions[i];
|
|
|
|
|
|
|
|
P("%d: id=%03d size=%03d no_stop_bit=%d tx_delay=%03d: %*ph\n",
|
|
|
|
i, rtx->i2c_dev_id, rtx->num_bytes,
|
|
|
|
rtx->no_stop_bit, rtx->i2c_transaction_delay,
|
|
|
|
rtx->num_bytes, rtx->bytes);
|
|
|
|
}
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_WRITE:
|
|
|
|
P("port=%d id=%d size=%d: %*ph\n",
|
|
|
|
req->u.i2c_write.port_number,
|
|
|
|
req->u.i2c_write.write_i2c_device_id,
|
|
|
|
req->u.i2c_write.num_bytes, req->u.i2c_write.num_bytes,
|
|
|
|
req->u.i2c_write.bytes);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
P("???\n");
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
#undef P
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL_FOR_TESTS_ONLY(drm_dp_dump_sideband_msg_req_body);
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
drm_dp_mst_dump_sideband_msg_tx(struct drm_printer *p,
|
|
|
|
const struct drm_dp_sideband_msg_tx *txmsg)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
char buf[64];
|
|
|
|
int ret;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
drm_dp_mst_rad_to_str(txmsg->dst->rad, txmsg->dst->lct, buf,
|
|
|
|
sizeof(buf));
|
|
|
|
drm_printf(p, "txmsg cur_offset=%x cur_len=%x seqno=%x state=%s path_msg=%d dst=%s\n",
|
|
|
|
txmsg->cur_offset, txmsg->cur_len, txmsg->seqno,
|
|
|
|
drm_dp_mst_sideband_tx_state_str(txmsg->state),
|
|
|
|
txmsg->path_msg, buf);
|
|
|
|
|
|
|
|
ret = drm_dp_decode_sideband_req(txmsg, &req);
|
|
|
|
if (ret) {
|
|
|
|
drm_printf(p, "<failed to decode sideband req: %d>\n", ret);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
drm_dp_dump_sideband_msg_req_body(&req, 1, p);
|
|
|
|
|
|
|
|
switch (req.req_type) {
|
|
|
|
case DP_REMOTE_DPCD_WRITE:
|
|
|
|
kfree(req.u.dpcd_write.bytes);
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_READ:
|
|
|
|
for (i = 0; i < req.u.i2c_read.num_transactions; i++)
|
|
|
|
kfree(req.u.i2c_read.transactions[i].bytes);
|
|
|
|
break;
|
|
|
|
case DP_REMOTE_I2C_WRITE:
|
|
|
|
kfree(req.u.i2c_write.bytes);
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
static void drm_dp_crc_sideband_chunk_req(u8 *msg, u8 len)
|
|
|
|
{
|
|
|
|
u8 crc4;
|
|
|
|
crc4 = drm_dp_msg_data_crc4(msg, len);
|
|
|
|
msg[len] = crc4;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_encode_sideband_reply(struct drm_dp_sideband_msg_reply_body *rep,
|
|
|
|
struct drm_dp_sideband_msg_tx *raw)
|
|
|
|
{
|
|
|
|
int idx = 0;
|
|
|
|
u8 *buf = raw->msg;
|
|
|
|
|
|
|
|
buf[idx++] = (rep->reply_type & 0x1) << 7 | (rep->req_type & 0x7f);
|
|
|
|
|
|
|
|
raw->cur_len = idx;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* this adds a chunk of msg to the builder to get the final msg */
|
|
|
|
static bool drm_dp_sideband_msg_build(struct drm_dp_sideband_msg_rx *msg,
|
|
|
|
u8 *replybuf, u8 replybuflen, bool hdr)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
u8 crc4;
|
|
|
|
|
|
|
|
if (hdr) {
|
|
|
|
u8 hdrlen;
|
|
|
|
struct drm_dp_sideband_msg_hdr recv_hdr;
|
|
|
|
ret = drm_dp_decode_sideband_msg_hdr(&recv_hdr, replybuf, replybuflen, &hdrlen);
|
|
|
|
if (ret == false) {
|
|
|
|
print_hex_dump(KERN_DEBUG, "failed hdr", DUMP_PREFIX_NONE, 16, 1, replybuf, replybuflen, false);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-07-19 21:46:32 +08:00
|
|
|
/*
|
|
|
|
* ignore out-of-order messages or messages that are part of a
|
|
|
|
* failed transaction
|
|
|
|
*/
|
|
|
|
if (!recv_hdr.somt && !msg->have_somt)
|
|
|
|
return false;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* get length contained in this portion */
|
|
|
|
msg->curchunk_len = recv_hdr.msg_len;
|
|
|
|
msg->curchunk_hdrlen = hdrlen;
|
|
|
|
|
|
|
|
/* we have already gotten an somt - don't bother parsing */
|
|
|
|
if (recv_hdr.somt && msg->have_somt)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
if (recv_hdr.somt) {
|
|
|
|
memcpy(&msg->initial_hdr, &recv_hdr, sizeof(struct drm_dp_sideband_msg_hdr));
|
|
|
|
msg->have_somt = true;
|
|
|
|
}
|
|
|
|
if (recv_hdr.eomt)
|
|
|
|
msg->have_eomt = true;
|
|
|
|
|
|
|
|
/* copy the bytes for the remainder of this header chunk */
|
|
|
|
msg->curchunk_idx = min(msg->curchunk_len, (u8)(replybuflen - hdrlen));
|
|
|
|
memcpy(&msg->chunk[0], replybuf + hdrlen, msg->curchunk_idx);
|
|
|
|
} else {
|
|
|
|
memcpy(&msg->chunk[msg->curchunk_idx], replybuf, replybuflen);
|
|
|
|
msg->curchunk_idx += replybuflen;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (msg->curchunk_idx >= msg->curchunk_len) {
|
|
|
|
/* do CRC */
|
|
|
|
crc4 = drm_dp_msg_data_crc4(msg->chunk, msg->curchunk_len - 1);
|
|
|
|
/* copy chunk into bigger msg */
|
|
|
|
memcpy(&msg->msg[msg->curlen], msg->chunk, msg->curchunk_len - 1);
|
|
|
|
msg->curlen += msg->curchunk_len - 1;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_link_address(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
int i;
|
|
|
|
memcpy(repmsg->u.link_addr.guid, &raw->msg[idx], 16);
|
|
|
|
idx += 16;
|
|
|
|
repmsg->u.link_addr.nports = raw->msg[idx] & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
for (i = 0; i < repmsg->u.link_addr.nports; i++) {
|
|
|
|
if (raw->msg[idx] & 0x80)
|
|
|
|
repmsg->u.link_addr.ports[i].input_port = 1;
|
|
|
|
|
|
|
|
repmsg->u.link_addr.ports[i].peer_device_type = (raw->msg[idx] >> 4) & 0x7;
|
|
|
|
repmsg->u.link_addr.ports[i].port_number = (raw->msg[idx] & 0xf);
|
|
|
|
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.link_addr.ports[i].mcs = (raw->msg[idx] >> 7) & 0x1;
|
|
|
|
repmsg->u.link_addr.ports[i].ddps = (raw->msg[idx] >> 6) & 0x1;
|
|
|
|
if (repmsg->u.link_addr.ports[i].input_port == 0)
|
|
|
|
repmsg->u.link_addr.ports[i].legacy_device_plug_status = (raw->msg[idx] >> 5) & 0x1;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
if (repmsg->u.link_addr.ports[i].input_port == 0) {
|
|
|
|
repmsg->u.link_addr.ports[i].dpcd_revision = (raw->msg[idx]);
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
memcpy(repmsg->u.link_addr.ports[i].peer_guid, &raw->msg[idx], 16);
|
|
|
|
idx += 16;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.link_addr.ports[i].num_sdp_streams = (raw->msg[idx] >> 4) & 0xf;
|
|
|
|
repmsg->u.link_addr.ports[i].num_sdp_stream_sinks = (raw->msg[idx] & 0xf);
|
|
|
|
idx++;
|
|
|
|
|
|
|
|
}
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
}
|
|
|
|
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("link address reply parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_remote_dpcd_read(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
repmsg->u.remote_dpcd_read_ack.port_number = raw->msg[idx] & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.remote_dpcd_read_ack.num_bytes = raw->msg[idx];
|
2018-08-27 16:07:42 +08:00
|
|
|
idx++;
|
2014-06-05 12:01:32 +08:00
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
|
|
|
|
memcpy(repmsg->u.remote_dpcd_read_ack.bytes, &raw->msg[idx], repmsg->u.remote_dpcd_read_ack.num_bytes);
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("link address reply parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_remote_dpcd_write(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
repmsg->u.remote_dpcd_write_ack.port_number = raw->msg[idx] & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_remote_i2c_read_ack(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
|
|
|
|
repmsg->u.remote_i2c_read_ack.port_number = (raw->msg[idx] & 0xf);
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.remote_i2c_read_ack.num_bytes = raw->msg[idx];
|
|
|
|
idx++;
|
|
|
|
/* TODO check */
|
|
|
|
memcpy(repmsg->u.remote_i2c_read_ack.bytes, &raw->msg[idx], repmsg->u.remote_i2c_read_ack.num_bytes);
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("remote i2c reply parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_enum_path_resources_ack(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
repmsg->u.path_resources.port_number = (raw->msg[idx] >> 4) & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.path_resources.full_payload_bw_number = (raw->msg[idx] << 8) | (raw->msg[idx+1]);
|
|
|
|
idx += 2;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.path_resources.avail_payload_bw_number = (raw->msg[idx] << 8) | (raw->msg[idx+1]);
|
|
|
|
idx += 2;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("enum resource parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_allocate_payload_ack(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
repmsg->u.allocate_payload.port_number = (raw->msg[idx] >> 4) & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.allocate_payload.vcpi = raw->msg[idx];
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.allocate_payload.allocated_pbn = (raw->msg[idx] << 8) | (raw->msg[idx+1]);
|
|
|
|
idx += 2;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("allocate payload parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_query_payload_ack(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
repmsg->u.query_payload.port_number = (raw->msg[idx] >> 4) & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
repmsg->u.query_payload.allocated_pbn = (raw->msg[idx] << 8) | (raw->msg[idx + 1]);
|
|
|
|
idx += 2;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("query payload parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
2017-09-07 08:14:58 +08:00
|
|
|
static bool drm_dp_sideband_parse_power_updown_phy_ack(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *repmsg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
|
|
|
|
repmsg->u.port_number.port_number = (raw->msg[idx] >> 4) & 0xf;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen) {
|
|
|
|
DRM_DEBUG_KMS("power up/down phy parse length fail %d %d\n",
|
|
|
|
idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static bool drm_dp_sideband_parse_reply(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_reply_body *msg)
|
|
|
|
{
|
|
|
|
memset(msg, 0, sizeof(*msg));
|
|
|
|
msg->reply_type = (raw->msg[0] & 0x80) >> 7;
|
|
|
|
msg->req_type = (raw->msg[0] & 0x7f);
|
|
|
|
|
2019-01-23 04:03:00 +08:00
|
|
|
if (msg->reply_type == DP_SIDEBAND_REPLY_NAK) {
|
2014-06-05 12:01:32 +08:00
|
|
|
memcpy(msg->u.nak.guid, &raw->msg[1], 16);
|
|
|
|
msg->u.nak.reason = raw->msg[17];
|
|
|
|
msg->u.nak.nak_data = raw->msg[18];
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
switch (msg->req_type) {
|
|
|
|
case DP_LINK_ADDRESS:
|
|
|
|
return drm_dp_sideband_parse_link_address(raw, msg);
|
|
|
|
case DP_QUERY_PAYLOAD:
|
|
|
|
return drm_dp_sideband_parse_query_payload_ack(raw, msg);
|
|
|
|
case DP_REMOTE_DPCD_READ:
|
|
|
|
return drm_dp_sideband_parse_remote_dpcd_read(raw, msg);
|
|
|
|
case DP_REMOTE_DPCD_WRITE:
|
|
|
|
return drm_dp_sideband_parse_remote_dpcd_write(raw, msg);
|
|
|
|
case DP_REMOTE_I2C_READ:
|
|
|
|
return drm_dp_sideband_parse_remote_i2c_read_ack(raw, msg);
|
|
|
|
case DP_ENUM_PATH_RESOURCES:
|
|
|
|
return drm_dp_sideband_parse_enum_path_resources_ack(raw, msg);
|
|
|
|
case DP_ALLOCATE_PAYLOAD:
|
|
|
|
return drm_dp_sideband_parse_allocate_payload_ack(raw, msg);
|
2017-09-07 08:14:58 +08:00
|
|
|
case DP_POWER_DOWN_PHY:
|
|
|
|
case DP_POWER_UP_PHY:
|
|
|
|
return drm_dp_sideband_parse_power_updown_phy_ack(raw, msg);
|
2014-06-05 12:01:32 +08:00
|
|
|
default:
|
2019-01-23 04:03:01 +08:00
|
|
|
DRM_ERROR("Got unknown reply 0x%02x (%s)\n", msg->req_type,
|
|
|
|
drm_dp_mst_req_type_str(msg->req_type));
|
2014-06-05 12:01:32 +08:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_connection_status_notify(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_req_body *msg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
|
|
|
|
msg->u.conn_stat.port_number = (raw->msg[idx] & 0xf0) >> 4;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
|
|
|
|
memcpy(msg->u.conn_stat.guid, &raw->msg[idx], 16);
|
|
|
|
idx += 16;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
|
|
|
|
msg->u.conn_stat.legacy_device_plug_status = (raw->msg[idx] >> 6) & 0x1;
|
|
|
|
msg->u.conn_stat.displayport_device_plug_status = (raw->msg[idx] >> 5) & 0x1;
|
|
|
|
msg->u.conn_stat.message_capability_status = (raw->msg[idx] >> 4) & 0x1;
|
|
|
|
msg->u.conn_stat.input_port = (raw->msg[idx] >> 3) & 0x1;
|
|
|
|
msg->u.conn_stat.peer_device_type = (raw->msg[idx] & 0x7);
|
|
|
|
idx++;
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("connection status reply parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_resource_status_notify(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_req_body *msg)
|
|
|
|
{
|
|
|
|
int idx = 1;
|
|
|
|
|
|
|
|
msg->u.resource_stat.port_number = (raw->msg[idx] & 0xf0) >> 4;
|
|
|
|
idx++;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
|
|
|
|
memcpy(msg->u.resource_stat.guid, &raw->msg[idx], 16);
|
|
|
|
idx += 16;
|
|
|
|
if (idx > raw->curlen)
|
|
|
|
goto fail_len;
|
|
|
|
|
|
|
|
msg->u.resource_stat.available_pbn = (raw->msg[idx] << 8) | (raw->msg[idx + 1]);
|
|
|
|
idx++;
|
|
|
|
return true;
|
|
|
|
fail_len:
|
|
|
|
DRM_DEBUG_KMS("resource status reply parse length fail %d %d\n", idx, raw->curlen);
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_sideband_parse_req(struct drm_dp_sideband_msg_rx *raw,
|
|
|
|
struct drm_dp_sideband_msg_req_body *msg)
|
|
|
|
{
|
|
|
|
memset(msg, 0, sizeof(*msg));
|
|
|
|
msg->req_type = (raw->msg[0] & 0x7f);
|
|
|
|
|
|
|
|
switch (msg->req_type) {
|
|
|
|
case DP_CONNECTION_STATUS_NOTIFY:
|
|
|
|
return drm_dp_sideband_parse_connection_status_notify(raw, msg);
|
|
|
|
case DP_RESOURCE_STATUS_NOTIFY:
|
|
|
|
return drm_dp_sideband_parse_resource_status_notify(raw, msg);
|
|
|
|
default:
|
2019-01-23 04:03:01 +08:00
|
|
|
DRM_ERROR("Got unknown request 0x%02x (%s)\n", msg->req_type,
|
|
|
|
drm_dp_mst_req_type_str(msg->req_type));
|
2014-06-05 12:01:32 +08:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
static int build_dpcd_write(struct drm_dp_sideband_msg_tx *msg, u8 port_num, u32 offset, u8 num_bytes, u8 *bytes)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
|
|
|
|
req.req_type = DP_REMOTE_DPCD_WRITE;
|
|
|
|
req.u.dpcd_write.port_number = port_num;
|
|
|
|
req.u.dpcd_write.dpcd_address = offset;
|
|
|
|
req.u.dpcd_write.num_bytes = num_bytes;
|
|
|
|
req.u.dpcd_write.bytes = bytes;
|
|
|
|
drm_dp_encode_sideband_req(&req, msg);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int build_link_address(struct drm_dp_sideband_msg_tx *msg)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
|
|
|
|
req.req_type = DP_LINK_ADDRESS;
|
|
|
|
drm_dp_encode_sideband_req(&req, msg);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int build_enum_path_resources(struct drm_dp_sideband_msg_tx *msg, int port_num)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
|
|
|
|
req.req_type = DP_ENUM_PATH_RESOURCES;
|
|
|
|
req.u.port_num.port_number = port_num;
|
|
|
|
drm_dp_encode_sideband_req(&req, msg);
|
|
|
|
msg->path_msg = true;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int build_allocate_payload(struct drm_dp_sideband_msg_tx *msg, int port_num,
|
2015-12-02 14:09:43 +08:00
|
|
|
u8 vcpi, uint16_t pbn,
|
|
|
|
u8 number_sdp_streams,
|
|
|
|
u8 *sdp_stream_sink)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
memset(&req, 0, sizeof(req));
|
|
|
|
req.req_type = DP_ALLOCATE_PAYLOAD;
|
|
|
|
req.u.allocate_payload.port_number = port_num;
|
|
|
|
req.u.allocate_payload.vcpi = vcpi;
|
|
|
|
req.u.allocate_payload.pbn = pbn;
|
2015-12-02 14:09:43 +08:00
|
|
|
req.u.allocate_payload.number_sdp_streams = number_sdp_streams;
|
|
|
|
memcpy(req.u.allocate_payload.sdp_stream_sink, sdp_stream_sink,
|
|
|
|
number_sdp_streams);
|
2014-06-05 12:01:32 +08:00
|
|
|
drm_dp_encode_sideband_req(&req, msg);
|
|
|
|
msg->path_msg = true;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2017-09-07 08:14:58 +08:00
|
|
|
static int build_power_updown_phy(struct drm_dp_sideband_msg_tx *msg,
|
|
|
|
int port_num, bool power_up)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
|
|
|
|
if (power_up)
|
|
|
|
req.req_type = DP_POWER_UP_PHY;
|
|
|
|
else
|
|
|
|
req.req_type = DP_POWER_DOWN_PHY;
|
|
|
|
|
|
|
|
req.u.port_num.port_number = port_num;
|
|
|
|
drm_dp_encode_sideband_req(&req, msg);
|
|
|
|
msg->path_msg = true;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static int drm_dp_mst_assign_payload_id(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_vcpi *vcpi)
|
|
|
|
{
|
2014-08-06 14:26:21 +08:00
|
|
|
int ret, vcpi_ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
mutex_lock(&mgr->payload_lock);
|
|
|
|
ret = find_first_zero_bit(&mgr->payload_mask, mgr->max_payloads + 1);
|
|
|
|
if (ret > mgr->max_payloads) {
|
|
|
|
ret = -EINVAL;
|
|
|
|
DRM_DEBUG_KMS("out of payload ids %d\n", ret);
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
2014-08-06 14:26:21 +08:00
|
|
|
vcpi_ret = find_first_zero_bit(&mgr->vcpi_mask, mgr->max_payloads + 1);
|
|
|
|
if (vcpi_ret > mgr->max_payloads) {
|
|
|
|
ret = -EINVAL;
|
|
|
|
DRM_DEBUG_KMS("out of vcpi ids %d\n", ret);
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
set_bit(ret, &mgr->payload_mask);
|
2014-08-06 14:26:21 +08:00
|
|
|
set_bit(vcpi_ret, &mgr->vcpi_mask);
|
|
|
|
vcpi->vcpi = vcpi_ret + 1;
|
2014-06-05 12:01:32 +08:00
|
|
|
mgr->proposed_vcpis[ret - 1] = vcpi;
|
|
|
|
out_unlock:
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_mst_put_payload_id(struct drm_dp_mst_topology_mgr *mgr,
|
2014-08-06 14:26:21 +08:00
|
|
|
int vcpi)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2014-08-06 14:26:21 +08:00
|
|
|
int i;
|
|
|
|
if (vcpi == 0)
|
2014-06-05 12:01:32 +08:00
|
|
|
return;
|
|
|
|
|
|
|
|
mutex_lock(&mgr->payload_lock);
|
2014-08-06 14:26:21 +08:00
|
|
|
DRM_DEBUG_KMS("putting payload %d\n", vcpi);
|
|
|
|
clear_bit(vcpi - 1, &mgr->vcpi_mask);
|
|
|
|
|
|
|
|
for (i = 0; i < mgr->max_payloads; i++) {
|
2019-09-25 22:14:40 +08:00
|
|
|
if (mgr->proposed_vcpis[i] &&
|
|
|
|
mgr->proposed_vcpis[i]->vcpi == vcpi) {
|
|
|
|
mgr->proposed_vcpis[i] = NULL;
|
|
|
|
clear_bit(i + 1, &mgr->payload_mask);
|
|
|
|
}
|
2014-08-06 14:26:21 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
static bool check_txmsg_state(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg)
|
|
|
|
{
|
2017-05-13 18:52:00 +08:00
|
|
|
unsigned int state;
|
2015-01-28 17:02:23 +08:00
|
|
|
|
|
|
|
/*
|
|
|
|
* All updates to txmsg->state are protected by mgr->qlock, and the two
|
|
|
|
* cases we check here are terminal states. For those the barriers
|
|
|
|
* provided by the wake_up/wait_event pair are enough.
|
|
|
|
*/
|
2017-05-13 18:52:00 +08:00
|
|
|
state = READ_ONCE(txmsg->state);
|
|
|
|
return (state == DRM_DP_SIDEBAND_TX_RX ||
|
|
|
|
state == DRM_DP_SIDEBAND_TX_TIMEOUT);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static int drm_dp_mst_wait_tx_reply(struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr = mstb->mgr;
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
ret = wait_event_timeout(mgr->tx_waitq,
|
|
|
|
check_txmsg_state(mgr, txmsg),
|
|
|
|
(4 * HZ));
|
|
|
|
mutex_lock(&mstb->mgr->qlock);
|
|
|
|
if (ret > 0) {
|
|
|
|
if (txmsg->state == DRM_DP_SIDEBAND_TX_TIMEOUT) {
|
|
|
|
ret = -EIO;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
DRM_DEBUG_KMS("timedout msg send %p %d %d\n", txmsg, txmsg->state, txmsg->seqno);
|
|
|
|
|
|
|
|
/* dump some state */
|
|
|
|
ret = -EIO;
|
|
|
|
|
|
|
|
/* remove from q */
|
|
|
|
if (txmsg->state == DRM_DP_SIDEBAND_TX_QUEUED ||
|
|
|
|
txmsg->state == DRM_DP_SIDEBAND_TX_START_SEND) {
|
|
|
|
list_del(&txmsg->next);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (txmsg->state == DRM_DP_SIDEBAND_TX_START_SEND ||
|
|
|
|
txmsg->state == DRM_DP_SIDEBAND_TX_SENT) {
|
|
|
|
mstb->tx_slots[txmsg->seqno] = NULL;
|
|
|
|
}
|
2020-01-13 17:36:49 +08:00
|
|
|
mgr->is_waiting_for_dwn_reply = false;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
out:
|
2019-10-01 22:06:14 +08:00
|
|
|
if (unlikely(ret == -EIO) && drm_debug_enabled(DRM_UT_DP)) {
|
2019-09-04 04:45:45 +08:00
|
|
|
struct drm_printer p = drm_debug_printer(DBG_PREFIX);
|
|
|
|
|
|
|
|
drm_dp_mst_dump_sideband_msg_tx(&p, txmsg);
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->qlock);
|
|
|
|
|
2020-01-13 17:36:49 +08:00
|
|
|
drm_dp_mst_kick_tx(mgr);
|
2014-06-05 12:01:32 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct drm_dp_mst_branch *drm_dp_add_mst_branch_device(u8 lct, u8 *rad)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
|
|
|
|
mstb = kzalloc(sizeof(*mstb), GFP_KERNEL);
|
|
|
|
if (!mstb)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
mstb->lct = lct;
|
|
|
|
if (lct > 1)
|
|
|
|
memcpy(mstb->rad, rad, lct / 2);
|
|
|
|
INIT_LIST_HEAD(&mstb->ports);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
kref_init(&mstb->topology_kref);
|
|
|
|
kref_init(&mstb->malloc_kref);
|
2014-06-05 12:01:32 +08:00
|
|
|
return mstb;
|
|
|
|
}
|
|
|
|
|
2016-01-27 22:39:36 +08:00
|
|
|
static void drm_dp_free_mst_branch_device(struct kref *kref)
|
|
|
|
{
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
struct drm_dp_mst_branch *mstb =
|
|
|
|
container_of(kref, struct drm_dp_mst_branch, malloc_kref);
|
|
|
|
|
|
|
|
if (mstb->port_parent)
|
|
|
|
drm_dp_mst_put_port_malloc(mstb->port_parent);
|
|
|
|
|
2016-01-27 22:39:36 +08:00
|
|
|
kfree(mstb);
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
/**
|
|
|
|
* DOC: Branch device and port refcounting
|
|
|
|
*
|
|
|
|
* Topology refcount overview
|
|
|
|
* ~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
*
|
|
|
|
* The refcounting schemes for &struct drm_dp_mst_branch and &struct
|
|
|
|
* drm_dp_mst_port are somewhat unusual. Both ports and branch devices have
|
|
|
|
* two different kinds of refcounts: topology refcounts, and malloc refcounts.
|
|
|
|
*
|
|
|
|
* Topology refcounts are not exposed to drivers, and are handled internally
|
|
|
|
* by the DP MST helpers. The helpers use them in order to prevent the
|
|
|
|
* in-memory topology state from being changed in the middle of critical
|
|
|
|
* operations like changing the internal state of payload allocations. This
|
|
|
|
* means each branch and port will be considered to be connected to the rest
|
2019-02-02 09:23:26 +08:00
|
|
|
* of the topology until its topology refcount reaches zero. Additionally,
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
* for ports this means that their associated &struct drm_connector will stay
|
|
|
|
* registered with userspace until the port's refcount reaches 0.
|
|
|
|
*
|
|
|
|
* Malloc refcount overview
|
|
|
|
* ~~~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
*
|
|
|
|
* Malloc references are used to keep a &struct drm_dp_mst_port or &struct
|
|
|
|
* drm_dp_mst_branch allocated even after all of its topology references have
|
|
|
|
* been dropped, so that the driver or MST helpers can safely access each
|
|
|
|
* branch's last known state before it was disconnected from the topology.
|
|
|
|
* When the malloc refcount of a port or branch reaches 0, the memory
|
|
|
|
* allocation containing the &struct drm_dp_mst_branch or &struct
|
|
|
|
* drm_dp_mst_port respectively will be freed.
|
|
|
|
*
|
|
|
|
* For &struct drm_dp_mst_branch, malloc refcounts are not currently exposed
|
|
|
|
* to drivers. As of writing this documentation, there are no drivers that
|
|
|
|
* have a usecase for accessing &struct drm_dp_mst_branch outside of the MST
|
|
|
|
* helpers. Exposing this API to drivers in a race-free manner would take more
|
|
|
|
* tweaking of the refcounting scheme, however patches are welcome provided
|
|
|
|
* there is a legitimate driver usecase for this.
|
|
|
|
*
|
|
|
|
* Refcount relationships in a topology
|
|
|
|
* ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
*
|
|
|
|
* Let's take a look at why the relationship between topology and malloc
|
|
|
|
* refcounts is designed the way it is.
|
|
|
|
*
|
|
|
|
* .. kernel-figure:: dp-mst/topology-figure-1.dot
|
|
|
|
*
|
|
|
|
* An example of topology and malloc refs in a DP MST topology with two
|
|
|
|
* active payloads. Topology refcount increments are indicated by solid
|
|
|
|
* lines, and malloc refcount increments are indicated by dashed lines.
|
|
|
|
* Each starts from the branch which incremented the refcount, and ends at
|
|
|
|
* the branch to which the refcount belongs to, i.e. the arrow points the
|
|
|
|
* same way as the C pointers used to reference a structure.
|
|
|
|
*
|
|
|
|
* As you can see in the above figure, every branch increments the topology
|
2019-02-02 09:23:26 +08:00
|
|
|
* refcount of its children, and increments the malloc refcount of its
|
|
|
|
* parent. Additionally, every payload increments the malloc refcount of its
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
* assigned port by 1.
|
|
|
|
*
|
|
|
|
* So, what would happen if MSTB #3 from the above figure was unplugged from
|
|
|
|
* the system, but the driver hadn't yet removed payload #2 from port #3? The
|
|
|
|
* topology would start to look like the figure below.
|
|
|
|
*
|
|
|
|
* .. kernel-figure:: dp-mst/topology-figure-2.dot
|
|
|
|
*
|
|
|
|
* Ports and branch devices which have been released from memory are
|
|
|
|
* colored grey, and references which have been removed are colored red.
|
|
|
|
*
|
|
|
|
* Whenever a port or branch device's topology refcount reaches zero, it will
|
|
|
|
* decrement the topology refcounts of all its children, the malloc refcount
|
|
|
|
* of its parent, and finally its own malloc refcount. For MSTB #4 and port
|
|
|
|
* #4, this means they both have been disconnected from the topology and freed
|
|
|
|
* from memory. But, because payload #2 is still holding a reference to port
|
2019-02-02 09:23:26 +08:00
|
|
|
* #3, port #3 is removed from the topology but its &struct drm_dp_mst_port
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
* is still accessible from memory. This also means port #3 has not yet
|
2019-02-02 09:23:26 +08:00
|
|
|
* decremented the malloc refcount of MSTB #3, so its &struct
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
* drm_dp_mst_branch will also stay allocated in memory until port #3's
|
|
|
|
* malloc refcount reaches 0.
