Currently the first thing we do in cxl_probe is to grab a reference
on the pci device. Later on, we call device_register on our adapter.
In our remove path, we call device_unregister, but we never call
pci_dev_put. We therefore leak the device every time we do a
reflash.
device_register/unregister is sufficient to hold the reference.
Therefore, drop the call to pci_dev_get.
Here's why this is safe.
The proposed cxl_probe(pdev) calls cxl_adapter_init:
a) init calls cxl_adapter_alloc, which creates a struct cxl,
conventionally called adapter. This struct contains a
device entry, adapter->dev.
b) init calls cxl_configure_adapter, where we set
adapter->dev.parent = &dev->dev (here dev is the pci dev)
So at this point, the cxl adapter's device's parent is the PCI
device that I want to be refcounted properly.
c) init calls cxl_register_adapter
*) cxl_register_adapter calls device_register(&adapter->dev)
So now we're in device_register, where dev is the adapter device, and
we want to know if the PCI device is safe after we return.
device_register(&adapter->dev) calls device_initialize() and then
device_add().
device_add() does a get_device(). device_add() also explicitly grabs
the device's parent, and calls get_device() on it:
parent = get_device(dev->parent);
So therefore, device_register() takes a lock on the parent PCI dev,
which is what pci_dev_get() was guarding. pci_dev_get() can therefore
be safely removed.
Fixes: f204e0b8ce ("cxl: Driver code for powernv PCIe based cards for userspace access")
Cc: stable@vger.kernel.org
Signed-off-by: Daniel Axtens <dja@axtens.net>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch configures the PSL Timebase function and enables it,
after the CAPP has been initialized by OPAL.
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Philippe Bergheaud <felix@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
cxl_reset currently PERSTs the slot, and then repeatedly tries to
read MMIO space in order to kick off EEH.
There are 2 problems with this: it's unnecessary, and it's racy.
It's unnecessary because the PERST will bring down the PHB link.
That will be picked up by the CAPP, which will send out an HMI.
Skiboot, noticing an HMI from the CAPP, will send an OPAL
notification to the kernel, which will trigger EEH recovery.
It's also racy: the EEH recovery triggered by the CAPP will
eventually cause the MMIO space to have its mapping invalidated
and the pointer NULLed out. This races with our attempt to read
the MMIO space. This is causing OOPSes in testing.
Simply drop all the attempts to force EEH detection, and trust
that Skiboot will send the notification and that we'll act on it.
The Skiboot code to send the EEH notification has been in Skiboot
for as long as CAPP recovery has been supported, so we don't need
to worry about breaking obscure setups with ancient firmware.
Cc: Ryan Grimm <grimm@linux.vnet.ibm.com>
Cc: stable@vger.kernel.org
Fixes: 62fa19d4b4 ("cxl: Add ability to reset the card")
Signed-off-by: Daniel Axtens <dja@axtens.net>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Macro DEFINE_PCI_DEVICE_TABLE is deprecated. So, here use
struct pci_device_id instead of DEFINE_PCI_DEVICE_TABLE with
the goal of getting rid of this macro completely.
The Coccinelle semantic patch that performs this transformation
is as follows:
@@
identifier a;
declarer name DEFINE_PCI_DEVICE_TABLE;
initializer i;
@@
- DEFINE_PCI_DEVICE_TABLE(a)
+ const struct pci_device_id a[]
= i;
Signed-off-by: Vaishali Thakkar <vthakkar1994@gmail.com>
Acked-by: Michael Neuling <mikey@neuling.org>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
EEH (Enhanced Error Handling) allows a driver to recover from the
temporary failure of an attached PCI card. Enable basic CXL support
for EEH.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Provide a kernel API and a sysfs entry which allow a user to specify
that when a card is PERSTed, it's image will stay the same, allowing
it to participate in EEH.
cxl_reset is used to reflash the card. In that case, we cannot safely
assert that the image will not change. Therefore, disallow cxl_reset
if the flag is set.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If the driver doesn't participate in EEH, the AFUs will be removed
by cxl_remove, which will be invoked by EEH.