|
|
|
|
*
|
|
|
|
* This relationship is necessary because in order to release payload #2, we
|
|
|
|
* need to be able to figure out the last relative of port #3 that's still
|
|
|
|
* connected to the topology. In this case, we would travel up the topology as
|
|
|
|
* shown below.
|
|
|
|
*
|
|
|
|
* .. kernel-figure:: dp-mst/topology-figure-3.dot
|
|
|
|
*
|
|
|
|
* And finally, remove payload #2 by communicating with port #2 through
|
|
|
|
* sideband transactions.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_get_mstb_malloc() - Increment the malloc refcount of a branch
|
|
|
|
* device
|
|
|
|
* @mstb: The &struct drm_dp_mst_branch to increment the malloc refcount of
|
|
|
|
*
|
|
|
|
* Increments &drm_dp_mst_branch.malloc_kref. When
|
|
|
|
* &drm_dp_mst_branch.malloc_kref reaches 0, the memory allocation for @mstb
|
|
|
|
* will be released and @mstb may no longer be used.
|
|
|
|
*
|
|
|
|
* See also: drm_dp_mst_put_mstb_malloc()
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
drm_dp_mst_get_mstb_malloc(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
kref_get(&mstb->malloc_kref);
|
|
|
|
DRM_DEBUG("mstb %p (%d)\n", mstb, kref_read(&mstb->malloc_kref));
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_put_mstb_malloc() - Decrement the malloc refcount of a branch
|
|
|
|
* device
|
|
|
|
* @mstb: The &struct drm_dp_mst_branch to decrement the malloc refcount of
|
|
|
|
*
|
|
|
|
* Decrements &drm_dp_mst_branch.malloc_kref. When
|
|
|
|
* &drm_dp_mst_branch.malloc_kref reaches 0, the memory allocation for @mstb
|
|
|
|
* will be released and @mstb may no longer be used.
|
|
|
|
*
|
|
|
|
* See also: drm_dp_mst_get_mstb_malloc()
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
drm_dp_mst_put_mstb_malloc(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
DRM_DEBUG("mstb %p (%d)\n", mstb, kref_read(&mstb->malloc_kref) - 1);
|
|
|
|
kref_put(&mstb->malloc_kref, drm_dp_free_mst_branch_device);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_free_mst_port(struct kref *kref)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port =
|
|
|
|
container_of(kref, struct drm_dp_mst_port, malloc_kref);
|
|
|
|
|
|
|
|
drm_dp_mst_put_mstb_malloc(port->parent);
|
|
|
|
kfree(port);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_get_port_malloc() - Increment the malloc refcount of an MST port
|
|
|
|
* @port: The &struct drm_dp_mst_port to increment the malloc refcount of
|
|
|
|
*
|
|
|
|
* Increments &drm_dp_mst_port.malloc_kref. When &drm_dp_mst_port.malloc_kref
|
|
|
|
* reaches 0, the memory allocation for @port will be released and @port may
|
|
|
|
* no longer be used.
|
|
|
|
*
|
|
|
|
* Because @port could potentially be freed at any time by the DP MST helpers
|
|
|
|
* if &drm_dp_mst_port.malloc_kref reaches 0, including during a call to this
|
|
|
|
* function, drivers that which to make use of &struct drm_dp_mst_port should
|
|
|
|
* ensure that they grab at least one main malloc reference to their MST ports
|
|
|
|
* in &drm_dp_mst_topology_cbs.add_connector. This callback is called before
|
|
|
|
* there is any chance for &drm_dp_mst_port.malloc_kref to reach 0.
|
|
|
|
*
|
|
|
|
* See also: drm_dp_mst_put_port_malloc()
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
drm_dp_mst_get_port_malloc(struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
kref_get(&port->malloc_kref);
|
|
|
|
DRM_DEBUG("port %p (%d)\n", port, kref_read(&port->malloc_kref));
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_get_port_malloc);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_put_port_malloc() - Decrement the malloc refcount of an MST port
|
|
|
|
* @port: The &struct drm_dp_mst_port to decrement the malloc refcount of
|
|
|
|
*
|
|
|
|
* Decrements &drm_dp_mst_port.malloc_kref. When &drm_dp_mst_port.malloc_kref
|
|
|
|
* reaches 0, the memory allocation for @port will be released and @port may
|
|
|
|
* no longer be used.
|
|
|
|
*
|
|
|
|
* See also: drm_dp_mst_get_port_malloc()
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
drm_dp_mst_put_port_malloc(struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
DRM_DEBUG("port %p (%d)\n", port, kref_read(&port->malloc_kref) - 1);
|
|
|
|
kref_put(&port->malloc_kref, drm_dp_free_mst_port);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_put_port_malloc);
|
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
#if IS_ENABLED(CONFIG_DRM_DEBUG_DP_MST_TOPOLOGY_REFS)
|
|
|
|
|
|
|
|
#define STACK_DEPTH 8
|
|
|
|
|
|
|
|
static noinline void
|
|
|
|
__topology_ref_save(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_topology_ref_history *history,
|
|
|
|
enum drm_dp_mst_topology_ref_type type)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
struct drm_dp_mst_topology_ref_entry *entry = NULL;
|
|
|
|
depot_stack_handle_t backtrace;
|
|
|
|
ulong stack_entries[STACK_DEPTH];
|
|
|
|
uint n;
|
|
|
|
int i;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
n = stack_trace_save(stack_entries, ARRAY_SIZE(stack_entries), 1);
|
|
|
|
backtrace = stack_depot_save(stack_entries, n, GFP_KERNEL);
|
|
|
|
if (!backtrace)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/* Try to find an existing entry for this backtrace */
|
|
|
|
for (i = 0; i < history->len; i++) {
|
|
|
|
if (history->entries[i].backtrace == backtrace) {
|
|
|
|
entry = &history->entries[i];
|
|
|
|
break;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
/* Otherwise add one */
|
|
|
|
if (!entry) {
|
|
|
|
struct drm_dp_mst_topology_ref_entry *new;
|
|
|
|
int new_len = history->len + 1;
|
|
|
|
|
|
|
|
new = krealloc(history->entries, sizeof(*new) * new_len,
|
|
|
|
GFP_KERNEL);
|
|
|
|
if (!new)
|
|
|
|
return;
|
|
|
|
|
|
|
|
entry = &new[history->len];
|
|
|
|
history->len = new_len;
|
|
|
|
history->entries = new;
|
|
|
|
|
|
|
|
entry->backtrace = backtrace;
|
|
|
|
entry->type = type;
|
|
|
|
entry->count = 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2019-06-21 05:59:25 +08:00
|
|
|
entry->count++;
|
|
|
|
entry->ts_nsec = ktime_get_ns();
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
|
|
|
topology_ref_history_cmp(const void *a, const void *b)
|
|
|
|
{
|
|
|
|
const struct drm_dp_mst_topology_ref_entry *entry_a = a, *entry_b = b;
|
|
|
|
|
|
|
|
if (entry_a->ts_nsec > entry_b->ts_nsec)
|
|
|
|
return 1;
|
|
|
|
else if (entry_a->ts_nsec < entry_b->ts_nsec)
|
|
|
|
return -1;
|
|
|
|
else
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline const char *
|
|
|
|
topology_ref_type_to_str(enum drm_dp_mst_topology_ref_type type)
|
|
|
|
{
|
|
|
|
if (type == DRM_DP_MST_TOPOLOGY_REF_GET)
|
|
|
|
return "get";
|
|
|
|
else
|
|
|
|
return "put";
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
__dump_topology_ref_history(struct drm_dp_mst_topology_ref_history *history,
|
|
|
|
void *ptr, const char *type_str)
|
|
|
|
{
|
|
|
|
struct drm_printer p = drm_debug_printer(DBG_PREFIX);
|
|
|
|
char *buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
|
|
|
|
int i;
|
|
|
|
|
|
|
|
if (!buf)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!history->len)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
/* First, sort the list so that it goes from oldest to newest
|
|
|
|
* reference entry
|
|
|
|
*/
|
|
|
|
sort(history->entries, history->len, sizeof(*history->entries),
|
|
|
|
topology_ref_history_cmp, NULL);
|
|
|
|
|
|
|
|
drm_printf(&p, "%s (%p) topology count reached 0, dumping history:\n",
|
|
|
|
type_str, ptr);
|
|
|
|
|
|
|
|
for (i = 0; i < history->len; i++) {
|
|
|
|
const struct drm_dp_mst_topology_ref_entry *entry =
|
|
|
|
&history->entries[i];
|
|
|
|
ulong *entries;
|
|
|
|
uint nr_entries;
|
|
|
|
u64 ts_nsec = entry->ts_nsec;
|
2019-11-07 03:41:15 +08:00
|
|
|
u32 rem_nsec = do_div(ts_nsec, 1000000000);
|
2019-06-21 05:59:25 +08:00
|
|
|
|
|
|
|
nr_entries = stack_depot_fetch(entry->backtrace, &entries);
|
|
|
|
stack_trace_snprint(buf, PAGE_SIZE, entries, nr_entries, 4);
|
|
|
|
|
2019-11-07 03:41:15 +08:00
|
|
|
drm_printf(&p, " %d %ss (last at %5llu.%06u):\n%s",
|
2019-06-21 05:59:25 +08:00
|
|
|
entry->count,
|
|
|
|
topology_ref_type_to_str(entry->type),
|
|
|
|
ts_nsec, rem_nsec / 1000, buf);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
/* Now free the history, since this is the only time we expose it */
|
|
|
|
kfree(history->entries);
|
|
|
|
out:
|
|
|
|
kfree(buf);
|
|
|
|
}
|
2016-01-27 22:39:36 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
static __always_inline void
|
|
|
|
drm_dp_mst_dump_mstb_topology_history(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
__dump_topology_ref_history(&mstb->topology_ref_history, mstb,
|
|
|
|
"MSTB");
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline void
|
|
|
|
drm_dp_mst_dump_port_topology_history(struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
__dump_topology_ref_history(&port->topology_ref_history, port,
|
|
|
|
"Port");
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline void
|
|
|
|
save_mstb_topology_ref(struct drm_dp_mst_branch *mstb,
|
|
|
|
enum drm_dp_mst_topology_ref_type type)
|
|
|
|
{
|
|
|
|
__topology_ref_save(mstb->mgr, &mstb->topology_ref_history, type);
|
|
|
|
}
|
|
|
|
|
|
|
|
static __always_inline void
|
|
|
|
save_port_topology_ref(struct drm_dp_mst_port *port,
|
|
|
|
enum drm_dp_mst_topology_ref_type type)
|
|
|
|
{
|
|
|
|
__topology_ref_save(port->mgr, &port->topology_ref_history, type);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
topology_ref_history_lock(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
mutex_lock(&mgr->topology_ref_history_lock);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
topology_ref_history_unlock(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
mutex_unlock(&mgr->topology_ref_history_lock);
|
|
|
|
}
|
|
|
|
#else
|
|
|
|
static inline void
|
|
|
|
topology_ref_history_lock(struct drm_dp_mst_topology_mgr *mgr) {}
|
|
|
|
static inline void
|
|
|
|
topology_ref_history_unlock(struct drm_dp_mst_topology_mgr *mgr) {}
|
|
|
|
static inline void
|
|
|
|
drm_dp_mst_dump_mstb_topology_history(struct drm_dp_mst_branch *mstb) {}
|
|
|
|
static inline void
|
|
|
|
drm_dp_mst_dump_port_topology_history(struct drm_dp_mst_port *port) {}
|
|
|
|
#define save_mstb_topology_ref(mstb, type)
|
|
|
|
#define save_port_topology_ref(port, type)
|
|
|
|
#endif
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static void drm_dp_destroy_mst_branch_device(struct kref *kref)
|
|
|
|
{
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
struct drm_dp_mst_branch *mstb =
|
|
|
|
container_of(kref, struct drm_dp_mst_branch, topology_kref);
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr = mstb->mgr;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
drm_dp_mst_dump_mstb_topology_history(mstb);
|
|
|
|
|
2019-02-20 06:41:02 +08:00
|
|
|
INIT_LIST_HEAD(&mstb->destroy_next);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-02-20 06:41:02 +08:00
|
|
|
/*
|
|
|
|
* This can get called under mgr->mutex, so we need to perform the
|
|
|
|
* actual destruction of the mstb in another worker
|
|
|
|
*/
|
|
|
|
mutex_lock(&mgr->delayed_destroy_lock);
|
|
|
|
list_add(&mstb->destroy_next, &mgr->destroy_branch_device_list);
|
|
|
|
mutex_unlock(&mgr->delayed_destroy_lock);
|
|
|
|
schedule_work(&mgr->delayed_destroy_work);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_try_get_mstb() - Increment the topology refcount of a
|
2019-02-02 09:23:26 +08:00
|
|
|
* branch device unless it's zero
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
* @mstb: &struct drm_dp_mst_branch to increment the topology refcount of
|
|
|
|
*
|
|
|
|
* Attempts to grab a topology reference to @mstb, if it hasn't yet been
|
|
|
|
* removed from the topology (e.g. &drm_dp_mst_branch.topology_kref has
|
|
|
|
* reached 0). Holding a topology reference implies that a malloc reference
|
|
|
|
* will be held to @mstb as long as the user holds the topology reference.
|
|
|
|
*
|
|
|
|
* Care should be taken to ensure that the user has at least one malloc
|
|
|
|
* reference to @mstb. If you already have a topology reference to @mstb, you
|
|
|
|
* should use drm_dp_mst_topology_get_mstb() instead.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_mst_topology_get_mstb()
|
|
|
|
* drm_dp_mst_topology_put_mstb()
|
|
|
|
*
|
|
|
|
* Returns:
|
|
|
|
* * 1: A topology reference was grabbed successfully
|
|
|
|
* * 0: @port is no longer in the topology, no reference was grabbed
|
|
|
|
*/
|
|
|
|
static int __must_check
|
|
|
|
drm_dp_mst_topology_try_get_mstb(struct drm_dp_mst_branch *mstb)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
int ret;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_lock(mstb->mgr);
|
|
|
|
ret = kref_get_unless_zero(&mstb->topology_kref);
|
|
|
|
if (ret) {
|
|
|
|
DRM_DEBUG("mstb %p (%d)\n",
|
|
|
|
mstb, kref_read(&mstb->topology_kref));
|
|
|
|
save_mstb_topology_ref(mstb, DRM_DP_MST_TOPOLOGY_REF_GET);
|
|
|
|
}
|
|
|
|
|
|
|
|
topology_ref_history_unlock(mstb->mgr);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
|
|
|
return ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_get_mstb() - Increment the topology refcount of a
|
|
|
|
* branch device
|
|
|
|
* @mstb: The &struct drm_dp_mst_branch to increment the topology refcount of
|
|
|
|
*
|
|
|
|
* Increments &drm_dp_mst_branch.topology_refcount without checking whether or
|
|
|
|
* not it's already reached 0. This is only valid to use in scenarios where
|
|
|
|
* you are already guaranteed to have at least one active topology reference
|
|
|
|
* to @mstb. Otherwise, drm_dp_mst_topology_try_get_mstb() must be used.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_mst_topology_try_get_mstb()
|
|
|
|
* drm_dp_mst_topology_put_mstb()
|
|
|
|
*/
|
|
|
|
static void drm_dp_mst_topology_get_mstb(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_lock(mstb->mgr);
|
|
|
|
|
|
|
|
save_mstb_topology_ref(mstb, DRM_DP_MST_TOPOLOGY_REF_GET);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
WARN_ON(kref_read(&mstb->topology_kref) == 0);
|
|
|
|
kref_get(&mstb->topology_kref);
|
|
|
|
DRM_DEBUG("mstb %p (%d)\n", mstb, kref_read(&mstb->topology_kref));
|
2019-06-21 05:59:25 +08:00
|
|
|
|
|
|
|
topology_ref_history_unlock(mstb->mgr);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_put_mstb() - release a topology reference to a branch
|
|
|
|
* device
|
|
|
|
* @mstb: The &struct drm_dp_mst_branch to release the topology reference from
|
|
|
|
*
|
|
|
|
* Releases a topology reference from @mstb by decrementing
|
|
|
|
* &drm_dp_mst_branch.topology_kref.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_mst_topology_try_get_mstb()
|
|
|
|
* drm_dp_mst_topology_get_mstb()
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
drm_dp_mst_topology_put_mstb(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_lock(mstb->mgr);
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
DRM_DEBUG("mstb %p (%d)\n",
|
|
|
|
mstb, kref_read(&mstb->topology_kref) - 1);
|
2019-06-21 05:59:25 +08:00
|
|
|
save_mstb_topology_ref(mstb, DRM_DP_MST_TOPOLOGY_REF_PUT);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_unlock(mstb->mgr);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
kref_put(&mstb->topology_kref, drm_dp_destroy_mst_branch_device);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_destroy_port(struct kref *kref)
|
|
|
|
{
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
struct drm_dp_mst_port *port =
|
|
|
|
container_of(kref, struct drm_dp_mst_port, topology_kref);
|
2014-06-05 12:01:32 +08:00
|
|
|
struct drm_dp_mst_topology_mgr *mgr = port->mgr;
|
2015-09-16 08:37:28 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
drm_dp_mst_dump_port_topology_history(port);
|
2015-06-15 08:34:28 +08:00
|
|
|
|
2019-03-11 06:59:10 +08:00
|
|
|
/* There's nothing that needs locking to destroy an input port yet */
|
|
|
|
if (port->input) {
|
|
|
|
drm_dp_mst_put_port_malloc(port);
|
|
|
|
return;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2019-03-11 06:59:10 +08:00
|
|
|
|
|
|
|
kfree(port->cached_edid);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* we can't destroy the connector here, as we might be holding the
|
|
|
|
* mode_config.mutex from an EDID retrieval
|
|
|
|
*/
|
|
|
|
mutex_lock(&mgr->delayed_destroy_lock);
|
|
|
|
list_add(&port->next, &mgr->destroy_port_list);
|
|
|
|
mutex_unlock(&mgr->delayed_destroy_lock);
|
|
|
|
schedule_work(&mgr->delayed_destroy_work);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_try_get_port() - Increment the topology refcount of a
|
2019-02-02 09:23:26 +08:00
|
|
|
* port unless it's zero
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
* @port: &struct drm_dp_mst_port to increment the topology refcount of
|
|
|
|
*
|
|
|
|
* Attempts to grab a topology reference to @port, if it hasn't yet been
|
|
|
|
* removed from the topology (e.g. &drm_dp_mst_port.topology_kref has reached
|
|
|
|
* 0). Holding a topology reference implies that a malloc reference will be
|
|
|
|
* held to @port as long as the user holds the topology reference.
|
|
|
|
*
|
|
|
|
* Care should be taken to ensure that the user has at least one malloc
|
|
|
|
* reference to @port. If you already have a topology reference to @port, you
|
|
|
|
* should use drm_dp_mst_topology_get_port() instead.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_mst_topology_get_port()
|
|
|
|
* drm_dp_mst_topology_put_port()
|
|
|
|
*
|
|
|
|
* Returns:
|
|
|
|
* * 1: A topology reference was grabbed successfully
|
|
|
|
* * 0: @port is no longer in the topology, no reference was grabbed
|
|
|
|
*/
|
|
|
|
static int __must_check
|
|
|
|
drm_dp_mst_topology_try_get_port(struct drm_dp_mst_port *port)
|
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
int ret;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_lock(port->mgr);
|
|
|
|
ret = kref_get_unless_zero(&port->topology_kref);
|
|
|
|
if (ret) {
|
|
|
|
DRM_DEBUG("port %p (%d)\n",
|
|
|
|
port, kref_read(&port->topology_kref));
|
|
|
|
save_port_topology_ref(port, DRM_DP_MST_TOPOLOGY_REF_GET);
|
|
|
|
}
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_unlock(port->mgr);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
return ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_get_port() - Increment the topology refcount of a port
|
|
|
|
* @port: The &struct drm_dp_mst_port to increment the topology refcount of
|
|
|
|
*
|
|
|
|
* Increments &drm_dp_mst_port.topology_refcount without checking whether or
|
|
|
|
* not it's already reached 0. This is only valid to use in scenarios where
|
|
|
|
* you are already guaranteed to have at least one active topology reference
|
|
|
|
* to @port. Otherwise, drm_dp_mst_topology_try_get_port() must be used.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_mst_topology_try_get_port()
|
|
|
|
* drm_dp_mst_topology_put_port()
|
|
|
|
*/
|
|
|
|
static void drm_dp_mst_topology_get_port(struct drm_dp_mst_port *port)
|
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_lock(port->mgr);
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
WARN_ON(kref_read(&port->topology_kref) == 0);
|
|
|
|
kref_get(&port->topology_kref);
|
|
|
|
DRM_DEBUG("port %p (%d)\n", port, kref_read(&port->topology_kref));
|
2019-06-21 05:59:25 +08:00
|
|
|
save_port_topology_ref(port, DRM_DP_MST_TOPOLOGY_REF_GET);
|
|
|
|
|
|
|
|
topology_ref_history_unlock(port->mgr);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_put_port() - release a topology reference to a port
|
|
|
|
* @port: The &struct drm_dp_mst_port to release the topology reference from
|
|
|
|
*
|
|
|
|
* Releases a topology reference from @port by decrementing
|
|
|
|
* &drm_dp_mst_port.topology_kref.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_mst_topology_try_get_port()
|
|
|
|
* drm_dp_mst_topology_get_port()
|
|
|
|
*/
|
2019-01-11 08:53:28 +08:00
|
|
|
static void drm_dp_mst_topology_put_port(struct drm_dp_mst_port *port)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-06-21 05:59:25 +08:00
|
|
|
topology_ref_history_lock(port->mgr);
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
DRM_DEBUG("port %p (%d)\n",
|
|
|
|
port, kref_read(&port->topology_kref) - 1);
|
2019-06-21 05:59:25 +08:00
|
|
|
save_port_topology_ref(port, DRM_DP_MST_TOPOLOGY_REF_PUT);
|
|
|
|
|
|
|
|
topology_ref_history_unlock(port->mgr);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
kref_put(&port->topology_kref, drm_dp_destroy_port);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
static struct drm_dp_mst_branch *
|
|
|
|
drm_dp_mst_topology_get_mstb_validated_locked(struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_mst_branch *to_find)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port;
|
|
|
|
struct drm_dp_mst_branch *rmstb;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
|
|
|
if (to_find == mstb)
|
2014-06-05 12:01:32 +08:00
|
|
|
return mstb;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
|
|
|
if (port->mstb) {
|
2019-01-11 08:53:28 +08:00
|
|
|
rmstb = drm_dp_mst_topology_get_mstb_validated_locked(
|
|
|
|
port->mstb, to_find);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (rmstb)
|
|
|
|
return rmstb;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
static struct drm_dp_mst_branch *
|
|
|
|
drm_dp_mst_topology_get_mstb_validated(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *rmstb = NULL;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
if (mgr->mst_primary) {
|
2019-01-11 08:53:28 +08:00
|
|
|
rmstb = drm_dp_mst_topology_get_mstb_validated_locked(
|
|
|
|
mgr->mst_primary, mstb);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
|
|
|
if (rmstb && !drm_dp_mst_topology_try_get_mstb(rmstb))
|
|
|
|
rmstb = NULL;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
return rmstb;
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
static struct drm_dp_mst_port *
|
|
|
|
drm_dp_mst_topology_get_port_validated_locked(struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_mst_port *to_find)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port, *mport;
|
|
|
|
|
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
if (port == to_find)
|
2014-06-05 12:01:32 +08:00
|
|
|
return port;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
if (port->mstb) {
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
mport = drm_dp_mst_topology_get_port_validated_locked(
|
|
|
|
port->mstb, to_find);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (mport)
|
|
|
|
return mport;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
static struct drm_dp_mst_port *
|
|
|
|
drm_dp_mst_topology_get_port_validated(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *rport = NULL;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
if (mgr->mst_primary) {
|
|
|
|
rport = drm_dp_mst_topology_get_port_validated_locked(
|
|
|
|
mgr->mst_primary, port);
|
|
|
|
|
|
|
|
if (rport && !drm_dp_mst_topology_try_get_port(rport))
|
|
|
|
rport = NULL;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
return rport;
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct drm_dp_mst_port *drm_dp_get_port(struct drm_dp_mst_branch *mstb, u8 port_num)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
int ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
|
|
|
if (port->port_num == port_num) {
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
ret = drm_dp_mst_topology_try_get_port(port);
|
|
|
|
return ret ? port : NULL;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* calculate a new RAD for this MST branch device
|
|
|
|
* if parent has an LCT of 2 then it has 1 nibble of RAD,
|
|
|
|
* if parent has an LCT of 3 then it has 2 nibbles of RAD,
|
|
|
|
*/
|
|
|
|
static u8 drm_dp_calculate_rad(struct drm_dp_mst_port *port,
|
|
|
|
u8 *rad)
|
|
|
|
{
|
2015-12-25 16:14:47 +08:00
|
|
|
int parent_lct = port->parent->lct;
|
2014-06-05 12:01:32 +08:00
|
|
|
int shift = 4;
|
2015-12-25 16:14:47 +08:00
|
|
|
int idx = (parent_lct - 1) / 2;
|
|
|
|
if (parent_lct > 1) {
|
|
|
|
memcpy(rad, port->parent->rad, idx + 1);
|
|
|
|
shift = (parent_lct % 2) ? 4 : 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
} else
|
|
|
|
rad[0] = 0;
|
|
|
|
|
|
|
|
rad[idx] |= port->port_num << shift;
|
2015-12-25 16:14:47 +08:00
|
|
|
return parent_lct + 1;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
static bool drm_dp_mst_is_dp_mst_end_device(u8 pdt, bool mcs)
|
|
|
|
{
|
|
|
|
switch (pdt) {
|
|
|
|
case DP_PEER_DEVICE_DP_LEGACY_CONV:
|
|
|
|
case DP_PEER_DEVICE_SST_SINK:
|
|
|
|
return true;
|
|
|
|
case DP_PEER_DEVICE_MST_BRANCHING:
|
|
|
|
/* For sst branch device */
|
|
|
|
if (!mcs)
|
|
|
|
return true;
|
|
|
|
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
return true;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
|
|
|
drm_dp_port_set_pdt(struct drm_dp_mst_port *port, u8 new_pdt,
|
|
|
|
bool new_mcs)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-03-14 01:51:41 +08:00
|
|
|
struct drm_dp_mst_topology_mgr *mgr = port->mgr;
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
u8 rad[8], lct;
|
|
|
|
int ret = 0;
|
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
if (port->pdt == new_pdt && port->mcs == new_mcs)
|
2019-03-14 01:51:41 +08:00
|
|
|
return 0;
|
|
|
|
|
|
|
|
/* Teardown the old pdt, if there is one */
|
2020-01-17 14:03:50 +08:00
|
|
|
if (port->pdt != DP_PEER_DEVICE_NONE) {
|
|
|
|
if (drm_dp_mst_is_dp_mst_end_device(port->pdt, port->mcs)) {
|
|
|
|
/*
|
|
|
|
* If the new PDT would also have an i2c bus,
|
|
|
|
* don't bother with reregistering it
|
|
|
|
*/
|
|
|
|
if (new_pdt != DP_PEER_DEVICE_NONE &&
|
|
|
|
drm_dp_mst_is_dp_mst_end_device(new_pdt, new_mcs)) {
|
|
|
|
port->pdt = new_pdt;
|
|
|
|
port->mcs = new_mcs;
|
|
|
|
return 0;
|
|
|
|
}
|
2019-03-14 01:51:41 +08:00
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
/* remove i2c over sideband */
|
|
|
|
drm_dp_mst_unregister_i2c_bus(&port->aux);
|
|
|
|
} else {
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
drm_dp_mst_topology_put_mstb(port->mstb);
|
|
|
|
port->mstb = NULL;
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
}
|
2019-03-14 01:51:41 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
port->pdt = new_pdt;
|
2020-01-17 14:03:50 +08:00
|
|
|
port->mcs = new_mcs;
|
2019-03-14 01:51:41 +08:00
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
if (port->pdt != DP_PEER_DEVICE_NONE) {
|
|
|
|
if (drm_dp_mst_is_dp_mst_end_device(port->pdt, port->mcs)) {
|
|
|
|
/* add i2c over sideband */
|
|
|
|
ret = drm_dp_mst_register_i2c_bus(&port->aux);
|
|
|
|
} else {
|
|
|
|
lct = drm_dp_calculate_rad(port, rad);
|
|
|
|
mstb = drm_dp_add_mst_branch_device(lct, rad);
|
|
|
|
if (!mstb) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
DRM_ERROR("Failed to create MSTB for port %p",
|
|
|
|
port);
|
|
|
|
goto out;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
port->mstb = mstb;
|
|
|
|
mstb->mgr = port->mgr;
|
|
|
|
mstb->port_parent = port;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
/*
|
|
|
|
* Make sure this port's memory allocation stays
|
|
|
|
* around until its child MSTB releases it
|
|
|
|
*/
|
|
|
|
drm_dp_mst_get_port_malloc(port);
|
|
|
|
mutex_unlock(&mgr->lock);
|
2019-03-14 01:51:41 +08:00
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
/* And make sure we send a link address for this */
|
|
|
|
ret = 1;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2019-03-14 01:51:41 +08:00
|
|
|
|
|
|
|
out:
|
|
|
|
if (ret < 0)
|
|
|
|
port->pdt = DP_PEER_DEVICE_NONE;
|
|
|
|
return ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_dpcd_read() - read a series of bytes from the DPCD via sideband
|
|
|
|
* @aux: Fake sideband AUX CH
|
|
|
|
* @offset: address of the (first) register to read
|
|
|
|
* @buffer: buffer to store the register values
|
|
|
|
* @size: number of bytes in @buffer
|
|
|
|
*
|
|
|
|
* Performs the same functionality for remote devices via
|
|
|
|
* sideband messaging as drm_dp_dpcd_read() does for local
|
|
|
|
* devices via actual AUX CH.
|
|
|
|
*
|
|
|
|
* Return: Number of bytes read, or negative error code on failure.
|
|
|
|
*/
|
|
|
|
ssize_t drm_dp_mst_dpcd_read(struct drm_dp_aux *aux,
|
|
|
|
unsigned int offset, void *buffer, size_t size)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port = container_of(aux, struct drm_dp_mst_port,
|
|
|
|
aux);
|
|
|
|
|
|
|
|
return drm_dp_send_dpcd_read(port->mgr, port,
|
|
|
|
offset, size, buffer);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_dpcd_write() - write a series of bytes to the DPCD via sideband
|
|
|
|
* @aux: Fake sideband AUX CH
|
|
|
|
* @offset: address of the (first) register to write
|
|
|
|
* @buffer: buffer containing the values to write
|
|
|
|
* @size: number of bytes in @buffer
|
|
|
|
*
|
|
|
|
* Performs the same functionality for remote devices via
|
|
|
|
* sideband messaging as drm_dp_dpcd_write() does for local
|
|
|
|
* devices via actual AUX CH.
|
|
|
|
*
|
|
|
|
* Return: 0 on success, negative error code on failure.
|
|
|
|
*/
|
|
|
|
ssize_t drm_dp_mst_dpcd_write(struct drm_dp_aux *aux,
|
|
|
|
unsigned int offset, void *buffer, size_t size)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port = container_of(aux, struct drm_dp_mst_port,
|
|
|
|
aux);
|
|
|
|
|
|
|
|
return drm_dp_send_dpcd_write(port->mgr, port,
|
|
|
|
offset, size, buffer);
|
|
|
|
}
|
|
|
|
|
2016-01-23 06:07:28 +08:00
|
|
|
static void drm_dp_check_mstb_guid(struct drm_dp_mst_branch *mstb, u8 *guid)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int ret;
|
2016-01-23 06:07:28 +08:00
|
|
|
|
|
|
|
memcpy(mstb->guid, guid, 16);
|
|
|
|
|
|
|
|
if (!drm_dp_validate_guid(mstb->mgr, mstb->guid)) {
|
|
|
|
if (mstb->port_parent) {
|
|
|
|
ret = drm_dp_send_dpcd_write(
|
|
|
|
mstb->mgr,
|
|
|
|
mstb->port_parent,
|
|
|
|
DP_GUID,
|
|
|
|
16,
|
|
|
|
mstb->guid);
|
|
|
|
} else {
|
|
|
|
|
|
|
|
ret = drm_dp_dpcd_write(
|
|
|
|
mstb->mgr->aux,
|
|
|
|
DP_GUID,
|
|
|
|
mstb->guid,
|
|
|
|
16);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2015-09-16 09:04:49 +08:00
|
|
|
static void build_mst_prop_path(const struct drm_dp_mst_branch *mstb,
|
|
|
|
int pnum,
|
2014-10-12 06:02:32 +08:00
|
|
|
char *proppath,
|
|
|
|
size_t proppath_size)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
char temp[8];
|
2014-10-12 06:02:32 +08:00
|
|
|
snprintf(proppath, proppath_size, "mst:%d", mstb->mgr->conn_base_id);
|
2014-06-05 12:01:32 +08:00
|
|
|
for (i = 0; i < (mstb->lct - 1); i++) {
|
|
|
|
int shift = (i % 2) ? 0 : 4;
|
2015-12-25 16:14:48 +08:00
|
|
|
int port_num = (mstb->rad[i / 2] >> shift) & 0xf;
|
2014-10-12 06:02:32 +08:00
|
|
|
snprintf(temp, sizeof(temp), "-%d", port_num);
|
|
|
|
strlcat(proppath, temp, proppath_size);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2015-09-16 09:04:49 +08:00
|
|
|
snprintf(temp, sizeof(temp), "-%d", pnum);
|
2014-10-12 06:02:32 +08:00
|
|
|
strlcat(proppath, temp, proppath_size);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_connector_late_register() - Late MST connector registration
|
2019-07-26 22:20:47 +08:00
|
|
|
* @connector: The MST connector
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
* @port: The MST port for this connector
|
|
|
|
*
|
|
|
|
* Helper to register the remote aux device for this MST port. Drivers should
|
|
|
|
* call this from their mst connector's late_register hook to enable MST aux
|
|
|
|
* devices.
|
|
|
|
*
|
|
|
|
* Return: 0 on success, negative error code on failure.
|
|
|
|
*/
|
|
|
|
int drm_dp_mst_connector_late_register(struct drm_connector *connector,
|
|
|
|
struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
DRM_DEBUG_KMS("registering %s remote bus for %s\n",
|
|
|
|
port->aux.name, connector->kdev->kobj.name);
|
|
|
|
|
|
|
|
port->aux.dev = connector->kdev;
|
|
|
|
return drm_dp_aux_register_devnode(&port->aux);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_connector_late_register);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_connector_early_unregister() - Early MST connector unregistration
|
2019-07-26 22:20:47 +08:00
|
|
|
* @connector: The MST connector
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
* @port: The MST port for this connector
|
|
|
|
*
|
|
|
|
* Helper to unregister the remote aux device for this MST port, registered by
|
|
|
|
* drm_dp_mst_connector_late_register(). Drivers should call this from their mst
|
|
|
|
* connector's early_unregister hook.
|
|
|
|
*/
|
|
|
|
void drm_dp_mst_connector_early_unregister(struct drm_connector *connector,
|
|
|
|
struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
DRM_DEBUG_KMS("unregistering %s remote bus for %s\n",
|
|
|
|
port->aux.name, connector->kdev->kobj.name);
|
|
|
|
drm_dp_aux_unregister_devnode(&port->aux);
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_connector_early_unregister);
|
|
|
|
|
2019-09-04 04:45:55 +08:00
|
|
|
static void
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_dp_mst_port_add_connector(struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr = port->mgr;
|
|
|
|
char proppath[255];
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
build_mst_prop_path(mstb, port->port_num, proppath, sizeof(proppath));
|
|
|
|
port->connector = mgr->cbs->add_connector(mgr, port, proppath);
|
|
|
|
if (!port->connector) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto error;
|
|
|
|
}
|
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
if (port->pdt != DP_PEER_DEVICE_NONE &&
|
|
|
|
drm_dp_mst_is_dp_mst_end_device(port->pdt, port->mcs)) {
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
port->cached_edid = drm_get_edid(port->connector,
|
|
|
|
&port->aux.ddc);
|
|
|
|
drm_connector_set_tile_property(port->connector);
|
|
|
|
}
|
|
|
|
|
|
|
|
mgr->cbs->register_connector(port->connector);
|
|
|
|
return;
|
|
|
|
|
|
|
|
error:
|
|
|
|
DRM_ERROR("Failed to create connector for port %p: %d\n", port, ret);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Drop a topology reference, and unlink the port from the in-memory topology
|
|
|
|
* layout
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
drm_dp_mst_topology_unlink_port(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
list_del(&port->next);
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
drm_dp_mst_topology_put_port(port);
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct drm_dp_mst_port *
|
|
|
|
drm_dp_mst_add_port(struct drm_device *dev,
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb, u8 port_number)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port = kzalloc(sizeof(*port), GFP_KERNEL);
|
|
|
|
|
|
|
|
if (!port)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
kref_init(&port->topology_kref);
|
|
|
|
kref_init(&port->malloc_kref);
|
|
|
|
port->parent = mstb;
|
|
|
|
port->port_num = port_number;
|
|
|
|
port->mgr = mgr;
|
|
|
|
port->aux.name = "DPMST";
|
|
|
|
port->aux.dev = dev->dev;
|
|
|
|
port->aux.is_remote = true;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Make sure the memory allocation for our parent branch stays
|
|
|
|
* around until our own memory allocation is released
|
|
|
|
*/
|
|
|
|
drm_dp_mst_get_mstb_malloc(mstb);
|
|
|
|
|
|
|
|
return port;
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
static int
|
2019-09-04 04:45:55 +08:00
|
|
|
drm_dp_mst_handle_link_address_port(struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_device *dev,
|
|
|
|
struct drm_dp_link_addr_reply_port *port_msg)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-03-14 01:51:41 +08:00
|
|
|
struct drm_dp_mst_topology_mgr *mgr = mstb->mgr;
|
2014-06-05 12:01:32 +08:00
|
|
|
struct drm_dp_mst_port *port;
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
int old_ddps = 0, ret;
|
|
|
|
u8 new_pdt = DP_PEER_DEVICE_NONE;
|
2020-01-17 14:03:50 +08:00
|
|
|
bool new_mcs = 0;
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
bool created = false, send_link_addr = false, changed = false;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
port = drm_dp_get_port(mstb, port_msg->port_number);
|
|
|
|
if (!port) {
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
port = drm_dp_mst_add_port(dev, mgr, mstb,
|
|
|
|
port_msg->port_number);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!port)
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
return -ENOMEM;
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
created = true;
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
changed = true;
|
|
|
|
} else if (!port->input && port_msg->input_port && port->connector) {
|
|
|
|
/* Since port->connector can't be changed here, we create a
|
|
|
|
* new port if input_port changes from 0 to 1
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
*/
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_dp_mst_topology_unlink_port(mgr, port);
|
|
|
|
drm_dp_mst_topology_put_port(port);
|
|
|
|
port = drm_dp_mst_add_port(dev, mgr, mstb,
|
|
|
|
port_msg->port_number);
|
|
|
|
if (!port)
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
changed = true;
|
2014-06-05 12:01:32 +08:00
|
|
|
created = true;
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
} else if (port->input && !port_msg->input_port) {
|
|
|
|
changed = true;
|
|
|
|
} else if (port->connector) {
|
|
|
|
/* We're updating a port that's exposed to userspace, so do it
|
|
|
|
* under lock
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
*/
|
|
|
|
drm_modeset_lock(&mgr->base.lock, NULL);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
old_ddps = port->ddps;
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
changed = port->ddps != port_msg->ddps ||
|
|
|
|
(port->ddps &&
|
|
|
|
(port->ldps != port_msg->legacy_device_plug_status ||
|
|
|
|
port->dpcd_rev != port_msg->dpcd_revision ||
|
|
|
|
port->mcs != port_msg->mcs ||
|
|
|
|
port->pdt != port_msg->peer_device_type ||
|
|
|
|
port->num_sdp_stream_sinks !=
|
|
|
|
port_msg->num_sdp_stream_sinks));
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
port->input = port_msg->input_port;
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
if (!port->input)
|
|
|
|
new_pdt = port_msg->peer_device_type;
|
2020-01-17 14:03:50 +08:00
|
|
|
new_mcs = port_msg->mcs;
|
2014-06-05 12:01:32 +08:00
|
|
|
port->ddps = port_msg->ddps;
|
|
|
|
port->ldps = port_msg->legacy_device_plug_status;
|
|
|
|
port->dpcd_rev = port_msg->dpcd_revision;
|
|
|
|
port->num_sdp_streams = port_msg->num_sdp_streams;
|
|
|
|
port->num_sdp_stream_sinks = port_msg->num_sdp_stream_sinks;
|
|
|
|
|
|
|
|
/* manage mstb port lists with mgr lock - take a reference
|
|
|
|
for this list */
|
|
|
|
if (created) {
|
2019-03-14 01:51:41 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
drm_dp_mst_topology_get_port(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
list_add(&port->next, &mstb->ports);
|
2019-03-14 01:51:41 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
if (old_ddps != port->ddps) {
|
|
|
|
if (port->ddps) {
|
2019-01-11 08:53:24 +08:00
|
|
|
if (!port->input) {
|
2019-03-14 01:51:41 +08:00
|
|
|
drm_dp_send_enum_path_resources(mgr, mstb,
|
|
|
|
port);
|
2019-01-11 08:53:24 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
} else {
|
|
|
|
port->available_pbn = 0;
|
2019-01-11 08:53:24 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
ret = drm_dp_port_set_pdt(port, new_pdt, new_mcs);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
if (ret == 1) {
|
|
|
|
send_link_addr = true;
|
|
|
|
} else if (ret < 0) {
|
|
|
|
DRM_ERROR("Failed to change PDT on port %p: %d\n",
|
|
|
|
port, ret);
|
|
|
|
goto fail;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
/*
|
|
|
|
* If this port wasn't just created, then we're reprobing because
|
|
|
|
* we're coming out of suspend. In this case, always resend the link
|
|
|
|
* address if there's an MSTB on this port
|
|
|
|
*/
|
2020-01-17 14:03:50 +08:00
|
|
|
if (!created && port->pdt == DP_PEER_DEVICE_MST_BRANCHING &&
|
|
|
|
port->mcs)
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
send_link_addr = true;
|
|
|
|
|
|
|
|
if (port->connector)
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_modeset_unlock(&mgr->base.lock);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
else if (!port->input)
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_dp_mst_port_add_connector(mstb, port);
|
2019-03-14 01:51:41 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
if (send_link_addr && port->mstb) {
|
|
|
|
ret = drm_dp_send_link_address(mgr, port->mstb);
|
|
|
|
if (ret == 1) /* MSTB below us changed */
|
|
|
|
changed = true;
|
|
|
|
else if (ret < 0)
|
|
|
|
goto fail_put;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2016-02-17 09:36:38 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* put reference to this port */
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
return changed;
|
2019-03-14 01:51:41 +08:00
|
|
|
|
|
|
|
fail:
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_dp_mst_topology_unlink_port(mgr, port);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
if (port->connector)
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_modeset_unlock(&mgr->base.lock);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
fail_put:
|
|
|
|
drm_dp_mst_topology_put_port(port);
|
|
|
|
return ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2019-09-04 04:45:55 +08:00
|
|
|
static void
|
|
|
|
drm_dp_mst_handle_conn_stat(struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_connection_status_notify *conn_stat)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
struct drm_dp_mst_topology_mgr *mgr = mstb->mgr;
|
2014-06-05 12:01:32 +08:00
|
|
|
struct drm_dp_mst_port *port;
|
drm/dp_mst: clear time slots for ports invalid
[Why]
When change the connection status in a MST topology, mst device
which detect the event will send out CONNECTION_STATUS_NOTIFY messgae.
e.g. src-mst-mst-sst => src-mst (unplug) mst-sst
Currently, under the above case of unplugging device, ports which have
been allocated payloads and are no longer in the topology still occupy
time slots and recorded in proposed_vcpi[] of topology manager.
If we don't clean up the proposed_vcpi[], when code flow goes to try to
update payload table by calling drm_dp_update_payload_part1(), we will
fail at checking port validation due to there are ports with proposed
time slots but no longer in the mst topology. As the result of that, we
will also stop updating the DPCD payload table of down stream port.
[How]
While handling the CONNECTION_STATUS_NOTIFY message, add a detection to
see if the event indicates that a device is unplugged to an output port.
If the detection is true, then iterrate over all proposed_vcpi[] to
see whether a port of the proposed_vcpi[] is still in the topology or
not. If the port is invalid, set its num_slots to 0.
Thereafter, when try to update payload table by calling
drm_dp_update_payload_part1(), we can successfully update the DPCD
payload table of down stream port and clear the proposed_vcpi[] to NULL.
Changes since v1:(https://patchwork.kernel.org/patch/11275801/)
* Invert the conditional to reduce the indenting
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Wayne Lin <Wayne.Lin@amd.com>
Signed-off-by: Lyude Paul <lyude@redhat.com>
[removed cc for stable - there's too many patches this depends on for
this to backport cleanly]
Link: https://patchwork.freedesktop.org/patch/msgid/20200106102158.28261-1-Wayne.Lin@amd.com
2020-01-06 18:21:58 +08:00
|
|
|
int old_ddps, old_input, ret, i;
|
2019-10-17 05:50:15 +08:00
|
|
|
u8 new_pdt;
|
2020-01-17 14:03:50 +08:00
|
|
|
bool new_mcs;
|
2019-10-17 05:50:15 +08:00
|
|
|
bool dowork = false, create_connector = false;
|
2019-03-14 01:51:41 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
port = drm_dp_get_port(mstb, conn_stat->port_number);
|
|
|
|
if (!port)
|
|
|
|
return;
|
|
|
|
|
2019-10-17 05:50:15 +08:00
|
|
|
if (port->connector) {
|
|
|
|
if (!port->input && conn_stat->input_port) {
|
|
|
|
/*
|
|
|
|
* We can't remove a connector from an already exposed
|
|
|
|
* port, so just throw the port out and make sure we
|
|
|
|
* reprobe the link address of it's parent MSTB
|
|
|
|
*/
|
|
|
|
drm_dp_mst_topology_unlink_port(mgr, port);
|
|
|
|
mstb->link_address_sent = false;
|
|
|
|
dowork = true;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Locking is only needed if the port's exposed to userspace */
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_modeset_lock(&mgr->base.lock, NULL);
|
2019-10-17 05:50:15 +08:00
|
|
|
} else if (port->input && !conn_stat->input_port) {
|
|
|
|
create_connector = true;
|
|
|
|
/* Reprobe link address so we get num_sdp_streams */
|
|
|
|
mstb->link_address_sent = false;
|
|
|
|
dowork = true;
|
|
|
|
}
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
old_ddps = port->ddps;
|
drm/dp_mst: clear time slots for ports invalid
[Why]
When change the connection status in a MST topology, mst device
which detect the event will send out CONNECTION_STATUS_NOTIFY messgae.
e.g. src-mst-mst-sst => src-mst (unplug) mst-sst
Currently, under the above case of unplugging device, ports which have
been allocated payloads and are no longer in the topology still occupy
time slots and recorded in proposed_vcpi[] of topology manager.
If we don't clean up the proposed_vcpi[], when code flow goes to try to
update payload table by calling drm_dp_update_payload_part1(), we will
fail at checking port validation due to there are ports with proposed
time slots but no longer in the mst topology. As the result of that, we
will also stop updating the DPCD payload table of down stream port.
[How]
While handling the CONNECTION_STATUS_NOTIFY message, add a detection to
see if the event indicates that a device is unplugged to an output port.
If the detection is true, then iterrate over all proposed_vcpi[] to
see whether a port of the proposed_vcpi[] is still in the topology or
not. If the port is invalid, set its num_slots to 0.
Thereafter, when try to update payload table by calling
drm_dp_update_payload_part1(), we can successfully update the DPCD
payload table of down stream port and clear the proposed_vcpi[] to NULL.
Changes since v1:(https://patchwork.kernel.org/patch/11275801/)
* Invert the conditional to reduce the indenting
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Wayne Lin <Wayne.Lin@amd.com>
Signed-off-by: Lyude Paul <lyude@redhat.com>
[removed cc for stable - there's too many patches this depends on for
this to backport cleanly]
Link: https://patchwork.freedesktop.org/patch/msgid/20200106102158.28261-1-Wayne.Lin@amd.com
2020-01-06 18:21:58 +08:00
|
|
|
old_input = port->input;
|
2019-10-17 05:50:15 +08:00
|
|
|
port->input = conn_stat->input_port;
|
2014-06-05 12:01:32 +08:00
|
|
|
port->ldps = conn_stat->legacy_device_plug_status;
|
|
|
|
port->ddps = conn_stat->displayport_device_plug_status;
|
|
|
|
|
|
|
|
if (old_ddps != port->ddps) {
|
|
|
|
if (port->ddps) {
|
2016-02-17 09:36:38 +08:00
|
|
|
dowork = true;
|
2014-06-05 12:01:32 +08:00
|
|
|
} else {
|
|
|
|
port->available_pbn = 0;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2019-10-17 05:50:15 +08:00
|
|
|
new_pdt = port->input ? DP_PEER_DEVICE_NONE : conn_stat->peer_device_type;
|
2020-01-17 14:03:50 +08:00
|
|
|
new_mcs = conn_stat->message_capability_status;
|
|
|
|
ret = drm_dp_port_set_pdt(port, new_pdt, new_mcs);
|
2019-10-17 05:50:15 +08:00
|
|
|
if (ret == 1) {
|
|
|
|
dowork = true;
|
|
|
|
} else if (ret < 0) {
|
|
|
|
DRM_ERROR("Failed to change PDT for port %p: %d\n",
|
|
|
|
port, ret);
|
|
|
|
dowork = false;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: clear time slots for ports invalid
[Why]
When change the connection status in a MST topology, mst device
which detect the event will send out CONNECTION_STATUS_NOTIFY messgae.
e.g. src-mst-mst-sst => src-mst (unplug) mst-sst
Currently, under the above case of unplugging device, ports which have
been allocated payloads and are no longer in the topology still occupy
time slots and recorded in proposed_vcpi[] of topology manager.
If we don't clean up the proposed_vcpi[], when code flow goes to try to
update payload table by calling drm_dp_update_payload_part1(), we will
fail at checking port validation due to there are ports with proposed
time slots but no longer in the mst topology. As the result of that, we
will also stop updating the DPCD payload table of down stream port.
[How]
While handling the CONNECTION_STATUS_NOTIFY message, add a detection to
see if the event indicates that a device is unplugged to an output port.
If the detection is true, then iterrate over all proposed_vcpi[] to
see whether a port of the proposed_vcpi[] is still in the topology or
not. If the port is invalid, set its num_slots to 0.
Thereafter, when try to update payload table by calling
drm_dp_update_payload_part1(), we can successfully update the DPCD
payload table of down stream port and clear the proposed_vcpi[] to NULL.
Changes since v1:(https://patchwork.kernel.org/patch/11275801/)
* Invert the conditional to reduce the indenting
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Wayne Lin <Wayne.Lin@amd.com>
Signed-off-by: Lyude Paul <lyude@redhat.com>
[removed cc for stable - there's too many patches this depends on for
this to backport cleanly]
Link: https://patchwork.freedesktop.org/patch/msgid/20200106102158.28261-1-Wayne.Lin@amd.com
2020-01-06 18:21:58 +08:00
|
|
|
if (!old_input && old_ddps != port->ddps && !port->ddps) {
|
|
|
|
for (i = 0; i < mgr->max_payloads; i++) {
|
|
|
|
struct drm_dp_vcpi *vcpi = mgr->proposed_vcpis[i];
|
|
|
|
struct drm_dp_mst_port *port_validated;
|
|
|
|
|
|
|
|
if (!vcpi)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
port_validated =
|
|
|
|
container_of(vcpi, struct drm_dp_mst_port, vcpi);
|
|
|
|
port_validated =
|
|
|
|
drm_dp_mst_topology_get_port_validated(mgr, port_validated);
|
|
|
|
if (!port_validated) {
|
|
|
|
mutex_lock(&mgr->payload_lock);
|
|
|
|
vcpi->num_slots = 0;
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
} else {
|
|
|
|
drm_dp_mst_topology_put_port(port_validated);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
if (port->connector)
|
|
|
|
drm_modeset_unlock(&mgr->base.lock);
|
2019-10-17 05:50:15 +08:00
|
|
|
else if (create_connector)
|
|
|
|
drm_dp_mst_port_add_connector(mstb, port);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
|
2019-10-17 05:50:15 +08:00
|
|
|
out:
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (dowork)
|
|
|
|
queue_work(system_long_wq, &mstb->mgr->work);
|
|
|
|
}
|
|
|
|
|
|
|
|
static struct drm_dp_mst_branch *drm_dp_get_mst_branch_device(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
u8 lct, u8 *rad)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
struct drm_dp_mst_port *port;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
int i, ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
/* find the port by iterating down */
|
2015-06-22 12:40:44 +08:00
|
|
|
|
|
|
|
mutex_lock(&mgr->lock);
|
2014-06-05 12:01:32 +08:00
|
|
|
mstb = mgr->mst_primary;
|
|
|
|
|
2018-11-09 17:00:12 +08:00
|
|
|
if (!mstb)
|
|
|
|
goto out;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
for (i = 0; i < lct - 1; i++) {
|
|
|
|
int shift = (i % 2) ? 0 : 4;
|
2015-12-25 16:14:48 +08:00
|
|
|
int port_num = (rad[i / 2] >> shift) & 0xf;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
|
|
|
if (port->port_num == port_num) {
|
2015-10-16 18:33:02 +08:00
|
|
|
mstb = port->mstb;
|
|
|
|
if (!mstb) {
|
2014-06-05 12:01:32 +08:00
|
|
|
DRM_ERROR("failed to lookup MSTB with lct %d, rad %02x\n", lct, rad[0]);
|
2015-10-16 18:33:02 +08:00
|
|
|
goto out;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
ret = drm_dp_mst_topology_try_get_mstb(mstb);
|
|
|
|
if (!ret)
|
|
|
|
mstb = NULL;
|
2015-10-16 18:33:02 +08:00
|
|
|
out:
|
2015-06-22 12:40:44 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
2014-06-05 12:01:32 +08:00
|
|
|
return mstb;
|
|
|
|
}
|
|
|
|
|
2015-12-19 06:14:42 +08:00
|
|
|
static struct drm_dp_mst_branch *get_mst_branch_device_by_guid_helper(
|
|
|
|
struct drm_dp_mst_branch *mstb,
|
2019-09-04 04:45:49 +08:00
|
|
|
const uint8_t *guid)
|
2015-12-19 06:14:42 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *found_mstb;
|
|
|
|
struct drm_dp_mst_port *port;
|
|
|
|
|
2016-01-23 06:07:28 +08:00
|
|
|
if (memcmp(mstb->guid, guid, 16) == 0)
|
|
|
|
return mstb;
|
|
|
|
|
|
|
|
|
2015-12-19 06:14:42 +08:00
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
|
|
|
if (!port->mstb)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
found_mstb = get_mst_branch_device_by_guid_helper(port->mstb, guid);
|
|
|
|
|
|
|
|
if (found_mstb)
|
|
|
|
return found_mstb;
|
|
|
|
}
|
|
|
|
|
|
|
|
return NULL;
|
|
|
|
}
|
|
|
|
|
2019-01-11 08:53:25 +08:00
|
|
|
static struct drm_dp_mst_branch *
|
|
|
|
drm_dp_get_mst_branch_device_by_guid(struct drm_dp_mst_topology_mgr *mgr,
|
2019-09-04 04:45:49 +08:00
|
|
|
const uint8_t *guid)
|
2015-12-19 06:14:42 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
int ret;
|
2015-12-19 06:14:42 +08:00
|
|
|
|
|
|
|
/* find the port by iterating down */
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
|
2016-01-23 06:07:28 +08:00
|
|
|
mstb = get_mst_branch_device_by_guid_helper(mgr->mst_primary, guid);
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
if (mstb) {
|
|
|
|
ret = drm_dp_mst_topology_try_get_mstb(mstb);
|
|
|
|
if (!ret)
|
|
|
|
mstb = NULL;
|
|
|
|
}
|
2015-12-19 06:14:42 +08:00
|
|
|
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
return mstb;
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
static int drm_dp_check_and_send_link_address(struct drm_dp_mst_topology_mgr *mgr,
|
2014-06-05 12:01:32 +08:00
|
|
|
struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port;
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
int ret;
|
|
|
|
bool changed = false;
|
2019-10-17 04:02:59 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
if (!mstb->link_address_sent) {
|
|
|
|
ret = drm_dp_send_link_address(mgr, mstb);
|
|
|
|
if (ret == 1)
|
|
|
|
changed = true;
|
|
|
|
else if (ret < 0)
|
|
|
|
return ret;
|
|
|
|
}
|
2015-09-06 16:53:00 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
2019-10-17 04:02:59 +08:00
|
|
|
struct drm_dp_mst_branch *mstb_child = NULL;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-17 04:02:59 +08:00
|
|
|
if (port->input || !port->ddps)
|
2014-06-05 12:01:32 +08:00
|
|
|
continue;
|
|
|
|
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
if (!port->available_pbn) {
|
|
|
|
drm_modeset_lock(&mgr->base.lock, NULL);
|
2014-06-05 12:01:32 +08:00
|
|
|
drm_dp_send_enum_path_resources(mgr, mstb, port);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_modeset_unlock(&mgr->base.lock);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
changed = true;
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-17 04:02:59 +08:00
|
|
|
if (port->mstb)
|
2019-01-11 08:53:28 +08:00
|
|
|
mstb_child = drm_dp_mst_topology_get_mstb_validated(
|
|
|
|
mgr, port->mstb);
|
2019-10-17 04:02:59 +08:00
|
|
|
|
|
|
|
if (mstb_child) {
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
ret = drm_dp_check_and_send_link_address(mgr,
|
|
|
|
mstb_child);
|
2019-10-17 04:02:59 +08:00
|
|
|
drm_dp_mst_topology_put_mstb(mstb_child);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
if (ret == 1)
|
|
|
|
changed = true;
|
|
|
|
else if (ret < 0)
|
|
|
|
return ret;
|
2015-06-22 15:31:59 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
|
|
|
|
return changed;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_mst_link_probe_work(struct work_struct *work)
|
|
|
|
{
|
2019-10-17 04:02:59 +08:00
|
|
|
struct drm_dp_mst_topology_mgr *mgr =
|
|
|
|
container_of(work, struct drm_dp_mst_topology_mgr, work);
|
|
|
|
struct drm_device *dev = mgr->dev;
|
2015-06-22 15:31:59 +08:00
|
|
|
struct drm_dp_mst_branch *mstb;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
int ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-17 04:02:59 +08:00
|
|
|
mutex_lock(&mgr->probe_lock);
|
|
|
|
|
2015-06-22 15:31:59 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
mstb = mgr->mst_primary;
|
|
|
|
if (mstb) {
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
ret = drm_dp_mst_topology_try_get_mstb(mstb);
|
|
|
|
if (!ret)
|
|
|
|
mstb = NULL;
|
2015-06-22 15:31:59 +08:00
|
|
|
}
|
|
|
|
mutex_unlock(&mgr->lock);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
if (!mstb) {
|
|
|
|
mutex_unlock(&mgr->probe_lock);
|
|
|
|
return;
|
2015-06-22 15:31:59 +08:00
|
|
|
}
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
ret = drm_dp_check_and_send_link_address(mgr, mstb);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
|
|
|
|
2019-10-17 04:02:59 +08:00
|
|
|
mutex_unlock(&mgr->probe_lock);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
if (ret)
|
|
|
|
drm_kms_helper_hotplug_event(dev);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static bool drm_dp_validate_guid(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
u8 *guid)
|
|
|
|
{
|
2017-07-12 23:52:54 +08:00
|
|
|
u64 salt;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2017-07-12 23:52:54 +08:00
|
|
|
if (memchr_inv(guid, 0, 16))
|
|
|
|
return true;
|
|
|
|
|
|
|
|
salt = get_jiffies_64();
|
|
|
|
|
|
|
|
memcpy(&guid[0], &salt, sizeof(u64));
|
|
|
|
memcpy(&guid[8], &salt, sizeof(u64));
|
|
|
|
|
|
|
|
return false;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static int build_dpcd_read(struct drm_dp_sideband_msg_tx *msg, u8 port_num, u32 offset, u8 num_bytes)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_req_body req;
|
|
|
|
|
|
|
|
req.req_type = DP_REMOTE_DPCD_READ;
|
|
|
|
req.u.dpcd_read.port_number = port_num;
|
|
|
|
req.u.dpcd_read.dpcd_address = offset;
|
|
|
|
req.u.dpcd_read.num_bytes = num_bytes;
|
|
|
|
drm_dp_encode_sideband_req(&req, msg);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int drm_dp_send_sideband_msg(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
bool up, u8 *msg, int len)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
int regbase = up ? DP_SIDEBAND_MSG_UP_REP_BASE : DP_SIDEBAND_MSG_DOWN_REQ_BASE;
|
|
|
|
int tosend, total, offset;
|
|
|
|
int retries = 0;
|
|
|
|
|
|
|
|
retry:
|
|
|
|
total = len;
|
|
|
|
offset = 0;
|
|
|
|
do {
|
|
|
|
tosend = min3(mgr->max_dpcd_transaction_bytes, 16, total);
|
|
|
|
|
|
|
|
ret = drm_dp_dpcd_write(mgr->aux, regbase + offset,
|
|
|
|
&msg[offset],
|
|
|
|
tosend);
|
|
|
|
if (ret != tosend) {
|
|
|
|
if (ret == -EIO && retries < 5) {
|
|
|
|
retries++;
|
|
|
|
goto retry;
|
|
|
|
}
|
|
|
|
DRM_DEBUG_KMS("failed to dpcd write %d %d\n", tosend, ret);
|
|
|
|
|
|
|
|
return -EIO;
|
|
|
|
}
|
|
|
|
offset += tosend;
|
|
|
|
total -= tosend;
|
|
|
|
} while (total > 0);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int set_hdr_from_dst_qlock(struct drm_dp_sideband_msg_hdr *hdr,
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *mstb = txmsg->dst;
|
2015-12-19 06:14:42 +08:00
|
|
|
u8 req_type;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
/* both msg slots are full */
|
|
|
|
if (txmsg->seqno == -1) {
|
|
|
|
if (mstb->tx_slots[0] && mstb->tx_slots[1]) {
|
|
|
|
DRM_DEBUG_KMS("%s: failed to find slot\n", __func__);
|
|
|
|
return -EAGAIN;
|
|
|
|
}
|
|
|
|
if (mstb->tx_slots[0] == NULL && mstb->tx_slots[1] == NULL) {
|
|
|
|
txmsg->seqno = mstb->last_seqno;
|
|
|
|
mstb->last_seqno ^= 1;
|
|
|
|
} else if (mstb->tx_slots[0] == NULL)
|
|
|
|
txmsg->seqno = 0;
|
|
|
|
else
|
|
|
|
txmsg->seqno = 1;
|
|
|
|
mstb->tx_slots[txmsg->seqno] = txmsg;
|
|
|
|
}
|
2015-12-19 06:14:42 +08:00
|
|
|
|
|
|
|
req_type = txmsg->msg[0] & 0x7f;
|
|
|
|
if (req_type == DP_CONNECTION_STATUS_NOTIFY ||
|
|
|
|
req_type == DP_RESOURCE_STATUS_NOTIFY)
|
|
|
|
hdr->broadcast = 1;
|
|
|
|
else
|
|
|
|
hdr->broadcast = 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
hdr->path_msg = txmsg->path_msg;
|
|
|
|
hdr->lct = mstb->lct;
|
|
|
|
hdr->lcr = mstb->lct - 1;
|
|
|
|
if (mstb->lct > 1)
|
|
|
|
memcpy(hdr->rad, mstb->rad, mstb->lct / 2);
|
|
|
|
hdr->seqno = txmsg->seqno;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
/*
|
|
|
|
* process a single block of the next message in the sideband queue
|
|
|
|
*/
|
|
|
|
static int process_single_tx_qlock(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg,
|
|
|
|
bool up)
|
|
|
|
{
|
|
|
|
u8 chunk[48];
|
|
|
|
struct drm_dp_sideband_msg_hdr hdr;
|
|
|
|
int len, space, idx, tosend;
|
|
|
|
int ret;
|
|
|
|
|
2014-07-14 19:13:18 +08:00
|
|
|
memset(&hdr, 0, sizeof(struct drm_dp_sideband_msg_hdr));
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
if (txmsg->state == DRM_DP_SIDEBAND_TX_QUEUED) {
|
|
|
|
txmsg->seqno = -1;
|
|
|
|
txmsg->state = DRM_DP_SIDEBAND_TX_START_SEND;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* make hdr from dst mst - for replies use seqno
|
|
|
|
otherwise assign one */
|
|
|
|
ret = set_hdr_from_dst_qlock(&hdr, txmsg);
|
|
|
|
if (ret < 0)
|
|
|
|
return ret;
|
|
|
|
|
|
|
|
/* amount left to send in this message */
|
|
|
|
len = txmsg->cur_len - txmsg->cur_offset;
|
|
|
|
|
|
|
|
/* 48 - sideband msg size - 1 byte for data CRC, x header bytes */
|
|
|
|
space = 48 - 1 - drm_dp_calc_sb_hdr_size(&hdr);
|
|
|
|
|
|
|
|
tosend = min(len, space);
|
|
|
|
if (len == txmsg->cur_len)
|
|
|
|
hdr.somt = 1;
|
|
|
|
if (space >= len)
|
|
|
|
hdr.eomt = 1;
|
|
|
|
|
|
|
|
|
|
|
|
hdr.msg_len = tosend + 1;
|
|
|
|
drm_dp_encode_sideband_msg_hdr(&hdr, chunk, &idx);
|
|
|
|
memcpy(&chunk[idx], &txmsg->msg[txmsg->cur_offset], tosend);
|
|
|
|
/* add crc at end */
|
|
|
|
drm_dp_crc_sideband_chunk_req(&chunk[idx], tosend);
|
|
|
|
idx += tosend + 1;
|
|
|
|
|
|
|
|
ret = drm_dp_send_sideband_msg(mgr, up, chunk, idx);
|
2019-10-01 22:06:14 +08:00
|
|
|
if (unlikely(ret) && drm_debug_enabled(DRM_UT_DP)) {
|
2019-09-04 04:45:45 +08:00
|
|
|
struct drm_printer p = drm_debug_printer(DBG_PREFIX);
|
|
|
|
|
|
|
|
drm_printf(&p, "sideband msg failed to send\n");
|
|
|
|
drm_dp_mst_dump_sideband_msg_tx(&p, txmsg);
|
2014-06-05 12:01:32 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
txmsg->cur_offset += tosend;
|
|
|
|
if (txmsg->cur_offset == txmsg->cur_len) {
|
|
|
|
txmsg->state = DRM_DP_SIDEBAND_TX_SENT;
|
|
|
|
return 1;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void process_single_down_tx_qlock(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
|
|
|
int ret;
|
|
|
|
|
2015-01-28 17:02:23 +08:00
|
|
|
WARN_ON(!mutex_is_locked(&mgr->qlock));
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* construct a chunk from the first msg in the tx_msg queue */
|
2016-07-16 03:48:03 +08:00
|
|
|
if (list_empty(&mgr->tx_msg_downq))
|
2014-06-05 12:01:32 +08:00
|
|
|
return;
|
|
|
|
|
|
|
|
txmsg = list_first_entry(&mgr->tx_msg_downq, struct drm_dp_sideband_msg_tx, next);
|
|
|
|
ret = process_single_tx_qlock(mgr, txmsg, false);
|
|
|
|
if (ret == 1) {
|
|
|
|
/* txmsg is sent it should be in the slots now */
|
2020-01-13 17:36:49 +08:00
|
|
|
mgr->is_waiting_for_dwn_reply = true;
|
2014-06-05 12:01:32 +08:00
|
|
|
list_del(&txmsg->next);
|
|
|
|
} else if (ret) {
|
|
|
|
DRM_DEBUG_KMS("failed to send msg in q %d\n", ret);
|
2020-01-13 17:36:49 +08:00
|
|
|
mgr->is_waiting_for_dwn_reply = false;
|
2014-06-05 12:01:32 +08:00
|
|
|
list_del(&txmsg->next);
|
|
|
|
if (txmsg->seqno != -1)
|
|
|
|
txmsg->dst->tx_slots[txmsg->seqno] = NULL;
|
|
|
|
txmsg->state = DRM_DP_SIDEBAND_TX_TIMEOUT;
|
2017-05-13 18:52:01 +08:00
|
|
|
wake_up_all(&mgr->tx_waitq);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/* called holding qlock */
|
2015-12-19 06:14:43 +08:00
|
|
|
static void process_single_up_tx_qlock(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
/* construct a chunk from the first msg in the tx_msg queue */
|
|
|
|
ret = process_single_tx_qlock(mgr, txmsg, true);
|
2015-12-19 06:14:43 +08:00
|
|
|
|
|
|
|
if (ret != 1)
|
2014-06-05 12:01:32 +08:00
|
|
|
DRM_DEBUG_KMS("failed to send msg in q %d\n", ret);
|
2015-12-19 06:14:43 +08:00
|
|
|
|
drm/mst: Fix MST sideband up-reply failure handling
Fix the breakage resulting in the stacktrace below, due to tx queue
being full when trying to send an up-reply. txmsg->seqno is -1 in this
case leading to a corruption of the mstb object by
txmsg->dst->tx_slots[txmsg->seqno] = NULL;
in process_single_up_tx_qlock().
[ +0,005162] [drm:process_single_tx_qlock [drm_kms_helper]] set_hdr_from_dst_qlock: failed to find slot
[ +0,000015] [drm:drm_dp_send_up_ack_reply.constprop.19 [drm_kms_helper]] failed to send msg in q -11
[ +0,000939] BUG: kernel NULL pointer dereference, address: 00000000000005a0
[ +0,006982] #PF: supervisor write access in kernel mode
[ +0,005223] #PF: error_code(0x0002) - not-present page
[ +0,005135] PGD 0 P4D 0
[ +0,002581] Oops: 0002 [#1] PREEMPT SMP NOPTI
[ +0,004359] CPU: 1 PID: 1200 Comm: kworker/u16:3 Tainted: G U 5.2.0-rc1+ #410
[ +0,008433] Hardware name: Intel Corporation Ice Lake Client Platform/IceLake U DDR4 SODIMM PD RVP, BIOS ICLSFWR1.R00.3175.A00.1904261428 04/26/2019
[ +0,013323] Workqueue: i915-dp i915_digport_work_func [i915]
[ +0,005676] RIP: 0010:queue_work_on+0x19/0x70
[ +0,004372] Code: ff ff ff 0f 1f 40 00 66 2e 0f 1f 84 00 00 00 00 00 41 56 49 89 f6 41 55 41 89 fd 41 54 55 53 48 89 d3 9c 5d fa e8 e7 81 0c 00 <f0> 48 0f ba 2b 00 73 31 45 31 e4 f7 c5 00 02 00 00 74 13 e8 cf 7f
[ +0,018750] RSP: 0018:ffffc900007dfc50 EFLAGS: 00010006
[ +0,005222] RAX: 0000000000000046 RBX: 00000000000005a0 RCX: 0000000000000001
[ +0,007133] RDX: 000000000001b608 RSI: 0000000000000000 RDI: ffffffff82121972
[ +0,007129] RBP: 0000000000000202 R08: 0000000000000000 R09: 0000000000000001
[ +0,007129] R10: 0000000000000000 R11: 0000000000000000 R12: ffff88847bfa5096
[ +0,007131] R13: 0000000000000010 R14: ffff88849c08f3f8 R15: 0000000000000000
[ +0,007128] FS: 0000000000000000(0000) GS:ffff88849dc80000(0000) knlGS:0000000000000000
[ +0,008083] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[ +0,005749] CR2: 00000000000005a0 CR3: 0000000005210006 CR4: 0000000000760ee0
[ +0,007128] PKRU: 55555554
[ +0,002722] Call Trace:
[ +0,002458] drm_dp_mst_handle_up_req+0x517/0x540 [drm_kms_helper]
[ +0,006197] ? drm_dp_mst_hpd_irq+0x5b/0x9c0 [drm_kms_helper]
[ +0,005764] drm_dp_mst_hpd_irq+0x5b/0x9c0 [drm_kms_helper]
[ +0,005623] ? intel_dp_hpd_pulse+0x205/0x370 [i915]
[ +0,005018] intel_dp_hpd_pulse+0x205/0x370 [i915]
[ +0,004836] i915_digport_work_func+0xbb/0x140 [i915]
[ +0,005108] process_one_work+0x245/0x610
[ +0,004027] worker_thread+0x37/0x380
[ +0,003684] ? process_one_work+0x610/0x610
[ +0,004184] kthread+0x119/0x130
[ +0,003240] ? kthread_park+0x80/0x80
[ +0,003668] ret_from_fork+0x24/0x50
Cc: Lyude Paul <lyude@redhat.com>
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190523212433.9058-1-imre.deak@intel.com
2019-05-24 05:24:33 +08:00
|
|
|
if (txmsg->seqno != -1) {
|
|
|
|
WARN_ON((unsigned int)txmsg->seqno >
|
|
|
|
ARRAY_SIZE(txmsg->dst->tx_slots));
|
|
|
|
txmsg->dst->tx_slots[txmsg->seqno] = NULL;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_queue_down_tx(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg)
|
|
|
|
{
|
|
|
|
mutex_lock(&mgr->qlock);
|
|
|
|
list_add_tail(&txmsg->next, &mgr->tx_msg_downq);
|
2019-09-04 04:45:45 +08:00
|
|
|
|
2019-10-01 22:06:14 +08:00
|
|
|
if (drm_debug_enabled(DRM_UT_DP)) {
|
2019-09-04 04:45:45 +08:00
|
|
|
struct drm_printer p = drm_debug_printer(DBG_PREFIX);
|
|
|
|
|
|
|
|
drm_dp_mst_dump_sideband_msg_tx(&p, txmsg);
|
|
|
|
}
|
|
|
|
|
2020-01-13 17:36:49 +08:00
|
|
|
if (list_is_singular(&mgr->tx_msg_downq) &&
|
|
|
|
!mgr->is_waiting_for_dwn_reply)
|
2014-06-05 12:01:32 +08:00
|
|
|
process_single_down_tx_qlock(mgr);
|
|
|
|
mutex_unlock(&mgr->qlock);
|
|
|
|
}
|
|
|
|
|
2019-09-04 04:45:39 +08:00
|
|
|
static void
|
|
|
|
drm_dp_dump_link_address(struct drm_dp_link_address_ack_reply *reply)
|
|
|
|
{
|
|
|
|
struct drm_dp_link_addr_reply_port *port_reply;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < reply->nports; i++) {
|
|
|
|
port_reply = &reply->ports[i];
|
|
|
|
DRM_DEBUG_KMS("port %d: input %d, pdt: %d, pn: %d, dpcd_rev: %02x, mcs: %d, ddps: %d, ldps %d, sdp %d/%d\n",
|
|
|
|
i,
|
|
|
|
port_reply->input_port,
|
|
|
|
port_reply->peer_device_type,
|
|
|
|
port_reply->port_number,
|
|
|
|
port_reply->dpcd_revision,
|
|
|
|
port_reply->mcs,
|
|
|
|
port_reply->ddps,
|
|
|
|
port_reply->legacy_device_plug_status,
|
|
|
|
port_reply->num_sdp_streams,
|
|
|
|
port_reply->num_sdp_stream_sinks);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
static int drm_dp_send_link_address(struct drm_dp_mst_topology_mgr *mgr,
|
2015-09-06 16:53:00 +08:00
|
|
|
struct drm_dp_mst_branch *mstb)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
2019-09-04 04:45:53 +08:00
|
|
|
struct drm_dp_link_address_ack_reply *reply;
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
struct drm_dp_mst_port *port, *tmp;
|
|
|
|
int i, len, ret, port_mask = 0;
|
|
|
|
bool changed = false;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg)
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
return -ENOMEM;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
txmsg->dst = mstb;
|
|
|
|
len = build_link_address(txmsg);
|
|
|
|
|
2015-09-06 16:53:00 +08:00
|
|
|
mstb->link_address_sent = true;
|
2014-06-05 12:01:32 +08:00
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
2019-09-04 04:45:53 +08:00
|
|
|
/* FIXME: Actually do some real error handling here */
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = drm_dp_mst_wait_tx_reply(mstb, txmsg);
|
2019-09-04 04:45:53 +08:00
|
|
|
if (ret <= 0) {
|
|
|
|
DRM_ERROR("Sending link address failed with %d\n", ret);
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK) {
|
|
|
|
DRM_ERROR("link address NAK received\n");
|
|
|
|
ret = -EIO;
|
|
|
|
goto out;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:53 +08:00
|
|
|
reply = &txmsg->reply.u.link_addr;
|
|
|
|
DRM_DEBUG_KMS("link address reply: %d\n", reply->nports);
|
|
|
|
drm_dp_dump_link_address(reply);
|
2016-01-23 06:07:28 +08:00
|
|
|
|
2019-09-04 04:45:53 +08:00
|
|
|
drm_dp_check_mstb_guid(mstb, reply->guid);
|
2016-01-23 06:07:28 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
for (i = 0; i < reply->nports; i++) {
|
|
|
|
port_mask |= BIT(reply->ports[i].port_number);
|
|
|
|
ret = drm_dp_mst_handle_link_address_port(mstb, mgr->dev,
|
|
|
|
&reply->ports[i]);
|
|
|
|
if (ret == 1)
|
|
|
|
changed = true;
|
|
|
|
else if (ret < 0)
|
|
|
|
goto out;
|
|
|
|
}
|
2019-09-04 04:45:53 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
/* Prune any ports that are currently a part of mstb in our in-memory
|
|
|
|
* topology, but were not seen in this link address. Usually this
|
|
|
|
* means that they were removed while the topology was out of sync,
|
|
|
|
* e.g. during suspend/resume
|
|
|
|
*/
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
list_for_each_entry_safe(port, tmp, &mstb->ports, next) {
|
|
|
|
if (port_mask & BIT(port->port_num))
|
|
|
|
continue;
|
|
|
|
|
|
|
|
DRM_DEBUG_KMS("port %d was not in link address, removing\n",
|
|
|
|
port->port_num);
|
|
|
|
list_del(&port->next);
|
|
|
|
drm_dp_mst_topology_put_port(port);
|
|
|
|
changed = true;
|
|
|
|
}
|
|
|
|
mutex_unlock(&mgr->lock);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:53 +08:00
|
|
|
out:
|
|
|
|
if (ret <= 0)
|
|
|
|
mstb->link_address_sent = false;
|
2014-06-05 12:01:32 +08:00
|
|
|
kfree(txmsg);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
return ret < 0 ? ret : changed;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2019-09-04 04:45:47 +08:00
|
|
|
static int
|
|
|
|
drm_dp_send_enum_path_resources(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb,
|
|
|
|
struct drm_dp_mst_port *port)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-09-04 04:45:47 +08:00
|
|
|
struct drm_dp_enum_path_resources_ack_reply *path_res;
|
2014-06-05 12:01:32 +08:00
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
2019-09-04 04:45:47 +08:00
|
|
|
int len;
|
2014-06-05 12:01:32 +08:00
|
|
|
int ret;
|
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
txmsg->dst = mstb;
|
|
|
|
len = build_enum_path_resources(txmsg, port->port_num);
|
|
|
|
|
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
|
|
|
ret = drm_dp_mst_wait_tx_reply(mstb, txmsg);
|
|
|
|
if (ret > 0) {
|
2019-09-04 04:45:47 +08:00
|
|
|
path_res = &txmsg->reply.u.path_resources;
|
|
|
|
|
2019-01-23 04:03:00 +08:00
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK) {
|
2014-06-05 12:01:32 +08:00
|
|
|
DRM_DEBUG_KMS("enum path resources nak received\n");
|
2019-01-23 04:03:00 +08:00
|
|
|
} else {
|
2019-09-04 04:45:47 +08:00
|
|
|
if (port->port_num != path_res->port_number)
|
2014-06-05 12:01:32 +08:00
|
|
|
DRM_ERROR("got incorrect port in response\n");
|
2019-09-04 04:45:47 +08:00
|
|
|
|
|
|
|
DRM_DEBUG_KMS("enum path resources %d: %d %d\n",
|
|
|
|
path_res->port_number,
|
|
|
|
path_res->full_payload_bw_number,
|
|
|
|
path_res->avail_payload_bw_number);
|
|
|
|
port->available_pbn =
|
|
|
|
path_res->avail_payload_bw_number;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
kfree(txmsg);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2016-01-27 22:39:36 +08:00
|
|
|
static struct drm_dp_mst_port *drm_dp_get_last_connected_port_to_mstb(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
if (!mstb->port_parent)
|
|
|
|
return NULL;
|
|
|
|
|
|
|
|
if (mstb->port_parent->mstb != mstb)
|
|
|
|
return mstb->port_parent;
|
|
|
|
|
|
|
|
return drm_dp_get_last_connected_port_to_mstb(mstb->port_parent->parent);
|
|
|
|
}
|
|
|
|
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
/*
|
|
|
|
* Searches upwards in the topology starting from mstb to try to find the
|
|
|
|
* closest available parent of mstb that's still connected to the rest of the
|
|
|
|
* topology. This can be used in order to perform operations like releasing
|
|
|
|
* payloads, where the branch device which owned the payload may no longer be
|
|
|
|
* around and thus would require that the payload on the last living relative
|
|
|
|
* be freed instead.
|
|
|
|
*/
|
|
|
|
static struct drm_dp_mst_branch *
|
|
|
|
drm_dp_get_last_connected_port_and_mstb(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb,
|
|
|
|
int *port_num)
|
2016-01-27 22:39:36 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *rmstb = NULL;
|
|
|
|
struct drm_dp_mst_port *found_port;
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
|
2016-01-27 22:39:36 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
if (!mgr->mst_primary)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
do {
|
2016-01-27 22:39:36 +08:00
|
|
|
found_port = drm_dp_get_last_connected_port_to_mstb(mstb);
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
if (!found_port)
|
|
|
|
break;
|
2016-01-27 22:39:36 +08:00
|
|
|
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
if (drm_dp_mst_topology_try_get_mstb(found_port->parent)) {
|
2016-01-27 22:39:36 +08:00
|
|
|
rmstb = found_port->parent;
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
*port_num = found_port->port_num;
|
|
|
|
} else {
|
|
|
|
/* Search again, starting from this parent */
|
|
|
|
mstb = found_port->parent;
|
2016-01-27 22:39:36 +08:00
|
|
|
}
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
} while (!rmstb);
|
|
|
|
out:
|
2016-01-27 22:39:36 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
return rmstb;
|
|
|
|
}
|
|
|
|
|
2014-07-21 19:23:57 +08:00
|
|
|
static int drm_dp_payload_send_msg(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
|
|
|
int id,
|
|
|
|
int pbn)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
2016-01-27 22:39:36 +08:00
|
|
|
int len, ret, port_num;
|
2015-12-02 14:09:43 +08:00
|
|
|
u8 sinks[DRM_DP_MAX_SDP_STREAMS];
|
|
|
|
int i;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2016-01-27 22:39:36 +08:00
|
|
|
port_num = port->port_num;
|
2019-01-11 08:53:28 +08:00
|
|
|
mstb = drm_dp_mst_topology_get_mstb_validated(mgr, port->parent);
|
2016-01-27 22:39:36 +08:00
|
|
|
if (!mstb) {
|
2019-01-11 08:53:25 +08:00
|
|
|
mstb = drm_dp_get_last_connected_port_and_mstb(mgr,
|
|
|
|
port->parent,
|
|
|
|
&port_num);
|
2016-01-27 22:39:36 +08:00
|
|
|
|
2019-01-11 08:53:32 +08:00
|
|
|
if (!mstb)
|
2016-01-27 22:39:36 +08:00
|
|
|
return -EINVAL;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto fail_put;
|
|
|
|
}
|
|
|
|
|
2015-12-02 14:09:43 +08:00
|
|
|
for (i = 0; i < port->num_sdp_streams; i++)
|
|
|
|
sinks[i] = i;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
txmsg->dst = mstb;
|
2016-01-27 22:39:36 +08:00
|
|
|
len = build_allocate_payload(txmsg, port_num,
|
2014-06-05 12:01:32 +08:00
|
|
|
id,
|
2015-12-02 14:09:43 +08:00
|
|
|
pbn, port->num_sdp_streams, sinks);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
drm/dp_mst: Restart last_connected_port_and_mstb() if topology ref fails
While this isn't a complete fix, this will improve the reliability of
drm_dp_get_last_connected_port_and_mstb() pretty significantly during
hotplug events, since there's a chance that the in-memory topology tree
may not be fully updated when drm_dp_get_last_connected_port_and_mstb()
is called and thus might end up causing our search to fail on an mstb
whose topology refcount has reached 0, but has not yet been removed from
it's parent.
Ideally, we should further fix this problem by ensuring that we deal
with the potential for racing with a hotplug event, which would look
like this:
* drm_dp_payload_send_msg() retrieves the last living relative of mstb
with drm_dp_get_last_connected_port_and_mstb()
* drm_dp_payload_send_msg() starts building payload message
At the same time, mstb gets unplugged from the topology and is no
longer the actual last living relative of the original mstb
* drm_dp_payload_send_msg() tries sending the payload message, hub times
out
* Hub timed out, we give up and run away-resulting in the payload being
leaked
This could be fixed by restarting the
drm_dp_get_last_connected_port_and_mstb() search whenever we get a
timeout, sending the payload to the new mstb, then repeating until
either the entire topology is removed from the system or
drm_dp_get_last_connected_port_and_mstb() fails. But since the above
race condition is not terribly likely, we'll address that in a later
patch series once we've improved the recovery handling for VCPI
allocations in the rest of the DP MST helpers.
Changes since v1:
* Convert kerneldoc for drm_dp_get_last_connected_port_and_mstb to
normal comment - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-8-lyude@redhat.com
2019-01-11 08:53:30 +08:00
|
|
|
/*
|
|
|
|
* FIXME: there is a small chance that between getting the last
|
|
|
|
* connected mstb and sending the payload message, the last connected
|
|
|
|
* mstb could also be removed from the topology. In the future, this
|
|
|
|
* needs to be fixed by restarting the
|
|
|
|
* drm_dp_get_last_connected_port_and_mstb() search in the event of a
|
|
|
|
* timeout if the topology is still connected to the system.
|
|
|
|
*/
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = drm_dp_mst_wait_tx_reply(mstb, txmsg);
|
|
|
|
if (ret > 0) {
|
2019-01-23 04:03:00 +08:00
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK)
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = -EINVAL;
|
2019-01-11 08:53:25 +08:00
|
|
|
else
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = 0;
|
|
|
|
}
|
|
|
|
kfree(txmsg);
|
|
|
|
fail_put:
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
2014-06-05 12:01:32 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2017-09-07 08:14:58 +08:00
|
|
|
int drm_dp_send_power_updown_phy(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port, bool power_up)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
|
|
|
int len, ret;
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
port = drm_dp_mst_topology_get_port_validated(mgr, port);
|
2017-09-07 08:14:58 +08:00
|
|
|
if (!port)
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg) {
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2017-09-07 08:14:58 +08:00
|
|
|
return -ENOMEM;
|
|
|
|
}
|
|
|
|
|
|
|
|
txmsg->dst = port->parent;
|
|
|
|
len = build_power_updown_phy(txmsg, port->port_num, power_up);
|
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
|
|
|
ret = drm_dp_mst_wait_tx_reply(port->parent, txmsg);
|
|
|
|
if (ret > 0) {
|
2019-01-23 04:03:00 +08:00
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK)
|
2017-09-07 08:14:58 +08:00
|
|
|
ret = -EINVAL;
|
|
|
|
else
|
|
|
|
ret = 0;
|
|
|
|
}
|
|
|
|
kfree(txmsg);
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2017-09-07 08:14:58 +08:00
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_send_power_updown_phy);
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static int drm_dp_create_payload_step1(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
int id,
|
|
|
|
struct drm_dp_payload *payload)
|
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
|
|
|
ret = drm_dp_dpcd_write_payload(mgr, id, payload);
|
|
|
|
if (ret < 0) {
|
|
|
|
payload->payload_state = 0;
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
payload->payload_state = DP_PAYLOAD_LOCAL;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-07-21 19:23:57 +08:00
|
|
|
static int drm_dp_create_payload_step2(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
|
|
|
int id,
|
|
|
|
struct drm_dp_payload *payload)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
ret = drm_dp_payload_send_msg(mgr, port, id, port->vcpi.pbn);
|
|
|
|
if (ret < 0)
|
|
|
|
return ret;
|
|
|
|
payload->payload_state = DP_PAYLOAD_REMOTE;
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
2014-07-21 19:23:57 +08:00
|
|
|
static int drm_dp_destroy_payload_step1(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
|
|
|
int id,
|
|
|
|
struct drm_dp_payload *payload)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
DRM_DEBUG_KMS("\n");
|
2019-02-02 09:23:26 +08:00
|
|
|
/* it's okay for these to fail */
|
2014-06-05 12:01:32 +08:00
|
|
|
if (port) {
|
|
|
|
drm_dp_payload_send_msg(mgr, port, id, 0);
|
|
|
|
}
|
|
|
|
|
|
|
|
drm_dp_dpcd_write_payload(mgr, id, payload);
|
2014-08-06 14:26:21 +08:00
|
|
|
payload->payload_state = DP_PAYLOAD_DELETE_LOCAL;
|
2014-06-05 12:01:32 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2014-07-21 19:23:57 +08:00
|
|
|
static int drm_dp_destroy_payload_step2(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
int id,
|
|
|
|
struct drm_dp_payload *payload)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
payload->payload_state = 0;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_update_payload_part1() - Execute payload update part 1
|
|
|
|
* @mgr: manager to use.
|
|
|
|
*
|
|
|
|
* This iterates over all proposed virtual channels, and tries to
|
|
|
|
* allocate space in the link for them. For 0->slots transitions,
|
|
|
|
* this step just writes the VCPI to the MST device. For slots->0
|
|
|
|
* transitions, this writes the updated VCPIs and removes the
|
|
|
|
* remote VC payloads.
|
|
|
|
*
|
|
|
|
* after calling this the driver should generate ACT and payload
|
|
|
|
* packets.
|
|
|
|
*/
|
|
|
|
int drm_dp_update_payload_part1(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
struct drm_dp_payload req_payload;
|
|
|
|
struct drm_dp_mst_port *port;
|
2019-01-11 08:53:32 +08:00
|
|
|
int i, j;
|
|
|
|
int cur_slots = 1;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
mutex_lock(&mgr->payload_lock);
|
|
|
|
for (i = 0; i < mgr->max_payloads; i++) {
|
2018-12-14 09:25:31 +08:00
|
|
|
struct drm_dp_vcpi *vcpi = mgr->proposed_vcpis[i];
|
|
|
|
struct drm_dp_payload *payload = &mgr->payloads[i];
|
2019-01-11 08:53:32 +08:00
|
|
|
bool put_port = false;
|
2018-12-14 09:25:31 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* solve the current payloads - compare to the hw ones
|
|
|
|
- update the hw view */
|
|
|
|
req_payload.start_slot = cur_slots;
|
2018-12-14 09:25:31 +08:00
|
|
|
if (vcpi) {
|
|
|
|
port = container_of(vcpi, struct drm_dp_mst_port,
|
|
|
|
vcpi);
|
2019-01-11 08:53:32 +08:00
|
|
|
|
|
|
|
/* Validated ports don't matter if we're releasing
|
|
|
|
* VCPI
|
|
|
|
*/
|
|
|
|
if (vcpi->num_slots) {
|
|
|
|
port = drm_dp_mst_topology_get_port_validated(
|
|
|
|
mgr, port);
|
|
|
|
if (!port) {
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
put_port = true;
|
2016-04-23 04:08:46 +08:00
|
|
|
}
|
2019-01-11 08:53:32 +08:00
|
|
|
|
2018-12-14 09:25:31 +08:00
|
|
|
req_payload.num_slots = vcpi->num_slots;
|
|
|
|
req_payload.vcpi = vcpi->vcpi;
|
2014-06-05 12:01:32 +08:00
|
|
|
} else {
|
|
|
|
port = NULL;
|
|
|
|
req_payload.num_slots = 0;
|
|
|
|
}
|
2014-08-06 14:26:21 +08:00
|
|
|
|
2018-12-14 09:25:31 +08:00
|
|
|
payload->start_slot = req_payload.start_slot;
|
2014-06-05 12:01:32 +08:00
|
|
|
/* work out what is required to happen with this payload */
|
2018-12-14 09:25:31 +08:00
|
|
|
if (payload->num_slots != req_payload.num_slots) {
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
/* need to push an update for this payload */
|
|
|
|
if (req_payload.num_slots) {
|
2018-12-14 09:25:31 +08:00
|
|
|
drm_dp_create_payload_step1(mgr, vcpi->vcpi,
|
|
|
|
&req_payload);
|
|
|
|
payload->num_slots = req_payload.num_slots;
|
|
|
|
payload->vcpi = req_payload.vcpi;
|
|
|
|
|
|
|
|
} else if (payload->num_slots) {
|
|
|
|
payload->num_slots = 0;
|
|
|
|
drm_dp_destroy_payload_step1(mgr, port,
|
|
|
|
payload->vcpi,
|
|
|
|
payload);
|
|
|
|
req_payload.payload_state =
|
|
|
|
payload->payload_state;
|
|
|
|
payload->start_slot = 0;
|
2014-08-06 14:26:21 +08:00
|
|
|
}
|
2018-12-14 09:25:31 +08:00
|
|
|
payload->payload_state = req_payload.payload_state;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
cur_slots += req_payload.num_slots;
|
2016-04-23 04:08:46 +08:00
|
|
|
|
2019-01-11 08:53:32 +08:00
|
|
|
if (put_port)
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2014-08-06 14:26:21 +08:00
|
|
|
|
2019-12-03 12:24:23 +08:00
|
|
|
for (i = 0; i < mgr->max_payloads; /* do nothing */) {
|
|
|
|
if (mgr->payloads[i].payload_state != DP_PAYLOAD_DELETE_LOCAL) {
|
|
|
|
i++;
|
2018-12-14 09:25:31 +08:00
|
|
|
continue;
|
2019-12-03 12:24:23 +08:00
|
|
|
}
|
2014-08-06 14:26:21 +08:00
|
|
|
|
2018-12-14 09:25:31 +08:00
|
|
|
DRM_DEBUG_KMS("removing payload %d\n", i);
|
|
|
|
for (j = i; j < mgr->max_payloads - 1; j++) {
|
|
|
|
mgr->payloads[j] = mgr->payloads[j + 1];
|
|
|
|
mgr->proposed_vcpis[j] = mgr->proposed_vcpis[j + 1];
|
|
|
|
|
|
|
|
if (mgr->proposed_vcpis[j] &&
|
|
|
|
mgr->proposed_vcpis[j]->num_slots) {
|
|
|
|
set_bit(j + 1, &mgr->payload_mask);
|
|
|
|
} else {
|
|
|
|
clear_bit(j + 1, &mgr->payload_mask);
|
|
|
|
}
|
2014-08-06 14:26:21 +08:00
|
|
|
}
|
2018-12-14 09:25:31 +08:00
|
|
|
|
|
|
|
memset(&mgr->payloads[mgr->max_payloads - 1], 0,
|
|
|
|
sizeof(struct drm_dp_payload));
|
|
|
|
mgr->proposed_vcpis[mgr->max_payloads - 1] = NULL;
|
|
|
|
clear_bit(mgr->max_payloads, &mgr->payload_mask);
|
2014-08-06 14:26:21 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_update_payload_part1);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_update_payload_part2() - Execute payload update part 2
|
|
|
|
* @mgr: manager to use.
|
|
|
|
*
|
|
|
|
* This iterates over all proposed virtual channels, and tries to
|
|
|
|
* allocate space in the link for them. For 0->slots transitions,
|
|
|
|
* this step writes the remote VC payload commands. For slots->0
|
|
|
|
* this just resets some internal state.
|
|
|
|
*/
|
|
|
|
int drm_dp_update_payload_part2(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port;
|
|
|
|
int i;
|
2014-07-14 18:53:44 +08:00
|
|
|
int ret = 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_lock(&mgr->payload_lock);
|
|
|
|
for (i = 0; i < mgr->max_payloads; i++) {
|
|
|
|
|
|
|
|
if (!mgr->proposed_vcpis[i])
|
|
|
|
continue;
|
|
|
|
|
|
|
|
port = container_of(mgr->proposed_vcpis[i], struct drm_dp_mst_port, vcpi);
|
|
|
|
|
|
|
|
DRM_DEBUG_KMS("payload %d %d\n", i, mgr->payloads[i].payload_state);
|
|
|
|
if (mgr->payloads[i].payload_state == DP_PAYLOAD_LOCAL) {
|
2014-08-06 14:26:21 +08:00
|
|
|
ret = drm_dp_create_payload_step2(mgr, port, mgr->proposed_vcpis[i]->vcpi, &mgr->payloads[i]);
|
2014-06-05 12:01:32 +08:00
|
|
|
} else if (mgr->payloads[i].payload_state == DP_PAYLOAD_DELETE_LOCAL) {
|
2014-08-06 14:26:21 +08:00
|
|
|
ret = drm_dp_destroy_payload_step2(mgr, mgr->proposed_vcpis[i]->vcpi, &mgr->payloads[i]);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
if (ret) {
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_update_payload_part2);
|
|
|
|
|
|
|
|
static int drm_dp_send_dpcd_read(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
int offset, int size, u8 *bytes)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int len;
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
int ret = 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
|
|
|
|
mstb = drm_dp_mst_topology_get_mstb_validated(mgr, port->parent);
|
|
|
|
if (!mstb)
|
|
|
|
return -EINVAL;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
if (!txmsg) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto fail_put;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
len = build_dpcd_read(txmsg, port->port_num, offset, size);
|
2014-06-05 12:01:32 +08:00
|
|
|
txmsg->dst = port->parent;
|
|
|
|
|
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
ret = drm_dp_mst_wait_tx_reply(mstb, txmsg);
|
|
|
|
if (ret < 0)
|
|
|
|
goto fail_free;
|
|
|
|
|
|
|
|
/* DPCD read should never be NACKed */
|
|
|
|
if (txmsg->reply.reply_type == 1) {
|
|
|
|
DRM_ERROR("mstb %p port %d: DPCD read on addr 0x%x for %d bytes NAKed\n",
|
|
|
|
mstb, port->port_num, offset, size);
|
|
|
|
ret = -EIO;
|
|
|
|
goto fail_free;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (txmsg->reply.u.remote_dpcd_read_ack.num_bytes != size) {
|
|
|
|
ret = -EPROTO;
|
|
|
|
goto fail_free;
|
|
|
|
}
|
|
|
|
|
|
|
|
ret = min_t(size_t, txmsg->reply.u.remote_dpcd_read_ack.num_bytes,
|
|
|
|
size);
|
|
|
|
memcpy(bytes, txmsg->reply.u.remote_dpcd_read_ack.bytes, ret);
|
|
|
|
|
|
|
|
fail_free:
|
|
|
|
kfree(txmsg);
|
|
|
|
fail_put:
|
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
|
|
|
|
|
|
|
return ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static int drm_dp_send_dpcd_write(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port,
|
|
|
|
int offset, int size, u8 *bytes)
|
|
|
|
{
|
|
|
|
int len;
|
|
|
|
int ret;
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
mstb = drm_dp_mst_topology_get_mstb_validated(mgr, port->parent);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!mstb)
|
|
|
|
return -EINVAL;
|
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto fail_put;
|
|
|
|
}
|
|
|
|
|
|
|
|
len = build_dpcd_write(txmsg, port->port_num, offset, size, bytes);
|
|
|
|
txmsg->dst = mstb;
|
|
|
|
|
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
|
|
|
ret = drm_dp_mst_wait_tx_reply(mstb, txmsg);
|
|
|
|
if (ret > 0) {
|
2019-01-23 04:03:00 +08:00
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK)
|
drm/dp_mst: Enable registration of AUX devices for MST ports
All available downstream ports - physical and logical - are exposed for
each MST device. They are listed in /dev/, following the same naming
scheme as SST devices by appending an incremental ID.
Although all downstream ports are exposed, only some will work as
expected. Consider the following topology:
+---------+
| ASIC |
+---------+
Conn-0|
|
+----v----+
+----| MST HUB |----+
| +---------+ |
| |
|Port-1 Port-2|
+-----v-----+ +-----v-----+
| MST | | SST |
| Display | | Display |
+-----------+ +-----------+
|Port-1
x
MST Path | MST Device
----------+----------------------------------
sst:0 | MST Hub
mst:0-1 | MST Display
mst:0-1-1 | MST Display's disconnected DP out
mst:0-1-8 | MST Display's internal sink
mst:0-2 | SST Display
On certain MST displays, the upstream physical port will ACK DPCD reads.
However, reads on the local logical port to the internal sink will
*NAK*. i.e. reading mst:0-1 ACKs, but mst:0-1-8 NAKs.
There may also be duplicates. Some displays will return the same GUID
when reading DPCD from both mst:0-1 and mst:0-1-8.
There are some device-dependent behavior as well. The MST hub used
during testing will actually *ACK* read requests on a disconnected
physical port, whereas the MST displays will NAK.
In light of these discrepancies, it's simpler to expose all downstream
ports - both physical and logical - and let the user decide what to use.
v3 changes:
* Change WARN_ON_ONCE -> DRM_ERROR on dpcd read errors
* Docstring and cosmetic fixes
v2 changes:
Moved remote aux device (un)registration to new mst connector late
register and early unregister helpers. Drivers should call these from
their own mst connector function hooks.
This is to solve an issue during driver unload, where mst connector
devices are unregistered before the remote aux devices are. In a setup
where aux devices are created as children of connector devices, the aux
device would be removed too early, and uncleanly. Doing so in
early_unregister solves this issue, as that is called before connector
unregistration.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190723232808.28128-3-sunpeng.li@amd.com
2019-07-24 07:28:01 +08:00
|
|
|
ret = -EIO;
|
2019-01-23 04:03:00 +08:00
|
|
|
else
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = 0;
|
|
|
|
}
|
|
|
|
kfree(txmsg);
|
|
|
|
fail_put:
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
2014-06-05 12:01:32 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int drm_dp_encode_up_ack_reply(struct drm_dp_sideband_msg_tx *msg, u8 req_type)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_reply_body reply;
|
|
|
|
|
2019-01-23 04:03:00 +08:00
|
|
|
reply.reply_type = DP_SIDEBAND_REPLY_ACK;
|
2014-06-05 12:01:32 +08:00
|
|
|
reply.req_type = req_type;
|
|
|
|
drm_dp_encode_sideband_reply(&reply, msg);
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static int drm_dp_send_up_ack_reply(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_branch *mstb,
|
|
|
|
int req_type, int seqno, bool broadcast)
|
|
|
|
{
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg)
|
|
|
|
return -ENOMEM;
|
|
|
|
|
|
|
|
txmsg->dst = mstb;
|
|
|
|
txmsg->seqno = seqno;
|
|
|
|
drm_dp_encode_up_ack_reply(txmsg, req_type);
|
|
|
|
|
|
|
|
mutex_lock(&mgr->qlock);
|
2015-12-19 06:14:43 +08:00
|
|
|
|
|
|
|
process_single_up_tx_qlock(mgr, txmsg);
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->qlock);
|
2015-12-19 06:14:43 +08:00
|
|
|
|
|
|
|
kfree(txmsg);
|
2014-06-05 12:01:32 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2019-09-25 22:14:41 +08:00
|
|
|
static int drm_dp_get_vc_payload_bw(u8 dp_link_bw, u8 dp_link_count)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-09-25 22:14:41 +08:00
|
|
|
if (dp_link_bw == 0 || dp_link_count == 0)
|
2014-11-12 18:13:37 +08:00
|
|
|
DRM_DEBUG_KMS("invalid link bandwidth in DPCD: %x (link count: %d)\n",
|
|
|
|
dp_link_bw, dp_link_count);
|
|
|
|
|
2019-09-25 22:14:41 +08:00
|
|
|
return dp_link_bw * dp_link_count / 2;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_mgr_set_mst() - Set the MST state for a topology manager
|
|
|
|
* @mgr: manager to set state for
|
|
|
|
* @mst_state: true to enable MST on this connector - false to disable.
|
|
|
|
*
|
|
|
|
* This is called by the driver when it detects an MST capable device plugged
|
|
|
|
* into a DP MST capable port, or when a DP MST capable device is unplugged.
|
|
|
|
*/
|
|
|
|
int drm_dp_mst_topology_mgr_set_mst(struct drm_dp_mst_topology_mgr *mgr, bool mst_state)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
struct drm_dp_mst_branch *mstb = NULL;
|
|
|
|
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
if (mst_state == mgr->mst_state)
|
|
|
|
goto out_unlock;
|
|
|
|
|
|
|
|
mgr->mst_state = mst_state;
|
|
|
|
/* set the device into MST mode */
|
|
|
|
if (mst_state) {
|
|
|
|
WARN_ON(mgr->mst_primary);
|
|
|
|
|
|
|
|
/* get dpcd info */
|
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_DPCD_REV, mgr->dpcd, DP_RECEIVER_CAP_SIZE);
|
|
|
|
if (ret != DP_RECEIVER_CAP_SIZE) {
|
|
|
|
DRM_DEBUG_KMS("failed to read DPCD\n");
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
2019-09-25 22:14:41 +08:00
|
|
|
mgr->pbn_div = drm_dp_get_vc_payload_bw(mgr->dpcd[1],
|
|
|
|
mgr->dpcd[2] & DP_MAX_LANE_COUNT_MASK);
|
|
|
|
if (mgr->pbn_div == 0) {
|
2014-11-12 18:13:37 +08:00
|
|
|
ret = -EINVAL;
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* add initial branch device at LCT 1 */
|
|
|
|
mstb = drm_dp_add_mst_branch_device(1, NULL);
|
|
|
|
if (mstb == NULL) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
mstb->mgr = mgr;
|
|
|
|
|
|
|
|
/* give this the main reference */
|
|
|
|
mgr->mst_primary = mstb;
|
drm/dp_mst: Introduce new refcounting scheme for mstbs and ports
The current way of handling refcounting in the DP MST helpers is really
confusing and probably just plain wrong because it's been hacked up many
times over the years without anyone actually going over the code and
seeing if things could be simplified.
To the best of my understanding, the current scheme works like this:
drm_dp_mst_port and drm_dp_mst_branch both have a single refcount. When
this refcount hits 0 for either of the two, they're removed from the
topology state, but not immediately freed. Both ports and branch devices
will reinitialize their kref once it's hit 0 before actually destroying
themselves. The intended purpose behind this is so that we can avoid
problems like not being able to free a remote payload that might still
be active, due to us having removed all of the port/branch device
structures in memory, as per:
commit 91a25e463130 ("drm/dp/mst: deallocate payload on port destruction")
Which may have worked, but then it caused use-after-free errors. Being
new to MST at the time, I tried fixing it;
commit 263efde31f97 ("drm/dp/mst: Get validated port ref in drm_dp_update_payload_part1()")
But, that was broken: both drm_dp_mst_port and drm_dp_mst_branch structs
are validated in almost every DP MST helper function. Simply put, this
means we go through the topology and try to see if the given
drm_dp_mst_branch or drm_dp_mst_port is still attached to something
before trying to use it in order to avoid dereferencing freed memory
(something that has happened a LOT in the past with this library).
Because of this it doesn't actually matter whether or not we keep keep
the ports and branches around in memory as that's not enough, because
any function that validates the branches and ports passed to it will
still reject them anyway since they're no longer in the topology
structure. So, use-after-free errors were fixed but payload deallocation
was completely broken.
Two years later, AMD informed me about this issue and I attempted to
come up with a temporary fix, pending a long-overdue cleanup of this
library:
commit c54c7374ff44 ("drm/dp_mst: Skip validating ports during destruction, just ref")
But then that introduced use-after-free errors, so I quickly reverted
it:
commit 9765635b3075 ("Revert "drm/dp_mst: Skip validating ports during destruction, just ref"")
And in the process, learned that there is just no simple fix for this:
the design is just broken. Unfortunately, the usage of these helpers are
quite broken as well. Some drivers like i915 have been smart enough to
avoid accessing any kind of information from MST port structures, but
others like nouveau have assumed, understandably so, that
drm_dp_mst_port structures are normal and can just be accessed at any
time without worrying about use-after-free errors.
After a lot of discussion, me and Daniel Vetter came up with a better
idea to replace all of this.
To summarize, since this is documented far more indepth in the
documentation this patch introduces, we make it so that drm_dp_mst_port
and drm_dp_mst_branch structures have two different classes of
refcounts: topology_kref, and malloc_kref. topology_kref corresponds to
the lifetime of the given drm_dp_mst_port or drm_dp_mst_branch in it's
given topology. Once it hits zero, any associated connectors are removed
and the branch or port can no longer be validated. malloc_kref
corresponds to the lifetime of the memory allocation for the actual
structure, and will always be non-zero so long as the topology_kref is
non-zero. This gives us a way to allow callers to hold onto port and
branch device structures past their topology lifetime, and dramatically
simplifies the lifetimes of both structures. This also finally fixes the
port deallocation problem, properly.
Additionally: since this now means that we can keep ports and branch
devices allocated in memory for however long we need, we no longer need
a significant amount of the port validation that we currently do.
Additionally, there is one last scenario that this fixes, which couldn't
have been fixed properly beforehand:
- CPU1 unrefs port from topology (refcount 1->0)
- CPU2 refs port in topology(refcount 0->1)
Since we now can guarantee memory safety for ports and branches
as-needed, we also can make our main reference counting functions fix
this problem by using kref_get_unless_zero() internally so that topology
refcounts can only ever reach 0 once.
Changes since v4:
* Change the kernel-figure summary for dp-mst/topology-figure-1.dot a
bit - danvet
* Remove figure numbers - danvet
Changes since v3:
* Remove rebase detritus - danvet
* Split out purely style changes into separate patches - hwentlan
Changes since v2:
* Fix commit message - checkpatch
* s/)-1/) - 1/g - checkpatch
Changes since v1:
* Remove forward declarations - danvet
* Move "Branch device and port refcounting" section from documentation
into kernel-doc comments - danvet
* Export internal topology lifetime functions into their own section in
the kernel-docs - danvet
* s/@/&/g for struct references in kernel-docs - danvet
* Drop the "when they are no longer being used" bits from the kernel
docs - danvet
* Modify diagrams to show how the DRM driver interacts with the topology
and payloads - danvet
* Make suggested documentation changes for
drm_dp_mst_topology_get_mstb() and drm_dp_mst_topology_get_port() -
danvet
* Better explain the relationship between malloc refs and topology krefs
in the documentation for drm_dp_mst_topology_get_port() and
drm_dp_mst_topology_get_mstb() - danvet
* Fix "See also" in drm_dp_mst_topology_get_mstb() - danvet
* Rename drm_dp_mst_topology_get_(port|mstb)() ->
drm_dp_mst_topology_try_get_(port|mstb)() and
drm_dp_mst_topology_ref_(port|mstb)() ->
drm_dp_mst_topology_get_(port|mstb)() - danvet
* s/should/must in docs - danvet
* WARN_ON(refcount == 0) in topology_get_(mstb|port) - danvet
* Move kdocs for mstb/port structs inline - danvet
* Split drm_dp_get_last_connected_port_and_mstb() changes into their own
commit - danvet
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-7-lyude@redhat.com
2019-01-11 08:53:29 +08:00
|
|
|
drm_dp_mst_topology_get_mstb(mgr->mst_primary);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2016-01-23 06:07:29 +08:00
|
|
|
ret = drm_dp_dpcd_writeb(mgr->aux, DP_MSTM_CTRL,
|
|
|
|
DP_MST_EN | DP_UP_REQ_EN | DP_UPSTREAM_IS_SRC);
|
|
|
|
if (ret < 0) {
|
|
|
|
goto out_unlock;
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_payload reset_pay;
|
|
|
|
reset_pay.start_slot = 0;
|
|
|
|
reset_pay.num_slots = 0x3f;
|
|
|
|
drm_dp_dpcd_write_payload(mgr, 0, &reset_pay);
|
|
|
|
}
|
|
|
|
|
|
|
|
queue_work(system_long_wq, &mgr->work);
|
|
|
|
|
|
|
|
ret = 0;
|
|
|
|
} else {
|
|
|
|
/* disable MST on the device */
|
|
|
|
mstb = mgr->mst_primary;
|
|
|
|
mgr->mst_primary = NULL;
|
|
|
|
/* this can fail if the device is gone */
|
|
|
|
drm_dp_dpcd_writeb(mgr->aux, DP_MSTM_CTRL, 0);
|
|
|
|
ret = 0;
|
|
|
|
memset(mgr->payloads, 0, mgr->max_payloads * sizeof(struct drm_dp_payload));
|
|
|
|
mgr->payload_mask = 0;
|
|
|
|
set_bit(0, &mgr->payload_mask);
|
2014-08-06 14:26:21 +08:00
|
|
|
mgr->vcpi_mask = 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
out_unlock:
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
if (mstb)
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
2014-06-05 12:01:32 +08:00
|
|
|
return ret;
|
|
|
|
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_topology_mgr_set_mst);
|
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
static void
|
|
|
|
drm_dp_mst_topology_mgr_invalidate_mstb(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port;
|
|
|
|
|
|
|
|
/* The link address will need to be re-sent on resume */
|
|
|
|
mstb->link_address_sent = false;
|
|
|
|
|
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
|
|
|
/* The PBN for each port will also need to be re-probed */
|
|
|
|
port->available_pbn = 0;
|
|
|
|
|
|
|
|
if (port->mstb)
|
|
|
|
drm_dp_mst_topology_mgr_invalidate_mstb(port->mstb);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_mgr_suspend() - suspend the MST manager
|
|
|
|
* @mgr: manager to suspend
|
|
|
|
*
|
|
|
|
* This function tells the MST device that we can't handle UP messages
|
|
|
|
* anymore. This should stop it from sending any since we are suspended.
|
|
|
|
*/
|
|
|
|
void drm_dp_mst_topology_mgr_suspend(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
drm_dp_dpcd_writeb(mgr->aux, DP_MSTM_CTRL,
|
|
|
|
DP_MST_EN | DP_UPSTREAM_IS_SRC);
|
|
|
|
mutex_unlock(&mgr->lock);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
flush_work(&mgr->up_req_work);
|
2015-09-30 08:39:42 +08:00
|
|
|
flush_work(&mgr->work);
|
2019-02-20 06:41:02 +08:00
|
|
|
flush_work(&mgr->delayed_destroy_work);
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
if (mgr->mst_state && mgr->mst_primary)
|
|
|
|
drm_dp_mst_topology_mgr_invalidate_mstb(mgr->mst_primary);
|
|
|
|
mutex_unlock(&mgr->lock);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_topology_mgr_suspend);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_mgr_resume() - resume the MST manager
|
|
|
|
* @mgr: manager to resume
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
* @sync: whether or not to perform topology reprobing synchronously
|
2014-06-05 12:01:32 +08:00
|
|
|
*
|
|
|
|
* This will fetch DPCD and see if the device is still there,
|
|
|
|
* if it is, it will rewrite the MSTM control bits, and return.
|
|
|
|
*
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
* If the device fails this returns -1, and the driver should do
|
2014-06-05 12:01:32 +08:00
|
|
|
* a full MST reprobe, in case we were undocked.
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
*
|
|
|
|
* During system resume (where it is assumed that the driver will be calling
|
|
|
|
* drm_atomic_helper_resume()) this function should be called beforehand with
|
|
|
|
* @sync set to true. In contexts like runtime resume where the driver is not
|
|
|
|
* expected to be calling drm_atomic_helper_resume(), this function should be
|
|
|
|
* called with @sync set to false in order to avoid deadlocking.
|
|
|
|
*
|
|
|
|
* Returns: -1 if the MST topology was removed while we were suspended, 0
|
|
|
|
* otherwise.
|
2014-06-05 12:01:32 +08:00
|
|
|
*/
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
int drm_dp_mst_topology_mgr_resume(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
bool sync)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-10-10 03:14:59 +08:00
|
|
|
int ret;
|
|
|
|
u8 guid[16];
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
mutex_lock(&mgr->lock);
|
2019-10-10 03:14:59 +08:00
|
|
|
if (!mgr->mst_primary)
|
|
|
|
goto out_fail;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-10 03:14:59 +08:00
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_DPCD_REV, mgr->dpcd,
|
|
|
|
DP_RECEIVER_CAP_SIZE);
|
|
|
|
if (ret != DP_RECEIVER_CAP_SIZE) {
|
|
|
|
DRM_DEBUG_KMS("dpcd read failed - undocked during suspend?\n");
|
|
|
|
goto out_fail;
|
|
|
|
}
|
2016-04-14 04:50:18 +08:00
|
|
|
|
2019-10-10 03:14:59 +08:00
|
|
|
ret = drm_dp_dpcd_writeb(mgr->aux, DP_MSTM_CTRL,
|
|
|
|
DP_MST_EN |
|
|
|
|
DP_UP_REQ_EN |
|
|
|
|
DP_UPSTREAM_IS_SRC);
|
|
|
|
if (ret < 0) {
|
|
|
|
DRM_DEBUG_KMS("mst write failed - undocked during suspend?\n");
|
|
|
|
goto out_fail;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-10 03:14:59 +08:00
|
|
|
/* Some hubs forget their guids after they resume */
|
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_GUID, guid, 16);
|
|
|
|
if (ret != 16) {
|
|
|
|
DRM_DEBUG_KMS("dpcd read failed - undocked during suspend?\n");
|
|
|
|
goto out_fail;
|
|
|
|
}
|
|
|
|
drm_dp_check_mstb_guid(mgr->mst_primary, guid);
|
2016-04-14 04:50:18 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
/*
|
|
|
|
* For the final step of resuming the topology, we need to bring the
|
|
|
|
* state of our in-memory topology back into sync with reality. So,
|
|
|
|
* restart the probing process as if we're probing a new hub
|
|
|
|
*/
|
|
|
|
queue_work(system_long_wq, &mgr->work);
|
2019-10-10 03:14:59 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
2016-04-14 04:50:18 +08:00
|
|
|
|
drm/dp_mst: Add basic topology reprobing when resuming
Finally! For a very long time, our MST helpers have had one very
annoying issue: They don't know how to reprobe the topology state when
coming out of suspend. This means that if a user has a machine connected
to an MST topology and decides to suspend their machine, we lose all
topology changes that happened during that period. That can be a big
problem if the machine was connected to a different topology on the same
port before resuming, as we won't bother reprobing any of the ports and
likely cause the user's monitors not to come back up as expected.
So, we start fixing this by teaching our MST helpers how to reprobe the
link addresses of each connected topology when resuming. As it turns
out, the behavior that we want here is identical to the behavior we want
when initially probing a newly connected MST topology, with a couple of
important differences:
- We need to be more careful about handling the potential races between
events from the MST hub that could change the topology state as we're
performing the link address reprobe
- We need to be more careful about handling unlikely state changes on
ports - such as an input port turning into an output port, something
that would be far more likely to happen in situations like the MST hub
we're connected to being changed while we're suspend
Both of which have been solved by previous commits. That leaves one
requirement:
- We need to prune any MST ports in our in-memory topology state that
were present when suspending, but have not appeared in the post-resume
link address response from their parent branch device
Which we can now handle in this commit by modifying
drm_dp_send_link_address(). We then introduce suspend/resume reprobing
by introducing drm_dp_mst_topology_mgr_invalidate_mstb(), which we call
in drm_dp_mst_topology_mgr_suspend() to traverse the in-memory topology
state to indicate that each mstb needs it's link address resent and PBN
resources reprobed.
On resume, we start back up &mgr->work and have it reprobe the topology
in the same way we would on a hotplug, removing any leftover ports that
no longer appear in the topology state.
Changes since v4:
* Split indenting changes in drm_dp_mst_topology_mgr_resume() into a
separate patch
* Only fire hotplugs when something has actually changed after a link
address probe
* Don't try to change port->connector at all on ports, just throw out
ports that need their connectors removed to make things easier.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-14-lyude@redhat.com
2019-06-18 07:57:33 +08:00
|
|
|
if (sync) {
|
|
|
|
DRM_DEBUG_KMS("Waiting for link probe work to finish re-syncing topology...\n");
|
|
|
|
flush_work(&mgr->work);
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-10 03:14:59 +08:00
|
|
|
return 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-10-10 03:14:59 +08:00
|
|
|
out_fail:
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_unlock(&mgr->lock);
|
2019-10-10 03:14:59 +08:00
|
|
|
return -1;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_topology_mgr_resume);
|
|
|
|
|
2017-07-19 21:46:32 +08:00
|
|
|
static bool drm_dp_get_one_sb_msg(struct drm_dp_mst_topology_mgr *mgr, bool up)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int len;
|
|
|
|
u8 replyblock[32];
|
|
|
|
int replylen, origlen, curreply;
|
|
|
|
int ret;
|
|
|
|
struct drm_dp_sideband_msg_rx *msg;
|
|
|
|
int basereg = up ? DP_SIDEBAND_MSG_UP_REQ_BASE : DP_SIDEBAND_MSG_DOWN_REP_BASE;
|
|
|
|
msg = up ? &mgr->up_req_recv : &mgr->down_rep_recv;
|
|
|
|
|
|
|
|
len = min(mgr->max_dpcd_transaction_bytes, 16);
|
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, basereg,
|
|
|
|
replyblock, len);
|
|
|
|
if (ret != len) {
|
|
|
|
DRM_DEBUG_KMS("failed to read DPCD down rep %d %d\n", len, ret);
|
2017-07-19 21:46:32 +08:00
|
|
|
return false;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
ret = drm_dp_sideband_msg_build(msg, replyblock, len, true);
|
|
|
|
if (!ret) {
|
|
|
|
DRM_DEBUG_KMS("sideband msg build failed %d\n", replyblock[0]);
|
2017-07-19 21:46:32 +08:00
|
|
|
return false;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
replylen = msg->curchunk_len + msg->curchunk_hdrlen;
|
|
|
|
|
|
|
|
origlen = replylen;
|
|
|
|
replylen -= len;
|
|
|
|
curreply = len;
|
|
|
|
while (replylen > 0) {
|
|
|
|
len = min3(replylen, mgr->max_dpcd_transaction_bytes, 16);
|
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, basereg + curreply,
|
|
|
|
replyblock, len);
|
|
|
|
if (ret != len) {
|
2017-07-19 19:43:28 +08:00
|
|
|
DRM_DEBUG_KMS("failed to read a chunk (len %d, ret %d)\n",
|
|
|
|
len, ret);
|
2017-07-19 21:46:32 +08:00
|
|
|
return false;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2017-07-19 19:43:28 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = drm_dp_sideband_msg_build(msg, replyblock, len, false);
|
2017-07-19 19:43:28 +08:00
|
|
|
if (!ret) {
|
2014-06-05 12:01:32 +08:00
|
|
|
DRM_DEBUG_KMS("failed to build sideband msg\n");
|
2017-07-19 21:46:32 +08:00
|
|
|
return false;
|
2017-07-19 19:43:28 +08:00
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
curreply += len;
|
|
|
|
replylen -= len;
|
|
|
|
}
|
2017-07-19 21:46:32 +08:00
|
|
|
return true;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static int drm_dp_mst_handle_down_rep(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
2019-09-04 04:45:51 +08:00
|
|
|
struct drm_dp_sideband_msg_tx *txmsg;
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
struct drm_dp_sideband_msg_hdr *hdr = &mgr->down_rep_recv.initial_hdr;
|
|
|
|
int slot = -1;
|
|
|
|
|
|
|
|
if (!drm_dp_get_one_sb_msg(mgr, false))
|
|
|
|
goto clear_down_rep_recv;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
if (!mgr->down_rep_recv.have_eomt)
|
2017-07-19 21:46:32 +08:00
|
|
|
return 0;
|
2019-09-04 04:45:51 +08:00
|
|
|
|
|
|
|
mstb = drm_dp_get_mst_branch_device(mgr, hdr->lct, hdr->rad);
|
|
|
|
if (!mstb) {
|
|
|
|
DRM_DEBUG_KMS("Got MST reply from unknown device %d\n",
|
|
|
|
hdr->lct);
|
|
|
|
goto clear_down_rep_recv;
|
2017-07-19 21:46:32 +08:00
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
/* find the message */
|
|
|
|
slot = hdr->seqno;
|
|
|
|
mutex_lock(&mgr->qlock);
|
|
|
|
txmsg = mstb->tx_slots[slot];
|
|
|
|
/* remove from slots */
|
|
|
|
mutex_unlock(&mgr->qlock);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
if (!txmsg) {
|
|
|
|
DRM_DEBUG_KMS("Got MST reply with no msg %p %d %d %02x %02x\n",
|
|
|
|
mstb, hdr->seqno, hdr->lct, hdr->rad[0],
|
|
|
|
mgr->down_rep_recv.msg[0]);
|
|
|
|
goto no_msg;
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
drm_dp_sideband_parse_reply(&mgr->down_rep_recv, &txmsg->reply);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK)
|
|
|
|
DRM_DEBUG_KMS("Got NAK reply: req 0x%02x (%s), reason 0x%02x (%s), nak data 0x%02x\n",
|
|
|
|
txmsg->reply.req_type,
|
|
|
|
drm_dp_mst_req_type_str(txmsg->reply.req_type),
|
|
|
|
txmsg->reply.u.nak.reason,
|
|
|
|
drm_dp_mst_nak_reason_str(txmsg->reply.u.nak.reason),
|
|
|
|
txmsg->reply.u.nak.nak_data);
|
2019-01-23 04:03:00 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
memset(&mgr->down_rep_recv, 0, sizeof(struct drm_dp_sideband_msg_rx));
|
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
mutex_lock(&mgr->qlock);
|
|
|
|
txmsg->state = DRM_DP_SIDEBAND_TX_RX;
|
|
|
|
mstb->tx_slots[slot] = NULL;
|
2020-01-13 17:36:49 +08:00
|
|
|
mgr->is_waiting_for_dwn_reply = false;
|
2019-09-04 04:45:51 +08:00
|
|
|
mutex_unlock(&mgr->qlock);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:51 +08:00
|
|
|
wake_up_all(&mgr->tx_waitq);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
|
|
|
no_msg:
|
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
|
|
|
clear_down_rep_recv:
|
2020-01-13 17:36:49 +08:00
|
|
|
mutex_lock(&mgr->qlock);
|
|
|
|
mgr->is_waiting_for_dwn_reply = false;
|
|
|
|
mutex_unlock(&mgr->qlock);
|
2019-09-04 04:45:51 +08:00
|
|
|
memset(&mgr->down_rep_recv, 0, sizeof(struct drm_dp_sideband_msg_rx));
|
|
|
|
|
|
|
|
return 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
static inline bool
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
drm_dp_mst_process_up_req(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_pending_up_req *up_req)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_branch *mstb = NULL;
|
|
|
|
struct drm_dp_sideband_msg_req_body *msg = &up_req->msg;
|
|
|
|
struct drm_dp_sideband_msg_hdr *hdr = &up_req->hdr;
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
bool hotplug = false;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
|
|
|
|
if (hdr->broadcast) {
|
|
|
|
const u8 *guid = NULL;
|
|
|
|
|
|
|
|
if (msg->req_type == DP_CONNECTION_STATUS_NOTIFY)
|
|
|
|
guid = msg->u.conn_stat.guid;
|
|
|
|
else if (msg->req_type == DP_RESOURCE_STATUS_NOTIFY)
|
|
|
|
guid = msg->u.resource_stat.guid;
|
|
|
|
|
|
|
|
mstb = drm_dp_get_mst_branch_device_by_guid(mgr, guid);
|
|
|
|
} else {
|
|
|
|
mstb = drm_dp_get_mst_branch_device(mgr, hdr->lct, hdr->rad);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!mstb) {
|
|
|
|
DRM_DEBUG_KMS("Got MST reply from unknown device %d\n",
|
|
|
|
hdr->lct);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
return false;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/* TODO: Add missing handler for DP_RESOURCE_STATUS_NOTIFY events */
|
|
|
|
if (msg->req_type == DP_CONNECTION_STATUS_NOTIFY) {
|
|
|
|
drm_dp_mst_handle_conn_stat(mstb, &msg->u.conn_stat);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
hotplug = true;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
return hotplug;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_mst_up_req_work(struct work_struct *work)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr =
|
|
|
|
container_of(work, struct drm_dp_mst_topology_mgr,
|
|
|
|
up_req_work);
|
|
|
|
struct drm_dp_pending_up_req *up_req;
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
bool send_hotplug = false;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
|
2019-10-17 04:02:59 +08:00
|
|
|
mutex_lock(&mgr->probe_lock);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
while (true) {
|
|
|
|
mutex_lock(&mgr->up_req_lock);
|
|
|
|
up_req = list_first_entry_or_null(&mgr->up_req_list,
|
|
|
|
struct drm_dp_pending_up_req,
|
|
|
|
next);
|
|
|
|
if (up_req)
|
|
|
|
list_del(&up_req->next);
|
|
|
|
mutex_unlock(&mgr->up_req_lock);
|
|
|
|
|
|
|
|
if (!up_req)
|
|
|
|
break;
|
|
|
|
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
send_hotplug |= drm_dp_mst_process_up_req(mgr, up_req);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
kfree(up_req);
|
|
|
|
}
|
2019-10-17 04:02:59 +08:00
|
|
|
mutex_unlock(&mgr->probe_lock);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
|
|
|
|
if (send_hotplug)
|
|
|
|
drm_kms_helper_hotplug_event(mgr->dev);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static int drm_dp_mst_handle_up_req(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
2019-09-04 04:45:50 +08:00
|
|
|
struct drm_dp_sideband_msg_hdr *hdr = &mgr->up_req_recv.initial_hdr;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
struct drm_dp_pending_up_req *up_req;
|
2019-09-04 04:45:48 +08:00
|
|
|
bool seqno;
|
2017-07-19 21:46:32 +08:00
|
|
|
|
2019-09-04 04:45:50 +08:00
|
|
|
if (!drm_dp_get_one_sb_msg(mgr, true))
|
|
|
|
goto out;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:48 +08:00
|
|
|
if (!mgr->up_req_recv.have_eomt)
|
|
|
|
return 0;
|
2015-12-19 06:14:42 +08:00
|
|
|
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
up_req = kzalloc(sizeof(*up_req), GFP_KERNEL);
|
|
|
|
if (!up_req) {
|
|
|
|
DRM_ERROR("Not enough memory to process MST up req\n");
|
|
|
|
return -ENOMEM;
|
2019-09-04 04:45:48 +08:00
|
|
|
}
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
INIT_LIST_HEAD(&up_req->next);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-09-04 04:45:50 +08:00
|
|
|
seqno = hdr->seqno;
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
drm_dp_sideband_parse_req(&mgr->up_req_recv, &up_req->msg);
|
2015-12-19 06:14:42 +08:00
|
|
|
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
if (up_req->msg.req_type != DP_CONNECTION_STATUS_NOTIFY &&
|
|
|
|
up_req->msg.req_type != DP_RESOURCE_STATUS_NOTIFY) {
|
|
|
|
DRM_DEBUG_KMS("Received unknown up req type, ignoring: %x\n",
|
|
|
|
up_req->msg.req_type);
|
|
|
|
kfree(up_req);
|
2019-09-04 04:45:50 +08:00
|
|
|
goto out;
|
|
|
|
}
|
2016-01-23 06:07:28 +08:00
|
|
|
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
drm_dp_send_up_ack_reply(mgr, mgr->mst_primary, up_req->msg.req_type,
|
|
|
|
seqno, false);
|
|
|
|
|
|
|
|
if (up_req->msg.req_type == DP_CONNECTION_STATUS_NOTIFY) {
|
|
|
|
const struct drm_dp_connection_status_notify *conn_stat =
|
|
|
|
&up_req->msg.u.conn_stat;
|
2016-02-17 09:36:38 +08:00
|
|
|
|
2019-09-04 04:45:50 +08:00
|
|
|
DRM_DEBUG_KMS("Got CSN: pn: %d ldps:%d ddps: %d mcs: %d ip: %d pdt: %d\n",
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
conn_stat->port_number,
|
|
|
|
conn_stat->legacy_device_plug_status,
|
|
|
|
conn_stat->displayport_device_plug_status,
|
|
|
|
conn_stat->message_capability_status,
|
|
|
|
conn_stat->input_port,
|
|
|
|
conn_stat->peer_device_type);
|
|
|
|
} else if (up_req->msg.req_type == DP_RESOURCE_STATUS_NOTIFY) {
|
|
|
|
const struct drm_dp_resource_status_notify *res_stat =
|
|
|
|
&up_req->msg.u.resource_stat;
|
2015-12-19 06:14:42 +08:00
|
|
|
|
2019-09-04 04:45:50 +08:00
|
|
|
DRM_DEBUG_KMS("Got RSN: pn: %d avail_pbn %d\n",
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
res_stat->port_number,
|
|
|
|
res_stat->available_pbn);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2019-09-04 04:45:48 +08:00
|
|
|
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
up_req->hdr = *hdr;
|
|
|
|
mutex_lock(&mgr->up_req_lock);
|
|
|
|
list_add_tail(&up_req->next, &mgr->up_req_list);
|
|
|
|
mutex_unlock(&mgr->up_req_lock);
|
|
|
|
queue_work(system_long_wq, &mgr->up_req_work);
|
|
|
|
|
2019-09-04 04:45:50 +08:00
|
|
|
out:
|
2019-09-04 04:45:48 +08:00
|
|
|
memset(&mgr->up_req_recv, 0, sizeof(struct drm_dp_sideband_msg_rx));
|
|
|
|
return 0;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_hpd_irq() - MST hotplug IRQ notify
|
|
|
|
* @mgr: manager to notify irq for.
|
|
|
|
* @esi: 4 bytes from SINK_COUNT_ESI
|
2014-07-30 20:23:44 +08:00
|
|
|
* @handled: whether the hpd interrupt was consumed or not
|
2014-06-05 12:01:32 +08:00
|
|
|
*
|
|
|
|
* This should be called from the driver when it detects a short IRQ,
|
|
|
|
* along with the value of the DEVICE_SERVICE_IRQ_VECTOR_ESI0. The
|
|
|
|
* topology manager will process the sideband messages received as a result
|
|
|
|
* of this.
|
|
|
|
*/
|
|
|
|
int drm_dp_mst_hpd_irq(struct drm_dp_mst_topology_mgr *mgr, u8 *esi, bool *handled)
|
|
|
|
{
|
|
|
|
int ret = 0;
|
|
|
|
int sc;
|
|
|
|
*handled = false;
|
|
|
|
sc = esi[0] & 0x3f;
|
|
|
|
|
|
|
|
if (sc != mgr->sink_count) {
|
|
|
|
mgr->sink_count = sc;
|
|
|
|
*handled = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (esi[1] & DP_DOWN_REP_MSG_RDY) {
|
|
|
|
ret = drm_dp_mst_handle_down_rep(mgr);
|
|
|
|
*handled = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (esi[1] & DP_UP_REQ_MSG_RDY) {
|
|
|
|
ret |= drm_dp_mst_handle_up_req(mgr);
|
|
|
|
*handled = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
drm_dp_mst_kick_tx(mgr);
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_hpd_irq);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_detect_port() - get connection status for an MST port
|
2016-07-16 03:48:04 +08:00
|
|
|
* @connector: DRM connector for this port
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
* @ctx: The acquisition context to use for grabbing locks
|
2014-06-05 12:01:32 +08:00
|
|
|
* @mgr: manager for this port
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
* @port: pointer to a port
|
2014-06-05 12:01:32 +08:00
|
|
|
*
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
* This returns the current connection state for a port.
|
2014-06-05 12:01:32 +08:00
|
|
|
*/
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
int
|
|
|
|
drm_dp_mst_detect_port(struct drm_connector *connector,
|
|
|
|
struct drm_modeset_acquire_ctx *ctx,
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
int ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-02-02 09:23:26 +08:00
|
|
|
/* we need to search for the port in the mgr in case it's gone */
|
2019-01-11 08:53:28 +08:00
|
|
|
port = drm_dp_mst_topology_get_port_validated(mgr, port);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!port)
|
|
|
|
return connector_status_disconnected;
|
|
|
|
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
ret = drm_modeset_lock(&mgr->base.lock, ctx);
|
|
|
|
if (ret)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
ret = connector_status_disconnected;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!port->ddps)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
switch (port->pdt) {
|
|
|
|
case DP_PEER_DEVICE_NONE:
|
|
|
|
case DP_PEER_DEVICE_MST_BRANCHING:
|
2020-01-17 14:03:50 +08:00
|
|
|
if (!port->mcs)
|
|
|
|
ret = connector_status_connected;
|
2014-06-05 12:01:32 +08:00
|
|
|
break;
|
|
|
|
|
|
|
|
case DP_PEER_DEVICE_SST_SINK:
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
ret = connector_status_connected;
|
2016-02-17 09:36:38 +08:00
|
|
|
/* for logical ports - cache the EDID */
|
|
|
|
if (port->port_num >= 8 && !port->cached_edid) {
|
|
|
|
port->cached_edid = drm_get_edid(connector, &port->aux.ddc);
|
|
|
|
}
|
2014-06-05 12:01:32 +08:00
|
|
|
break;
|
|
|
|
case DP_PEER_DEVICE_DP_LEGACY_CONV:
|
|
|
|
if (port->ldps)
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
ret = connector_status_connected;
|
2014-06-05 12:01:32 +08:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
out:
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
drm/dp_mst: Protect drm_dp_mst_port members with locking
This is a complicated one. Essentially, there's currently a problem in the MST
core that hasn't really caused any issues that we're aware of (emphasis on "that
we're aware of"): locking.
When we go through and probe the link addresses and path resources in a
topology, we hold no locks when updating ports with said information. The
members I'm referring to in particular are:
- ldps
- ddps
- mcs
- pdt
- dpcd_rev
- num_sdp_streams
- num_sdp_stream_sinks
- available_pbn
- input
- connector
Now that we're handling UP requests asynchronously and will be using some of
the struct members mentioned above in atomic modesetting in the future for
features such as PBN validation, this is going to become a lot more important.
As well, the next few commits that prepare us for and introduce suspend/resume
reprobing will also need clear locking in order to prevent from additional
racing hilarities that we never could have hit in the past.
So, let's solve this issue by using &mgr->base.lock, the modesetting
lock which currently only protects &mgr->base.state. This works
perfectly because it allows us to avoid blocking connection_mutex
unnecessarily, and we can grab this in connector detection paths since
it's a ww mutex. We start by having drm_dp_mst_handle_up_req() hold this
when updating ports. For drm_dp_mst_handle_link_address_port() things
are a bit more complicated. As I've learned the hard way, we can grab
&mgr->lock.base for everything except for port->connector. See, our
normal driver probing paths end up generating this rather obvious
lockdep chain:
&drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
However, sysfs grabs &drm->mode_config.mutex in order to protect itself
from connector state changing under it. Because this entails grabbing
kn->count, e.g. the lock that the kernel provides for protecting sysfs
contexts, we end up grabbing kn->count followed by
&drm->mode_config.mutex. This ends up creating an extremely rude chain:
&kn->count
-> &drm->mode_config.mutex
-> crtc_ww_class_mutex/crtc_ww_class_acquire
-> &connector->mutex
I mean, look at that thing! It's just evil!!! This gross thing ends up
making any calls to drm_connector_register()/drm_connector_unregister()
impossible when holding any kind of modesetting lock. This is annoying
because ideally, we always want to ensure that
drm_dp_mst_port->connector never changes when doing an atomic commit or
check that would affect the atomic topology state so that it can
reliably and easily be used from future DRM DP MST helpers to assist
with tasks such as scanning through the current VCPI allocations and
adding connectors which need to have their allocations updated in
response to a bandwidth change or the like.
Being able to hold &mgr->base.lock throughout the entire link probe
process would have been _great_, since we could prevent userspace from
ever seeing any states in-between individual port changes and as a
result likely end up with a much faster probe and more consistent
results from said probes. But without some rework of how we handle
connector probing in sysfs it's not at all currently possible. In the
future, maybe we can try using the sysfs locks to protect updates to
connector probing state and fix this mess.
So for now, to protect everything other than port->connector under
&mgr->base.lock and ensure that we still have the guarantee that atomic
check/commit contexts will never see port->connector change we use a
silly trick. See: port->connector only needs to change in order to
ensure that input ports (see the MST spec) never have a ghost connector
associated with them. But, there's nothing stopping us from simply
throwing the entire port out and creating a new one in order to maintain
that requirement while still keeping port->connector consistent across
the lifetime of the port in atomic check/commit contexts. For all
intended purposes this works fine, as we validate ports in any contexts
we care about before using them and as such will end up reporting the
connector as disconnected until it's port's destruction finalizes. So,
we just do that in cases where we detect port->input has transitioned
from true->false. We don't need to worry about the other direction,
since a port without a connector isn't visible to userspace and as such
doesn't need to be protected by &mgr->base.lock until we finish
registering a connector for it.
For updating members of drm_dp_mst_port other than port->connector, we
simply grab &mgr->base.lock in drm_dp_mst_link_probe_work() for already
registered ports, update said members and drop the lock before
potentially registering a connector and probing the link address of it's
children.
Finally, we modify drm_dp_mst_detect_port() to take a modesetting lock
acquisition context in order to acquire &mgr->base.lock under
&connection_mutex and convert all it's users over to using the
.detect_ctx probe hooks.
With that, we finally have well defined locking.
Changes since v4:
* Get rid of port->mutex, stop using connection_mutex and just use our own
modesetting lock - mgr->base.lock. Also, add a probe_lock that comes
before this patch.
* Just throw out ports that get changed from an output to an input, and
replace them with new ports. This lets us ensure that modesetting
contexts never see port->connector go from having a connector to being
NULL.
* Write an extremely detailed explanation of what problems this is
trying to fix, since there's a _lot_ of context here and I honestly
forgot some of it myself a couple times.
* Don't grab mgr->lock when reading port->mstb in
drm_dp_mst_handle_link_address_port(). It's not needed.
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-7-lyude@redhat.com
2019-06-18 05:59:29 +08:00
|
|
|
return ret;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_detect_port);
|
|
|
|
|
2015-12-02 14:09:43 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_port_has_audio() - Check whether port has audio capability or not
|
|
|
|
* @mgr: manager for this port
|
|
|
|
* @port: unverified pointer to a port.
|
|
|
|
*
|
|
|
|
* This returns whether the port supports audio or not.
|
|
|
|
*/
|
|
|
|
bool drm_dp_mst_port_has_audio(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
bool ret = false;
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
port = drm_dp_mst_topology_get_port_validated(mgr, port);
|
2015-12-02 14:09:43 +08:00
|
|
|
if (!port)
|
|
|
|
return ret;
|
|
|
|
ret = port->has_audio;
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2015-12-02 14:09:43 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_port_has_audio);
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_get_edid() - get EDID for an MST port
|
|
|
|
* @connector: toplevel connector to get EDID for
|
|
|
|
* @mgr: manager for this port
|
|
|
|
* @port: unverified pointer to a port.
|
|
|
|
*
|
|
|
|
* This returns an EDID for the port connected to a connector,
|
|
|
|
* It validates the pointer still exists so the caller doesn't require a
|
|
|
|
* reference.
|
|
|
|
*/
|
|
|
|
struct edid *drm_dp_mst_get_edid(struct drm_connector *connector, struct drm_dp_mst_topology_mgr *mgr, struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
struct edid *edid = NULL;
|
|
|
|
|
2019-02-02 09:23:26 +08:00
|
|
|
/* we need to search for the port in the mgr in case it's gone */
|
2019-01-11 08:53:28 +08:00
|
|
|
port = drm_dp_mst_topology_get_port_validated(mgr, port);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!port)
|
|
|
|
return NULL;
|
|
|
|
|
2014-10-20 14:28:02 +08:00
|
|
|
if (port->cached_edid)
|
|
|
|
edid = drm_edid_duplicate(port->cached_edid);
|
2016-02-17 09:36:38 +08:00
|
|
|
else {
|
|
|
|
edid = drm_get_edid(connector, &port->aux.ddc);
|
|
|
|
}
|
2015-12-02 14:09:43 +08:00
|
|
|
port->has_audio = drm_detect_monitor_audio(edid);
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
return edid;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_get_edid);
|
|
|
|
|
|
|
|
/**
|
2018-10-24 07:12:46 +08:00
|
|
|
* drm_dp_find_vcpi_slots() - Find VCPI slots for this PBN value
|
2014-06-05 12:01:32 +08:00
|
|
|
* @mgr: manager to use
|
|
|
|
* @pbn: payload bandwidth to convert into slots.
|
2018-10-24 07:12:46 +08:00
|
|
|
*
|
|
|
|
* Calculate the number of VCPI slots that will be required for the given PBN
|
|
|
|
* value. This function is deprecated, and should not be used in atomic
|
|
|
|
* drivers.
|
|
|
|
*
|
|
|
|
* RETURNS:
|
|
|
|
* The total slots required for this port, or error.
|
2014-06-05 12:01:32 +08:00
|
|
|
*/
|
|
|
|
int drm_dp_find_vcpi_slots(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
int pbn)
|
|
|
|
{
|
|
|
|
int num_slots;
|
|
|
|
|
|
|
|
num_slots = DIV_ROUND_UP(pbn, mgr->pbn_div);
|
|
|
|
|
2017-03-16 15:10:25 +08:00
|
|
|
/* max. time slots - one slot for MTP header */
|
|
|
|
if (num_slots > 63)
|
2014-06-05 12:01:32 +08:00
|
|
|
return -ENOSPC;
|
|
|
|
return num_slots;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_find_vcpi_slots);
|
|
|
|
|
|
|
|
static int drm_dp_init_vcpi(struct drm_dp_mst_topology_mgr *mgr,
|
2017-03-16 15:10:26 +08:00
|
|
|
struct drm_dp_vcpi *vcpi, int pbn, int slots)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
2017-03-16 15:10:25 +08:00
|
|
|
/* max. time slots - one slot for MTP header */
|
2017-03-16 15:10:26 +08:00
|
|
|
if (slots > 63)
|
2014-06-05 12:01:32 +08:00
|
|
|
return -ENOSPC;
|
|
|
|
|
|
|
|
vcpi->pbn = pbn;
|
2017-03-16 15:10:26 +08:00
|
|
|
vcpi->aligned_pbn = slots * mgr->pbn_div;
|
|
|
|
vcpi->num_slots = slots;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
ret = drm_dp_mst_assign_payload_id(mgr, vcpi);
|
|
|
|
if (ret < 0)
|
|
|
|
return ret;
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2017-04-21 13:51:32 +08:00
|
|
|
/**
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
* drm_dp_atomic_find_vcpi_slots() - Find and add VCPI slots to the state
|
2017-04-21 13:51:32 +08:00
|
|
|
* @state: global atomic state
|
|
|
|
* @mgr: MST topology manager for the port
|
|
|
|
* @port: port to find vcpi slots for
|
|
|
|
* @pbn: bandwidth required for the mode in PBN
|
|
|
|
*
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
* Allocates VCPI slots to @port, replacing any previous VCPI allocations it
|
|
|
|
* may have had. Any atomic drivers which support MST must call this function
|
|
|
|
* in their &drm_encoder_helper_funcs.atomic_check() callback to change the
|
|
|
|
* current VCPI allocation for the new state, but only when
|
|
|
|
* &drm_crtc_state.mode_changed or &drm_crtc_state.connectors_changed is set
|
|
|
|
* to ensure compatibility with userspace applications that still use the
|
|
|
|
* legacy modesetting UAPI.
|
|
|
|
*
|
|
|
|
* Allocations set by this function are not checked against the bandwidth
|
|
|
|
* restraints of @mgr until the driver calls drm_dp_mst_atomic_check().
|
|
|
|
*
|
|
|
|
* Additionally, it is OK to call this function multiple times on the same
|
|
|
|
* @port as needed. It is not OK however, to call this function and
|
|
|
|
* drm_dp_atomic_release_vcpi_slots() in the same atomic check phase.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_atomic_release_vcpi_slots()
|
|
|
|
* drm_dp_mst_atomic_check()
|
|
|
|
*
|
|
|
|
* Returns:
|
|
|
|
* Total slots in the atomic state assigned for this port, or a negative error
|
|
|
|
* code if the port no longer exists
|
2017-04-21 13:51:32 +08:00
|
|
|
*/
|
|
|
|
int drm_dp_atomic_find_vcpi_slots(struct drm_atomic_state *state,
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port, int pbn)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_state *topology_state;
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
struct drm_dp_vcpi_allocation *pos, *vcpi = NULL;
|
2019-10-19 15:18:40 +08:00
|
|
|
int prev_slots, req_slots;
|
2017-04-21 13:51:32 +08:00
|
|
|
|
|
|
|
topology_state = drm_atomic_get_mst_topology_state(state, mgr);
|
2017-07-12 23:51:00 +08:00
|
|
|
if (IS_ERR(topology_state))
|
|
|
|
return PTR_ERR(topology_state);
|
2017-04-21 13:51:32 +08:00
|
|
|
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
/* Find the current allocation for this port, if any */
|
|
|
|
list_for_each_entry(pos, &topology_state->vcpis, next) {
|
|
|
|
if (pos->port == port) {
|
|
|
|
vcpi = pos;
|
|
|
|
prev_slots = vcpi->vcpi;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This should never happen, unless the driver tries
|
|
|
|
* releasing and allocating the same VCPI allocation,
|
|
|
|
* which is an error
|
|
|
|
*/
|
|
|
|
if (WARN_ON(!prev_slots)) {
|
|
|
|
DRM_ERROR("cannot allocate and release VCPI on [MST PORT:%p] in the same state\n",
|
|
|
|
port);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
break;
|
|
|
|
}
|
2017-04-21 13:51:32 +08:00
|
|
|
}
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
if (!vcpi)
|
|
|
|
prev_slots = 0;
|
|
|
|
|
|
|
|
req_slots = DIV_ROUND_UP(pbn, mgr->pbn_div);
|
|
|
|
|
|
|
|
DRM_DEBUG_ATOMIC("[CONNECTOR:%d:%s] [MST PORT:%p] VCPI %d -> %d\n",
|
|
|
|
port->connector->base.id, port->connector->name,
|
|
|
|
port, prev_slots, req_slots);
|
|
|
|
|
|
|
|
/* Add the new allocation to the state */
|
|
|
|
if (!vcpi) {
|
|
|
|
vcpi = kzalloc(sizeof(*vcpi), GFP_KERNEL);
|
2019-02-02 08:20:02 +08:00
|
|
|
if (!vcpi)
|
|
|
|
return -ENOMEM;
|
2017-04-21 13:51:32 +08:00
|
|
|
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
drm_dp_mst_get_port_malloc(port);
|
|
|
|
vcpi->port = port;
|
|
|
|
list_add(&vcpi->next, &topology_state->vcpis);
|
|
|
|
}
|
|
|
|
vcpi->vcpi = req_slots;
|
2017-04-21 13:51:32 +08:00
|
|
|
|
2019-10-19 15:18:40 +08:00
|
|
|
return req_slots;
|
2017-04-21 13:51:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_atomic_find_vcpi_slots);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_atomic_release_vcpi_slots() - Release allocated vcpi slots
|
|
|
|
* @state: global atomic state
|
|
|
|
* @mgr: MST topology manager for the port
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
* @port: The port to release the VCPI slots from
|
2017-04-21 13:51:32 +08:00
|
|
|
*
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
* Releases any VCPI slots that have been allocated to a port in the atomic
|
|
|
|
* state. Any atomic drivers which support MST must call this function in
|
|
|
|
* their &drm_connector_helper_funcs.atomic_check() callback when the
|
2019-02-02 09:23:26 +08:00
|
|
|
* connector will no longer have VCPI allocated (e.g. because its CRTC was
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
* removed) when it had VCPI allocated in the previous atomic state.
|
|
|
|
*
|
|
|
|
* It is OK to call this even if @port has been removed from the system.
|
|
|
|
* Additionally, it is OK to call this function multiple times on the same
|
|
|
|
* @port as needed. It is not OK however, to call this function and
|
|
|
|
* drm_dp_atomic_find_vcpi_slots() on the same @port in a single atomic check
|
|
|
|
* phase.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_atomic_find_vcpi_slots()
|
|
|
|
* drm_dp_mst_atomic_check()
|
|
|
|
*
|
|
|
|
* Returns:
|
|
|
|
* 0 if all slots for this port were added back to
|
|
|
|
* &drm_dp_mst_topology_state.avail_slots or negative error code
|
2017-04-21 13:51:32 +08:00
|
|
|
*/
|
|
|
|
int drm_dp_atomic_release_vcpi_slots(struct drm_atomic_state *state,
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr,
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
struct drm_dp_mst_port *port)
|
2017-04-21 13:51:32 +08:00
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_state *topology_state;
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
struct drm_dp_vcpi_allocation *pos;
|
|
|
|
bool found = false;
|
2017-04-21 13:51:32 +08:00
|
|
|
|
|
|
|
topology_state = drm_atomic_get_mst_topology_state(state, mgr);
|
2017-07-12 23:51:00 +08:00
|
|
|
if (IS_ERR(topology_state))
|
|
|
|
return PTR_ERR(topology_state);
|
2017-04-21 13:51:32 +08:00
|
|
|
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
list_for_each_entry(pos, &topology_state->vcpis, next) {
|
|
|
|
if (pos->port == port) {
|
|
|
|
found = true;
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (WARN_ON(!found)) {
|
|
|
|
DRM_ERROR("no VCPI for [MST PORT:%p] found in mst state %p\n",
|
|
|
|
port, &topology_state->base);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
|
|
|
|
|
|
|
DRM_DEBUG_ATOMIC("[MST PORT:%p] VCPI %d -> 0\n", port, pos->vcpi);
|
|
|
|
if (pos->vcpi) {
|
|
|
|
drm_dp_mst_put_port_malloc(port);
|
|
|
|
pos->vcpi = 0;
|
|
|
|
}
|
2017-04-21 13:51:32 +08:00
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_atomic_release_vcpi_slots);
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_allocate_vcpi() - Allocate a virtual channel
|
|
|
|
* @mgr: manager for this port
|
|
|
|
* @port: port to allocate a virtual channel for.
|
|
|
|
* @pbn: payload bandwidth number to request
|
|
|
|
* @slots: returned number of slots for this PBN.
|
|
|
|
*/
|
2017-03-16 15:10:26 +08:00
|
|
|
bool drm_dp_mst_allocate_vcpi(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port, int pbn, int slots)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
|
|
|
int ret;
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
port = drm_dp_mst_topology_get_port_validated(mgr, port);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!port)
|
|
|
|
return false;
|
|
|
|
|
2017-03-16 15:10:26 +08:00
|
|
|
if (slots < 0)
|
|
|
|
return false;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
if (port->vcpi.vcpi > 0) {
|
2019-01-11 08:53:26 +08:00
|
|
|
DRM_DEBUG_KMS("payload: vcpi %d already allocated for pbn %d - requested pbn %d\n",
|
|
|
|
port->vcpi.vcpi, port->vcpi.pbn, pbn);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (pbn == port->vcpi.pbn) {
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
return true;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2017-03-16 15:10:26 +08:00
|
|
|
ret = drm_dp_init_vcpi(mgr, &port->vcpi, pbn, slots);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (ret) {
|
2017-03-16 15:10:25 +08:00
|
|
|
DRM_DEBUG_KMS("failed to init vcpi slots=%d max=63 ret=%d\n",
|
2019-01-11 08:53:26 +08:00
|
|
|
DIV_ROUND_UP(pbn, mgr->pbn_div), ret);
|
2014-06-05 12:01:32 +08:00
|
|
|
goto out;
|
|
|
|
}
|
2017-03-16 15:10:25 +08:00
|
|
|
DRM_DEBUG_KMS("initing vcpi for pbn=%d slots=%d\n",
|
2019-01-11 08:53:26 +08:00
|
|
|
pbn, port->vcpi.num_slots);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2019-02-02 09:23:26 +08:00
|
|
|
/* Keep port allocated until its payload has been removed */
|
2019-01-11 08:53:32 +08:00
|
|
|
drm_dp_mst_get_port_malloc(port);
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
return true;
|
|
|
|
out:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_allocate_vcpi);
|
|
|
|
|
2015-02-24 07:23:55 +08:00
|
|
|
int drm_dp_mst_get_vcpi_slots(struct drm_dp_mst_topology_mgr *mgr, struct drm_dp_mst_port *port)
|
|
|
|
{
|
|
|
|
int slots = 0;
|
2019-01-11 08:53:28 +08:00
|
|
|
port = drm_dp_mst_topology_get_port_validated(mgr, port);
|
2015-02-24 07:23:55 +08:00
|
|
|
if (!port)
|
|
|
|
return slots;
|
|
|
|
|
|
|
|
slots = port->vcpi.num_slots;
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_port(port);
|
2015-02-24 07:23:55 +08:00
|
|
|
return slots;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_get_vcpi_slots);
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_reset_vcpi_slots() - Reset number of slots to 0 for VCPI
|
|
|
|
* @mgr: manager for this port
|
|
|
|
* @port: unverified pointer to a port.
|
|
|
|
*
|
|
|
|
* This just resets the number of slots for the ports VCPI for later programming.
|
|
|
|
*/
|
|
|
|
void drm_dp_mst_reset_vcpi_slots(struct drm_dp_mst_topology_mgr *mgr, struct drm_dp_mst_port *port)
|
|
|
|
{
|
2019-01-11 08:53:32 +08:00
|
|
|
/*
|
2019-02-02 09:23:26 +08:00
|
|
|
* A port with VCPI will remain allocated until its VCPI is
|
2019-01-11 08:53:32 +08:00
|
|
|
* released, no verified ref needed
|
|
|
|
*/
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
port->vcpi.num_slots = 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_reset_vcpi_slots);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_deallocate_vcpi() - deallocate a VCPI
|
|
|
|
* @mgr: manager for this port
|
2019-02-02 08:20:01 +08:00
|
|
|
* @port: port to deallocate vcpi for
|
|
|
|
*
|
|
|
|
* This can be called unconditionally, regardless of whether
|
|
|
|
* drm_dp_mst_allocate_vcpi() succeeded or not.
|
2014-06-05 12:01:32 +08:00
|
|
|
*/
|
2019-01-11 08:53:27 +08:00
|
|
|
void drm_dp_mst_deallocate_vcpi(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port)
|
2014-06-05 12:01:32 +08:00
|
|
|
{
|
2019-02-02 08:20:01 +08:00
|
|
|
if (!port->vcpi.vcpi)
|
|
|
|
return;
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
drm_dp_mst_put_payload_id(mgr, port->vcpi.vcpi);
|
|
|
|
port->vcpi.num_slots = 0;
|
|
|
|
port->vcpi.pbn = 0;
|
|
|
|
port->vcpi.aligned_pbn = 0;
|
|
|
|
port->vcpi.vcpi = 0;
|
2019-01-11 08:53:32 +08:00
|
|
|
drm_dp_mst_put_port_malloc(port);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_deallocate_vcpi);
|
|
|
|
|
|
|
|
static int drm_dp_dpcd_write_payload(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
int id, struct drm_dp_payload *payload)
|
|
|
|
{
|
|
|
|
u8 payload_alloc[3], status;
|
|
|
|
int ret;
|
|
|
|
int retries = 0;
|
|
|
|
|
|
|
|
drm_dp_dpcd_writeb(mgr->aux, DP_PAYLOAD_TABLE_UPDATE_STATUS,
|
|
|
|
DP_PAYLOAD_TABLE_UPDATED);
|
|
|
|
|
|
|
|
payload_alloc[0] = id;
|
|
|
|
payload_alloc[1] = payload->start_slot;
|
|
|
|
payload_alloc[2] = payload->num_slots;
|
|
|
|
|
|
|
|
ret = drm_dp_dpcd_write(mgr->aux, DP_PAYLOAD_ALLOCATE_SET, payload_alloc, 3);
|
|
|
|
if (ret != 3) {
|
|
|
|
DRM_DEBUG_KMS("failed to write payload allocation %d\n", ret);
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
|
|
|
|
retry:
|
|
|
|
ret = drm_dp_dpcd_readb(mgr->aux, DP_PAYLOAD_TABLE_UPDATE_STATUS, &status);
|
|
|
|
if (ret < 0) {
|
|
|
|
DRM_DEBUG_KMS("failed to read payload table status %d\n", ret);
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (!(status & DP_PAYLOAD_TABLE_UPDATED)) {
|
|
|
|
retries++;
|
|
|
|
if (retries < 20) {
|
|
|
|
usleep_range(10000, 20000);
|
|
|
|
goto retry;
|
|
|
|
}
|
|
|
|
DRM_DEBUG_KMS("status not set after read payload table status %d\n", status);
|
|
|
|
ret = -EINVAL;
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
ret = 0;
|
|
|
|
fail:
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_check_act_status() - Check ACT handled status.
|
|
|
|
* @mgr: manager to use
|
|
|
|
*
|
|
|
|
* Check the payload status bits in the DPCD for ACT handled completion.
|
|
|
|
*/
|
|
|
|
int drm_dp_check_act_status(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
u8 status;
|
|
|
|
int ret;
|
|
|
|
int count = 0;
|
|
|
|
|
|
|
|
do {
|
|
|
|
ret = drm_dp_dpcd_readb(mgr->aux, DP_PAYLOAD_TABLE_UPDATE_STATUS, &status);
|
|
|
|
|
|
|
|
if (ret < 0) {
|
|
|
|
DRM_DEBUG_KMS("failed to read payload table status %d\n", ret);
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (status & DP_PAYLOAD_ACT_HANDLED)
|
|
|
|
break;
|
|
|
|
count++;
|
|
|
|
udelay(100);
|
|
|
|
|
|
|
|
} while (count < 30);
|
|
|
|
|
|
|
|
if (!(status & DP_PAYLOAD_ACT_HANDLED)) {
|
|
|
|
DRM_DEBUG_KMS("failed to get ACT bit %d after %d retries\n", status, count);
|
|
|
|
ret = -EINVAL;
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
return 0;
|
|
|
|
fail:
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_check_act_status);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_calc_pbn_mode() - Calculate the PBN for a mode.
|
|
|
|
* @clock: dot clock for the mode
|
|
|
|
* @bpp: bpp for the mode.
|
|
|
|
*
|
|
|
|
* This uses the formula in the spec to calculate the PBN value for a mode.
|
|
|
|
*/
|
|
|
|
int drm_dp_calc_pbn_mode(int clock, int bpp)
|
|
|
|
{
|
2016-01-23 06:07:26 +08:00
|
|
|
/*
|
|
|
|
* margin 5300ppm + 300ppm ~ 0.6% as per spec, factor is 1.006
|
|
|
|
* The unit of 54/64Mbytes/sec is an arbitrary unit chosen based on
|
|
|
|
* common multiplier to render an integer PBN for all link rate/lane
|
|
|
|
* counts combinations
|
|
|
|
* calculate
|
|
|
|
* peak_kbps *= (1006/1000)
|
|
|
|
* peak_kbps *= (64/54)
|
|
|
|
* peak_kbps *= 8 convert to bytes
|
|
|
|
*/
|
2019-09-25 22:14:42 +08:00
|
|
|
return DIV_ROUND_UP_ULL(mul_u32_u32(clock * bpp, 64 * 1006),
|
|
|
|
8 * 54 * 1000 * 1000);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_calc_pbn_mode);
|
|
|
|
|
|
|
|
/* we want to kick the TX after we've ack the up/down IRQs. */
|
|
|
|
static void drm_dp_mst_kick_tx(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
queue_work(system_long_wq, &mgr->tx_work);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_mst_dump_mstb(struct seq_file *m,
|
|
|
|
struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_port *port;
|
|
|
|
int tabs = mstb->lct;
|
|
|
|
char prefix[10];
|
|
|
|
int i;
|
|
|
|
|
|
|
|
for (i = 0; i < tabs; i++)
|
|
|
|
prefix[i] = '\t';
|
|
|
|
prefix[i] = '\0';
|
|
|
|
|
|
|
|
seq_printf(m, "%smst: %p, %d\n", prefix, mstb, mstb->num_ports);
|
|
|
|
list_for_each_entry(port, &mstb->ports, next) {
|
2016-04-15 01:18:36 +08:00
|
|
|
seq_printf(m, "%sport: %d: input: %d: pdt: %d, ddps: %d ldps: %d, sdp: %d/%d, %p, conn: %p\n", prefix, port->port_num, port->input, port->pdt, port->ddps, port->ldps, port->num_sdp_streams, port->num_sdp_stream_sinks, port, port->connector);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (port->mstb)
|
|
|
|
drm_dp_mst_dump_mstb(m, port->mstb);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2018-03-19 22:19:32 +08:00
|
|
|
#define DP_PAYLOAD_TABLE_SIZE 64
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
static bool dump_dp_payload_table(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
char *buf)
|
|
|
|
{
|
|
|
|
int i;
|
2017-05-31 07:35:37 +08:00
|
|
|
|
2018-03-19 22:19:32 +08:00
|
|
|
for (i = 0; i < DP_PAYLOAD_TABLE_SIZE; i += 16) {
|
2017-05-31 07:35:37 +08:00
|
|
|
if (drm_dp_dpcd_read(mgr->aux,
|
|
|
|
DP_PAYLOAD_TABLE_UPDATE_STATUS + i,
|
|
|
|
&buf[i], 16) != 16)
|
|
|
|
return false;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
2017-05-31 07:35:37 +08:00
|
|
|
return true;
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
2016-04-15 01:18:36 +08:00
|
|
|
static void fetch_monitor_name(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_port *port, char *name,
|
|
|
|
int namelen)
|
|
|
|
{
|
|
|
|
struct edid *mst_edid;
|
|
|
|
|
|
|
|
mst_edid = drm_dp_mst_get_edid(port->connector, mgr, port);
|
|
|
|
drm_edid_get_monitor_name(mst_edid, name, namelen);
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_dump_topology(): dump topology to seq file.
|
|
|
|
* @m: seq_file to dump output to
|
|
|
|
* @mgr: manager to dump current topology for.
|
|
|
|
*
|
|
|
|
* helper to dump MST topology to a seq file for debugfs.
|
|
|
|
*/
|
|
|
|
void drm_dp_mst_dump_topology(struct seq_file *m,
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
struct drm_dp_mst_port *port;
|
2016-04-15 01:18:36 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
if (mgr->mst_primary)
|
|
|
|
drm_dp_mst_dump_mstb(m, mgr->mst_primary);
|
|
|
|
|
|
|
|
/* dump VCPIs */
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
|
|
|
|
mutex_lock(&mgr->payload_lock);
|
2016-04-15 01:18:36 +08:00
|
|
|
seq_printf(m, "vcpi: %lx %lx %d\n", mgr->payload_mask, mgr->vcpi_mask,
|
|
|
|
mgr->max_payloads);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
|
|
|
for (i = 0; i < mgr->max_payloads; i++) {
|
|
|
|
if (mgr->proposed_vcpis[i]) {
|
2016-04-15 01:18:36 +08:00
|
|
|
char name[14];
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
port = container_of(mgr->proposed_vcpis[i], struct drm_dp_mst_port, vcpi);
|
2016-04-15 01:18:36 +08:00
|
|
|
fetch_monitor_name(mgr, port, name, sizeof(name));
|
|
|
|
seq_printf(m, "vcpi %d: %d %d %d sink name: %s\n", i,
|
|
|
|
port->port_num, port->vcpi.vcpi,
|
|
|
|
port->vcpi.num_slots,
|
|
|
|
(*name != 0) ? name : "Unknown");
|
2014-06-05 12:01:32 +08:00
|
|
|
} else
|
2016-04-15 01:18:36 +08:00
|
|
|
seq_printf(m, "vcpi %d:unused\n", i);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
for (i = 0; i < mgr->max_payloads; i++) {
|
|
|
|
seq_printf(m, "payload %d: %d, %d, %d\n",
|
|
|
|
i,
|
|
|
|
mgr->payloads[i].payload_state,
|
|
|
|
mgr->payloads[i].start_slot,
|
|
|
|
mgr->payloads[i].num_slots);
|
|
|
|
|
|
|
|
|
|
|
|
}
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
if (mgr->mst_primary) {
|
2018-03-19 22:19:32 +08:00
|
|
|
u8 buf[DP_PAYLOAD_TABLE_SIZE];
|
2014-06-05 12:01:32 +08:00
|
|
|
int ret;
|
2017-05-31 07:35:37 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_DPCD_REV, buf, DP_RECEIVER_CAP_SIZE);
|
2017-05-31 07:35:37 +08:00
|
|
|
seq_printf(m, "dpcd: %*ph\n", DP_RECEIVER_CAP_SIZE, buf);
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_FAUX_CAP, buf, 2);
|
2017-05-31 07:35:37 +08:00
|
|
|
seq_printf(m, "faux/mst: %*ph\n", 2, buf);
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_MSTM_CTRL, buf, 1);
|
2017-05-31 07:35:37 +08:00
|
|
|
seq_printf(m, "mst ctrl: %*ph\n", 1, buf);
|
2014-06-05 12:01:32 +08:00
|
|
|
|
2015-07-14 09:33:31 +08:00
|
|
|
/* dump the standard OUI branch header */
|
|
|
|
ret = drm_dp_dpcd_read(mgr->aux, DP_BRANCH_OUI, buf, DP_BRANCH_OUI_HEADER_SIZE);
|
2017-05-31 07:35:37 +08:00
|
|
|
seq_printf(m, "branch oui: %*phN devid: ", 3, buf);
|
2016-04-15 01:18:36 +08:00
|
|
|
for (i = 0x3; i < 0x8 && buf[i]; i++)
|
2015-07-14 09:33:31 +08:00
|
|
|
seq_printf(m, "%c", buf[i]);
|
2017-05-31 07:35:37 +08:00
|
|
|
seq_printf(m, " revision: hw: %x.%x sw: %x.%x\n",
|
|
|
|
buf[0x9] >> 4, buf[0x9] & 0xf, buf[0xa], buf[0xb]);
|
|
|
|
if (dump_dp_payload_table(mgr, buf))
|
2018-03-19 22:19:32 +08:00
|
|
|
seq_printf(m, "payload table: %*ph\n", DP_PAYLOAD_TABLE_SIZE, buf);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_dump_topology);
|
|
|
|
|
|
|
|
static void drm_dp_tx_work(struct work_struct *work)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr = container_of(work, struct drm_dp_mst_topology_mgr, tx_work);
|
|
|
|
|
|
|
|
mutex_lock(&mgr->qlock);
|
2020-01-13 17:36:49 +08:00
|
|
|
if (!list_empty(&mgr->tx_msg_downq) && !mgr->is_waiting_for_dwn_reply)
|
2014-06-05 12:01:32 +08:00
|
|
|
process_single_down_tx_qlock(mgr);
|
|
|
|
mutex_unlock(&mgr->qlock);
|
|
|
|
}
|
|
|
|
|
2019-02-20 06:41:02 +08:00
|
|
|
static inline void
|
|
|
|
drm_dp_delayed_destroy_port(struct drm_dp_mst_port *port)
|
2015-06-15 08:34:28 +08:00
|
|
|
{
|
2019-03-11 06:59:10 +08:00
|
|
|
if (port->connector)
|
|
|
|
port->mgr->cbs->destroy_connector(port->mgr, port->connector);
|
2019-02-20 06:41:02 +08:00
|
|
|
|
2020-01-17 14:03:50 +08:00
|
|
|
drm_dp_port_set_pdt(port, DP_PEER_DEVICE_NONE, port->mcs);
|
2019-02-20 06:41:02 +08:00
|
|
|
drm_dp_mst_put_port_malloc(port);
|
|
|
|
}
|
|
|
|
|
|
|
|
static inline void
|
|
|
|
drm_dp_delayed_destroy_mstb(struct drm_dp_mst_branch *mstb)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr = mstb->mgr;
|
|
|
|
struct drm_dp_mst_port *port, *tmp;
|
|
|
|
bool wake_tx = false;
|
|
|
|
|
|
|
|
mutex_lock(&mgr->lock);
|
|
|
|
list_for_each_entry_safe(port, tmp, &mstb->ports, next) {
|
|
|
|
list_del(&port->next);
|
|
|
|
drm_dp_mst_topology_put_port(port);
|
|
|
|
}
|
|
|
|
mutex_unlock(&mgr->lock);
|
|
|
|
|
|
|
|
/* drop any tx slots msg */
|
|
|
|
mutex_lock(&mstb->mgr->qlock);
|
|
|
|
if (mstb->tx_slots[0]) {
|
|
|
|
mstb->tx_slots[0]->state = DRM_DP_SIDEBAND_TX_TIMEOUT;
|
|
|
|
mstb->tx_slots[0] = NULL;
|
|
|
|
wake_tx = true;
|
|
|
|
}
|
|
|
|
if (mstb->tx_slots[1]) {
|
|
|
|
mstb->tx_slots[1]->state = DRM_DP_SIDEBAND_TX_TIMEOUT;
|
|
|
|
mstb->tx_slots[1] = NULL;
|
|
|
|
wake_tx = true;
|
|
|
|
}
|
|
|
|
mutex_unlock(&mstb->mgr->qlock);
|
|
|
|
|
|
|
|
if (wake_tx)
|
|
|
|
wake_up_all(&mstb->mgr->tx_waitq);
|
|
|
|
|
|
|
|
drm_dp_mst_put_mstb_malloc(mstb);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void drm_dp_delayed_destroy_work(struct work_struct *work)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr =
|
|
|
|
container_of(work, struct drm_dp_mst_topology_mgr,
|
|
|
|
delayed_destroy_work);
|
|
|
|
bool send_hotplug = false, go_again;
|
|
|
|
|
2015-06-15 08:34:28 +08:00
|
|
|
/*
|
|
|
|
* Not a regular list traverse as we have to drop the destroy
|
2019-02-20 06:41:02 +08:00
|
|
|
* connector lock before destroying the mstb/port, to avoid AB->BA
|
2015-06-15 08:34:28 +08:00
|
|
|
* ordering between this lock and the config mutex.
|
|
|
|
*/
|
2019-02-20 06:41:02 +08:00
|
|
|
do {
|
|
|
|
go_again = false;
|
|
|
|
|
|
|
|
for (;;) {
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
|
|
|
|
mutex_lock(&mgr->delayed_destroy_lock);
|
|
|
|
mstb = list_first_entry_or_null(&mgr->destroy_branch_device_list,
|
|
|
|
struct drm_dp_mst_branch,
|
|
|
|
destroy_next);
|
|
|
|
if (mstb)
|
|
|
|
list_del(&mstb->destroy_next);
|
|
|
|
mutex_unlock(&mgr->delayed_destroy_lock);
|
|
|
|
|
|
|
|
if (!mstb)
|
|
|
|
break;
|
|
|
|
|
|
|
|
drm_dp_delayed_destroy_mstb(mstb);
|
|
|
|
go_again = true;
|
2015-06-15 08:34:28 +08:00
|
|
|
}
|
|
|
|
|
2019-02-20 06:41:02 +08:00
|
|
|
for (;;) {
|
|
|
|
struct drm_dp_mst_port *port;
|
2015-08-11 15:54:29 +08:00
|
|
|
|
2019-02-20 06:41:02 +08:00
|
|
|
mutex_lock(&mgr->delayed_destroy_lock);
|
|
|
|
port = list_first_entry_or_null(&mgr->destroy_port_list,
|
|
|
|
struct drm_dp_mst_port,
|
|
|
|
next);
|
|
|
|
if (port)
|
|
|
|
list_del(&port->next);
|
|
|
|
mutex_unlock(&mgr->delayed_destroy_lock);
|
|
|
|
|
|
|
|
if (!port)
|
|
|
|
break;
|
|
|
|
|
|
|
|
drm_dp_delayed_destroy_port(port);
|
|
|
|
send_hotplug = true;
|
|
|
|
go_again = true;
|
|
|
|
}
|
|
|
|
} while (go_again);
|
2015-08-11 15:54:29 +08:00
|
|
|
|
2015-09-16 08:37:28 +08:00
|
|
|
if (send_hotplug)
|
2018-11-29 06:12:34 +08:00
|
|
|
drm_kms_helper_hotplug_event(mgr->dev);
|
2015-06-15 08:34:28 +08:00
|
|
|
}
|
|
|
|
|
2017-07-12 23:51:02 +08:00
|
|
|
static struct drm_private_state *
|
|
|
|
drm_dp_mst_duplicate_state(struct drm_private_obj *obj)
|
2017-04-21 13:51:31 +08:00
|
|
|
{
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
struct drm_dp_mst_topology_state *state, *old_state =
|
|
|
|
to_dp_mst_topology_state(obj->state);
|
|
|
|
struct drm_dp_vcpi_allocation *pos, *vcpi;
|
2017-04-21 13:51:31 +08:00
|
|
|
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
state = kmemdup(old_state, sizeof(*state), GFP_KERNEL);
|
2017-07-12 23:51:02 +08:00
|
|
|
if (!state)
|
2017-04-21 13:51:31 +08:00
|
|
|
return NULL;
|
|
|
|
|
2017-07-12 23:51:02 +08:00
|
|
|
__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
|
2017-04-21 13:51:31 +08:00
|
|
|
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
INIT_LIST_HEAD(&state->vcpis);
|
|
|
|
|
|
|
|
list_for_each_entry(pos, &old_state->vcpis, next) {
|
|
|
|
/* Prune leftover freed VCPI allocations */
|
|
|
|
if (!pos->vcpi)
|
|
|
|
continue;
|
|
|
|
|
|
|
|
vcpi = kmemdup(pos, sizeof(*vcpi), GFP_KERNEL);
|
|
|
|
if (!vcpi)
|
|
|
|
goto fail;
|
|
|
|
|
|
|
|
drm_dp_mst_get_port_malloc(vcpi->port);
|
|
|
|
list_add(&vcpi->next, &state->vcpis);
|
|
|
|
}
|
|
|
|
|
2017-07-12 23:51:02 +08:00
|
|
|
return &state->base;
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
|
|
|
|
fail:
|
|
|
|
list_for_each_entry_safe(pos, vcpi, &state->vcpis, next) {
|
|
|
|
drm_dp_mst_put_port_malloc(pos->port);
|
|
|
|
kfree(pos);
|
|
|
|
}
|
|
|
|
kfree(state);
|
|
|
|
|
|
|
|
return NULL;
|
2017-04-21 13:51:31 +08:00
|
|
|
}
|
|
|
|
|
2017-07-12 23:51:02 +08:00
|
|
|
static void drm_dp_mst_destroy_state(struct drm_private_obj *obj,
|
|
|
|
struct drm_private_state *state)
|
2017-04-21 13:51:31 +08:00
|
|
|
{
|
2017-07-12 23:51:02 +08:00
|
|
|
struct drm_dp_mst_topology_state *mst_state =
|
|
|
|
to_dp_mst_topology_state(state);
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
struct drm_dp_vcpi_allocation *pos, *tmp;
|
|
|
|
|
|
|
|
list_for_each_entry_safe(pos, tmp, &mst_state->vcpis, next) {
|
|
|
|
/* We only keep references to ports with non-zero VCPIs */
|
|
|
|
if (pos->vcpi)
|
|
|
|
drm_dp_mst_put_port_malloc(pos->port);
|
|
|
|
kfree(pos);
|
|
|
|
}
|
2017-07-12 23:51:02 +08:00
|
|
|
|
|
|
|
kfree(mst_state);
|
2017-04-21 13:51:31 +08:00
|
|
|
}
|
|
|
|
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
static inline int
|
|
|
|
drm_dp_mst_atomic_check_topology_state(struct drm_dp_mst_topology_mgr *mgr,
|
|
|
|
struct drm_dp_mst_topology_state *mst_state)
|
|
|
|
{
|
|
|
|
struct drm_dp_vcpi_allocation *vcpi;
|
2019-01-11 08:53:42 +08:00
|
|
|
int avail_slots = 63, payload_count = 0;
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
|
|
|
|
list_for_each_entry(vcpi, &mst_state->vcpis, next) {
|
|
|
|
/* Releasing VCPI is always OK-even if the port is gone */
|
|
|
|
if (!vcpi->vcpi) {
|
|
|
|
DRM_DEBUG_ATOMIC("[MST PORT:%p] releases all VCPI slots\n",
|
|
|
|
vcpi->port);
|
|
|
|
continue;
|
|
|
|
}
|
|
|
|
|
|
|
|
DRM_DEBUG_ATOMIC("[MST PORT:%p] requires %d vcpi slots\n",
|
|
|
|
vcpi->port, vcpi->vcpi);
|
|
|
|
|
|
|
|
avail_slots -= vcpi->vcpi;
|
|
|
|
if (avail_slots < 0) {
|
|
|
|
DRM_DEBUG_ATOMIC("[MST PORT:%p] not enough VCPI slots in mst state %p (avail=%d)\n",
|
|
|
|
vcpi->port, mst_state,
|
|
|
|
avail_slots + vcpi->vcpi);
|
|
|
|
return -ENOSPC;
|
|
|
|
}
|
2019-01-11 08:53:42 +08:00
|
|
|
|
|
|
|
if (++payload_count > mgr->max_payloads) {
|
|
|
|
DRM_DEBUG_ATOMIC("[MST MGR:%p] state %p has too many payloads (max=%d)\n",
|
|
|
|
mgr, mst_state, mgr->max_payloads);
|
|
|
|
return -EINVAL;
|
|
|
|
}
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
}
|
|
|
|
DRM_DEBUG_ATOMIC("[MST MGR:%p] mst state %p VCPI avail=%d used=%d\n",
|
|
|
|
mgr, mst_state, avail_slots,
|
|
|
|
63 - avail_slots);
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_atomic_check - Check that the new state of an MST topology in an
|
|
|
|
* atomic update is valid
|
|
|
|
* @state: Pointer to the new &struct drm_dp_mst_topology_state
|
|
|
|
*
|
|
|
|
* Checks the given topology state for an atomic update to ensure that it's
|
|
|
|
* valid. This includes checking whether there's enough bandwidth to support
|
|
|
|
* the new VCPI allocations in the atomic update.
|
|
|
|
*
|
|
|
|
* Any atomic drivers supporting DP MST must make sure to call this after
|
|
|
|
* checking the rest of their state in their
|
|
|
|
* &drm_mode_config_funcs.atomic_check() callback.
|
|
|
|
*
|
|
|
|
* See also:
|
|
|
|
* drm_dp_atomic_find_vcpi_slots()
|
|
|
|
* drm_dp_atomic_release_vcpi_slots()
|
|
|
|
*
|
|
|
|
* Returns:
|
|
|
|
*
|
|
|
|
* 0 if the new state is valid, negative error code otherwise.
|
|
|
|
*/
|
|
|
|
int drm_dp_mst_atomic_check(struct drm_atomic_state *state)
|
|
|
|
{
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr;
|
|
|
|
struct drm_dp_mst_topology_state *mst_state;
|
|
|
|
int i, ret = 0;
|
|
|
|
|
|
|
|
for_each_new_mst_mgr_in_state(state, mgr, mst_state, i) {
|
|
|
|
ret = drm_dp_mst_atomic_check_topology_state(mgr, mst_state);
|
|
|
|
if (ret)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_atomic_check);
|
|
|
|
|
2019-01-11 08:53:40 +08:00
|
|
|
const struct drm_private_state_funcs drm_dp_mst_topology_state_funcs = {
|
2017-07-12 23:51:02 +08:00
|
|
|
.atomic_duplicate_state = drm_dp_mst_duplicate_state,
|
|
|
|
.atomic_destroy_state = drm_dp_mst_destroy_state,
|
2017-04-21 13:51:31 +08:00
|
|
|
};
|
2019-01-11 08:53:40 +08:00
|
|
|
EXPORT_SYMBOL(drm_dp_mst_topology_state_funcs);
|
2017-04-21 13:51:31 +08:00
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_atomic_get_mst_topology_state: get MST topology state
|
|
|
|
*
|
|
|
|
* @state: global atomic state
|
|
|
|
* @mgr: MST topology manager, also the private object in this case
|
|
|
|
*
|
|
|
|
* This function wraps drm_atomic_get_priv_obj_state() passing in the MST atomic
|
|
|
|
* state vtable so that the private object state returned is that of a MST
|
|
|
|
* topology object. Also, drm_atomic_get_private_obj_state() expects the caller
|
|
|
|
* to care of the locking, so warn if don't hold the connection_mutex.
|
|
|
|
*
|
|
|
|
* RETURNS:
|
|
|
|
*
|
|
|
|
* The MST topology state or error pointer.
|
|
|
|
*/
|
|
|
|
struct drm_dp_mst_topology_state *drm_atomic_get_mst_topology_state(struct drm_atomic_state *state,
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
2017-07-12 23:51:02 +08:00
|
|
|
return to_dp_mst_topology_state(drm_atomic_get_private_obj_state(state, &mgr->base));
|
2017-04-21 13:51:31 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_atomic_get_mst_topology_state);
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_mgr_init - initialise a topology manager
|
|
|
|
* @mgr: manager struct to initialise
|
|
|
|
* @dev: device providing this structure - for i2c addition.
|
|
|
|
* @aux: DP helper aux channel to talk to this device
|
|
|
|
* @max_dpcd_transaction_bytes: hw specific DPCD transaction limit
|
|
|
|
* @max_payloads: maximum number of payloads this GPU can source
|
|
|
|
* @conn_base_id: the connector object ID the MST device is connected to.
|
|
|
|
*
|
|
|
|
* Return 0 for success, or negative error code on failure
|
|
|
|
*/
|
|
|
|
int drm_dp_mst_topology_mgr_init(struct drm_dp_mst_topology_mgr *mgr,
|
2017-01-25 07:49:29 +08:00
|
|
|
struct drm_device *dev, struct drm_dp_aux *aux,
|
2014-06-05 12:01:32 +08:00
|
|
|
int max_dpcd_transaction_bytes,
|
|
|
|
int max_payloads, int conn_base_id)
|
|
|
|
{
|
2017-07-12 23:51:02 +08:00
|
|
|
struct drm_dp_mst_topology_state *mst_state;
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_init(&mgr->lock);
|
|
|
|
mutex_init(&mgr->qlock);
|
|
|
|
mutex_init(&mgr->payload_lock);
|
2019-02-20 06:41:02 +08:00
|
|
|
mutex_init(&mgr->delayed_destroy_lock);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
mutex_init(&mgr->up_req_lock);
|
2019-10-17 04:02:59 +08:00
|
|
|
mutex_init(&mgr->probe_lock);
|
2019-06-21 05:59:25 +08:00
|
|
|
#if IS_ENABLED(CONFIG_DRM_DEBUG_DP_MST_TOPOLOGY_REFS)
|
|
|
|
mutex_init(&mgr->topology_ref_history_lock);
|
|
|
|
#endif
|
2014-06-05 12:01:32 +08:00
|
|
|
INIT_LIST_HEAD(&mgr->tx_msg_downq);
|
2019-02-20 06:41:02 +08:00
|
|
|
INIT_LIST_HEAD(&mgr->destroy_port_list);
|
|
|
|
INIT_LIST_HEAD(&mgr->destroy_branch_device_list);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
INIT_LIST_HEAD(&mgr->up_req_list);
|
2014-06-05 12:01:32 +08:00
|
|
|
INIT_WORK(&mgr->work, drm_dp_mst_link_probe_work);
|
|
|
|
INIT_WORK(&mgr->tx_work, drm_dp_tx_work);
|
2019-02-20 06:41:02 +08:00
|
|
|
INIT_WORK(&mgr->delayed_destroy_work, drm_dp_delayed_destroy_work);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
INIT_WORK(&mgr->up_req_work, drm_dp_mst_up_req_work);
|
2014-06-05 12:01:32 +08:00
|
|
|
init_waitqueue_head(&mgr->tx_waitq);
|
|
|
|
mgr->dev = dev;
|
|
|
|
mgr->aux = aux;
|
|
|
|
mgr->max_dpcd_transaction_bytes = max_dpcd_transaction_bytes;
|
|
|
|
mgr->max_payloads = max_payloads;
|
|
|
|
mgr->conn_base_id = conn_base_id;
|
2016-01-29 20:44:29 +08:00
|
|
|
if (max_payloads + 1 > sizeof(mgr->payload_mask) * 8 ||
|
|
|
|
max_payloads + 1 > sizeof(mgr->vcpi_mask) * 8)
|
|
|
|
return -EINVAL;
|
2014-06-05 12:01:32 +08:00
|
|
|
mgr->payloads = kcalloc(max_payloads, sizeof(struct drm_dp_payload), GFP_KERNEL);
|
|
|
|
if (!mgr->payloads)
|
|
|
|
return -ENOMEM;
|
|
|
|
mgr->proposed_vcpis = kcalloc(max_payloads, sizeof(struct drm_dp_vcpi *), GFP_KERNEL);
|
|
|
|
if (!mgr->proposed_vcpis)
|
|
|
|
return -ENOMEM;
|
|
|
|
set_bit(0, &mgr->payload_mask);
|
2016-01-29 20:44:28 +08:00
|
|
|
|
2017-07-12 23:51:02 +08:00
|
|
|
mst_state = kzalloc(sizeof(*mst_state), GFP_KERNEL);
|
|
|
|
if (mst_state == NULL)
|
2017-04-21 13:51:31 +08:00
|
|
|
return -ENOMEM;
|
2017-07-12 23:51:02 +08:00
|
|
|
|
|
|
|
mst_state->mgr = mgr;
|
drm/dp_mst: Start tracking per-port VCPI allocations
There has been a TODO waiting for quite a long time in
drm_dp_mst_topology.c:
/* We cannot rely on port->vcpi.num_slots to update
* topology_state->avail_slots as the port may not exist if the parent
* branch device was unplugged. This should be fixed by tracking
* per-port slot allocation in drm_dp_mst_topology_state instead of
* depending on the caller to tell us how many slots to release.
*/
That's not the only reason we should fix this: forcing the driver to
track the VCPI allocations throughout a state's atomic check is
error prone, because it means that extra care has to be taken with the
order that drm_dp_atomic_find_vcpi_slots() and
drm_dp_atomic_release_vcpi_slots() are called in in order to ensure
idempotency. Currently the only driver actually using these helpers,
i915, doesn't even do this correctly: multiple ->best_encoder() checks
with i915's current implementation would not be idempotent and would
over-allocate VCPI slots, something I learned trying to implement
fallback retraining in MST.
So: simplify this whole mess, and teach drm_dp_atomic_find_vcpi_slots()
and drm_dp_atomic_release_vcpi_slots() to track the VCPI allocations for
each port. This allows us to ensure idempotency without having to rely
on the driver as much. Additionally: the driver doesn't need to do any
kind of VCPI slot tracking anymore if it doesn't need it for it's own
internal state.
Additionally; this adds a new drm_dp_mst_atomic_check() helper which
must be used by atomic drivers to perform validity checks for the new
VCPI allocations incurred by a state.
Also: update the documentation and make it more obvious that these
/must/ be called by /all/ atomic drivers supporting MST.
Changes since v9:
* Add some missing changes that were requested by danvet that I forgot
about after I redid all of the kref stuff:
* Remove unnecessary state changes in intel_dp_mst_atomic_check
* Cleanup atomic check logic for VCPI allocations - all we need to check in
compute_config is whether or not this state disables a CRTC, then free
VCPI based off that
Changes since v8:
* Fix compile errors, whoops!
Changes since v7:
- Don't check for mixed stale/valid VCPI allocations, just rely on
connector registration to stop such erroneous modesets
Changes since v6:
- Keep a kref to all of the ports we have allocations on. This required
a good bit of changing to when we call drm_dp_find_vcpi_slots(),
mainly that we need to ensure that we only redo VCPI allocations on
actual mode or CRTC changes, not crtc_state->active changes.
Additionally, we no longer take the registration of the DRM connector
for each port into account because so long as we have a kref to the
port in the new or previous atomic state, the connector will stay
registered.
- Use the small changes to drm_dp_put_port() to add even more error
checking to make misusage of the helpers more obvious. I added this
after having to chase down various use-after-free conditions that
started popping up from the new helpers so no one else has to
troubleshoot that.
- Move some accidental DRM_DEBUG_KMS() calls to DRM_DEBUG_ATOMIC()
- Update documentation again, note that find/release() should both not be
called on the same port in a single atomic check phase (but multiple
calls to one or the other is OK)
Changes since v4:
- Don't skip the atomic checks for VCPI allocations if no new VCPI
allocations happen in a state. This makes the next change I'm about
to list here a lot easier to implement.
- Don't ignore VCPI allocations on destroyed ports, instead ensure that
when ports are destroyed and still have VCPI allocations in the
topology state, the only state changes allowed are releasing said
ports' VCPI. This prevents a state with a mix of VCPI allocations
from destroyed ports, and allocations from valid ports.
Changes since v3:
- Don't release VCPI allocations in the topology state immediately in
drm_dp_atomic_release_vcpi_slots(), instead mark them as 0 and skip
over them in drm_dp_mst_duplicate_state(). This makes it so
drm_dp_atomic_release_vcpi_slots() is still idempotent while also
throwing warnings if the driver messes up it's book keeping and tries
to release VCPI slots on a port that doesn't have any pre-existing
VCPI allocation - danvet
- Change mst_state/state in some debugging messages to "mst state"
Changes since v2:
- Use kmemdup() for duplicating MST state - danvet
- Move port validation out of duplicate state callback - danvet
- Handle looping through MST topology states in
drm_dp_mst_atomic_check() so the driver doesn't have to do it
- Fix documentation in drm_dp_atomic_find_vcpi_slots()
- Move the atomic check for each individual topology state into it's
own function, reduces indenting
- Don't consider "stale" MST ports when calculating the bandwidth
requirements. This is needed because originally we relied on the
state duplication functions to prune any stale ports from the new
state, which would prevent us from incorrectly considering their
bandwidth requirements alongside legitimate new payloads.
- Add function references in drm_dp_atomic_release_vcpi_slots() - danvet
- Annotate atomic VCPI and atomic check functions with __must_check
- danvet
Changes since v1:
- Don't use the now-removed ->atomic_check() for private objects hook,
just give drivers a function to call themselves
Signed-off-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Cc: David Airlie <airlied@redhat.com>
Cc: Jerry Zuo <Jerry.Zuo@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Juston Li <juston.li@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190111005343.17443-19-lyude@redhat.com
2019-01-11 08:53:41 +08:00
|
|
|
INIT_LIST_HEAD(&mst_state->vcpis);
|
2017-07-12 23:51:02 +08:00
|
|
|
|
2018-10-22 20:31:22 +08:00
|
|
|
drm_atomic_private_obj_init(dev, &mgr->base,
|
2017-07-12 23:51:02 +08:00
|
|
|
&mst_state->base,
|
2019-01-11 08:53:40 +08:00
|
|
|
&drm_dp_mst_topology_state_funcs);
|
2017-04-21 13:51:31 +08:00
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_topology_mgr_init);
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_topology_mgr_destroy() - destroy topology manager.
|
|
|
|
* @mgr: manager to destroy
|
|
|
|
*/
|
|
|
|
void drm_dp_mst_topology_mgr_destroy(struct drm_dp_mst_topology_mgr *mgr)
|
|
|
|
{
|
2018-12-12 07:50:26 +08:00
|
|
|
drm_dp_mst_topology_mgr_set_mst(mgr, false);
|
2015-09-30 08:39:42 +08:00
|
|
|
flush_work(&mgr->work);
|
2019-02-20 06:41:02 +08:00
|
|
|
cancel_work_sync(&mgr->delayed_destroy_work);
|
2014-06-05 12:01:32 +08:00
|
|
|
mutex_lock(&mgr->payload_lock);
|
|
|
|
kfree(mgr->payloads);
|
|
|
|
mgr->payloads = NULL;
|
|
|
|
kfree(mgr->proposed_vcpis);
|
|
|
|
mgr->proposed_vcpis = NULL;
|
|
|
|
mutex_unlock(&mgr->payload_lock);
|
|
|
|
mgr->dev = NULL;
|
|
|
|
mgr->aux = NULL;
|
2017-07-12 23:51:02 +08:00
|
|
|
drm_atomic_private_obj_fini(&mgr->base);
|
2017-04-21 13:51:31 +08:00
|
|
|
mgr->funcs = NULL;
|
2019-09-04 04:45:52 +08:00
|
|
|
|
2019-02-20 06:41:02 +08:00
|
|
|
mutex_destroy(&mgr->delayed_destroy_lock);
|
2019-09-04 04:45:52 +08:00
|
|
|
mutex_destroy(&mgr->payload_lock);
|
|
|
|
mutex_destroy(&mgr->qlock);
|
|
|
|
mutex_destroy(&mgr->lock);
|
drm/dp_mst: Handle UP requests asynchronously
Once upon a time, hotplugging devices on MST branches actually worked in
DRM. Now, it only works in amdgpu (likely because of how it's hotplug
handlers are implemented). On both i915 and nouveau, hotplug
notifications from MST branches are noticed - but trying to respond to
them causes messaging timeouts and causes the whole topology state to go
out of sync with reality, usually resulting in the user needing to
replug the entire topology in hopes that it actually fixes things.
The reason for this is because the way we currently handle UP requests
in MST is completely bogus. drm_dp_mst_handle_up_req() is called from
drm_dp_mst_hpd_irq(), which is usually called from the driver's hotplug
handler. Because we handle sending the hotplug event from this function,
we actually cause the driver's hotplug handler (and in turn, all
sideband transactions) to block on
drm_device->mode_config.connection_mutex. This makes it impossible to
send any sideband messages from the driver's connector probing
functions, resulting in the aforementioned sideband message timeout.
There's even more problems with this beyond breaking hotplugging on MST
branch devices. It also makes it almost impossible to protect
drm_dp_mst_port struct members under a lock because we then have to
worry about dealing with all of the lock dependency issues that ensue.
So, let's finally actually fix this issue by handling the processing of
up requests asyncronously. This way we can send sideband messages from
most contexts without having to deal with getting blocked if we hold
connection_mutex. This also fixes MST branch device hotplugging on i915,
finally!
Cc: Juston Li <juston.li@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Harry Wentland <hwentlan@amd.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Sean Paul <sean@poorly.run>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022023641.8026-5-lyude@redhat.com
2019-06-18 04:37:18 +08:00
|
|
|
mutex_destroy(&mgr->up_req_lock);
|
2019-10-17 04:02:59 +08:00
|
|
|
mutex_destroy(&mgr->probe_lock);
|
2019-06-21 05:59:25 +08:00
|
|
|
#if IS_ENABLED(CONFIG_DRM_DEBUG_DP_MST_TOPOLOGY_REFS)
|
|
|
|
mutex_destroy(&mgr->topology_ref_history_lock);
|
|
|
|
#endif
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
EXPORT_SYMBOL(drm_dp_mst_topology_mgr_destroy);
|
|
|
|
|
2018-09-29 02:04:00 +08:00
|
|
|
static bool remote_i2c_read_ok(const struct i2c_msg msgs[], int num)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
if (num - 1 > DP_REMOTE_I2C_READ_MAX_TRANSACTIONS)
|
|
|
|
return false;
|
|
|
|
|
|
|
|
for (i = 0; i < num - 1; i++) {
|
|
|
|
if (msgs[i].flags & I2C_M_RD ||
|
|
|
|
msgs[i].len > 0xff)
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
return msgs[num - 1].flags & I2C_M_RD &&
|
|
|
|
msgs[num - 1].len <= 0xff;
|
|
|
|
}
|
|
|
|
|
2014-06-05 12:01:32 +08:00
|
|
|
/* I2C device */
|
|
|
|
static int drm_dp_mst_i2c_xfer(struct i2c_adapter *adapter, struct i2c_msg *msgs,
|
|
|
|
int num)
|
|
|
|
{
|
|
|
|
struct drm_dp_aux *aux = adapter->algo_data;
|
|
|
|
struct drm_dp_mst_port *port = container_of(aux, struct drm_dp_mst_port, aux);
|
|
|
|
struct drm_dp_mst_branch *mstb;
|
|
|
|
struct drm_dp_mst_topology_mgr *mgr = port->mgr;
|
|
|
|
unsigned int i;
|
|
|
|
struct drm_dp_sideband_msg_req_body msg;
|
|
|
|
struct drm_dp_sideband_msg_tx *txmsg = NULL;
|
|
|
|
int ret;
|
|
|
|
|
2019-01-11 08:53:28 +08:00
|
|
|
mstb = drm_dp_mst_topology_get_mstb_validated(mgr, port->parent);
|
2014-06-05 12:01:32 +08:00
|
|
|
if (!mstb)
|
|
|
|
return -EREMOTEIO;
|
|
|
|
|
2018-09-29 02:04:00 +08:00
|
|
|
if (!remote_i2c_read_ok(msgs, num)) {
|
2014-06-05 12:01:32 +08:00
|
|
|
DRM_DEBUG_KMS("Unsupported I2C transaction for MST device\n");
|
|
|
|
ret = -EIO;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
2015-10-14 16:51:17 +08:00
|
|
|
memset(&msg, 0, sizeof(msg));
|
2014-06-05 12:01:32 +08:00
|
|
|
msg.req_type = DP_REMOTE_I2C_READ;
|
|
|
|
msg.u.i2c_read.num_transactions = num - 1;
|
|
|
|
msg.u.i2c_read.port_number = port->port_num;
|
|
|
|
for (i = 0; i < num - 1; i++) {
|
|
|
|
msg.u.i2c_read.transactions[i].i2c_dev_id = msgs[i].addr;
|
|
|
|
msg.u.i2c_read.transactions[i].num_bytes = msgs[i].len;
|
|
|
|
msg.u.i2c_read.transactions[i].bytes = msgs[i].buf;
|
2018-09-29 02:03:59 +08:00
|
|
|
msg.u.i2c_read.transactions[i].no_stop_bit = !(msgs[i].flags & I2C_M_STOP);
|
2014-06-05 12:01:32 +08:00
|
|
|
}
|
|
|
|
msg.u.i2c_read.read_i2c_device_id = msgs[num - 1].addr;
|
|
|
|
msg.u.i2c_read.num_bytes_read = msgs[num - 1].len;
|
|
|
|
|
|
|
|
txmsg = kzalloc(sizeof(*txmsg), GFP_KERNEL);
|
|
|
|
if (!txmsg) {
|
|
|
|
ret = -ENOMEM;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
txmsg->dst = mstb;
|
|
|
|
drm_dp_encode_sideband_req(&msg, txmsg);
|
|
|
|
|
|
|
|
drm_dp_queue_down_tx(mgr, txmsg);
|
|
|
|
|
|
|
|
ret = drm_dp_mst_wait_tx_reply(mstb, txmsg);
|
|
|
|
if (ret > 0) {
|
|
|
|
|
2019-01-23 04:03:00 +08:00
|
|
|
if (txmsg->reply.reply_type == DP_SIDEBAND_REPLY_NAK) {
|
2014-06-05 12:01:32 +08:00
|
|
|
ret = -EREMOTEIO;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
if (txmsg->reply.u.remote_i2c_read_ack.num_bytes != msgs[num - 1].len) {
|
|
|
|
ret = -EIO;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
memcpy(msgs[num - 1].buf, txmsg->reply.u.remote_i2c_read_ack.bytes, msgs[num - 1].len);
|
|
|
|
ret = num;
|
|
|
|
}
|
|
|
|
out:
|
|
|
|
kfree(txmsg);
|
2019-01-11 08:53:28 +08:00
|
|
|
drm_dp_mst_topology_put_mstb(mstb);
|
2014-06-05 12:01:32 +08:00
|
|
|
return ret;
|
|
|
|
}
|
|
|
|
|
|
|
|
static u32 drm_dp_mst_i2c_functionality(struct i2c_adapter *adapter)
|
|
|
|
{
|
|
|
|
return I2C_FUNC_I2C | I2C_FUNC_SMBUS_EMUL |
|
|
|
|
I2C_FUNC_SMBUS_READ_BLOCK_DATA |
|
|
|
|
I2C_FUNC_SMBUS_BLOCK_PROC_CALL |
|
|
|
|
I2C_FUNC_10BIT_ADDR;
|
|
|
|
}
|
|
|
|
|
|
|
|
static const struct i2c_algorithm drm_dp_mst_i2c_algo = {
|
|
|
|
.functionality = drm_dp_mst_i2c_functionality,
|
|
|
|
.master_xfer = drm_dp_mst_i2c_xfer,
|
|
|
|
};
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_register_i2c_bus() - register an I2C adapter for I2C-over-AUX
|
|
|
|
* @aux: DisplayPort AUX channel
|
|
|
|
*
|
|
|
|
* Returns 0 on success or a negative error code on failure.
|
|
|
|
*/
|
|
|
|
static int drm_dp_mst_register_i2c_bus(struct drm_dp_aux *aux)
|
|
|
|
{
|
|
|
|
aux->ddc.algo = &drm_dp_mst_i2c_algo;
|
|
|
|
aux->ddc.algo_data = aux;
|
|
|
|
aux->ddc.retries = 3;
|
|
|
|
|
|
|
|
aux->ddc.class = I2C_CLASS_DDC;
|
|
|
|
aux->ddc.owner = THIS_MODULE;
|
|
|
|
aux->ddc.dev.parent = aux->dev;
|
|
|
|
aux->ddc.dev.of_node = aux->dev->of_node;
|
|
|
|
|
|
|
|
strlcpy(aux->ddc.name, aux->name ? aux->name : dev_name(aux->dev),
|
|
|
|
sizeof(aux->ddc.name));
|
|
|
|
|
|
|
|
return i2c_add_adapter(&aux->ddc);
|
|
|
|
}
|
|
|
|
|
|
|
|
/**
|
|
|
|
* drm_dp_mst_unregister_i2c_bus() - unregister an I2C-over-AUX adapter
|
|
|
|
* @aux: DisplayPort AUX channel
|
|
|
|
*/
|
|
|
|
static void drm_dp_mst_unregister_i2c_bus(struct drm_dp_aux *aux)
|
|
|
|
{
|
|
|
|
i2c_del_adapter(&aux->ddc);
|
|
|
|
}
|