If the driver does particpate in EEH, the vPHB needs to stick around
so that the it can particpate.
In both cases, we shouldn't remove the AFU/vPHB.
Reviewed-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
As with an adapter, some aspects of initialisation are done only once
in the lifetime of an AFU: for example, allocating memory, or setting
up sysfs/debugfs files.
However, we may want to be able to do some parts of the initialisation
multiple times: for example, in error recovery we want to be able to
tear down and then re-map IO memory and IRQs.
Therefore, refactor AFU init/teardown as follows.
- Create two new functions: 'cxl_configure_afu', and its pair
'cxl_deconfigure_afu'. As with the adapter functions,
these (de)configure resources that do not need to last the entire
lifetime of the AFU.
- Allocating and releasing memory remain the task of 'cxl_alloc_afu'
and 'cxl_release_afu'.
- Once-only functions that do not involve allocating/releasing memory
stay in the overarching 'cxl_init_afu'/'cxl_remove_afu' pair.
However, the task of picking an AFU mode and activating it has been
broken out.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Some aspects of initialisation are done only once in the lifetime of
an adapter: for example, allocating memory for the adapter,
allocating the adapter number, or setting up sysfs/debugfs files.
However, we may want to be able to do some parts of the
initialisation multiple times: for example, in error recovery we
want to be able to tear down and then re-map IO memory and IRQs.
Therefore, refactor CXL init/teardown as follows.
- Keep the overarching functions 'cxl_init_adapter' and its pair,
'cxl_remove_adapter'.
- Move all 'once only' allocation/freeing steps to the existing
'cxl_alloc_adapter' function, and its pair 'cxl_release_adapter'
(This involves moving allocation of the adapter number out of
cxl_init_adapter.)
- Create two new functions: 'cxl_configure_adapter', and its pair
'cxl_deconfigure_adapter'. These two functions 'wire up' the
hardware --- they (de)configure resources that do not need to
last the entire lifetime of the adapter
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
- MMIO pointer unmapping is guarded by a null pointer check.
However, iounmap doesn't null the pointer, just invalidate it.
Therefore, explicitly null the pointer after unmapping.
- afu_desc_mmio also needs to be unmapped.
- PCI regions are allocated in cxl_map_adapter_regs.
Therefore they should be released in unmap, not elsewhere.
Acked-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Previously the SPA was allocated and freed upon entering and leaving
AFU-directed mode. This causes some issues for error recovery - contexts
hold a pointer inside the SPA, and they may persist after the AFU has
been detached.
We would ideally like to allocate the SPA when the AFU is allocated, and
release it until the AFU is released. However, we don't know how big the
SPA needs to be until we read the AFU descriptor.
Therefore, restructure the code:
- Allocate the SPA only once, on the first attach.
- Release the SPA only when the entire AFU is being released (not
detached). Guard the release with a NULL check, so we don't free
if it was never allocated (e.g. dedicated mode)
Acked-by: Cyril Bur <cyrilbur@gmail.com>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Destroy afu->contexts_idr on release of an afu, reclaiming the allocated
memory.
Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
A precision of 16 (%.16llx) has the same effect as a field width of 16
along with passing the 0 flag (%016llx), but the latter is much more
common in the kernel tree. Update cxl to use that.
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Before freeing p2n, test p2n, not p1n.
Signed-off-by: Daniel Axtens <dja@axtens.net>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Fix typo in debug print. p1_base() should be p2_base(). No change other
than to the debug output.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This patch does two things.
Firstly it presents the Accelerator Function Unit (AFUs) behind the POWER
Service Layer (PSL) as PCI devices on a virtual PCI Host Bridge (vPHB). This
in in addition to the PSL being a PCI device itself.
As part of the Coherent Accelerator Interface Architecture (CAIA) AFUs can
provide an AFU configuration. This AFU configuration recored is architected to
be the same as a PCI config space.
This patch sets discovers the AFU configuration records, provides AFU config
space read/write functions to these configuration records. It then enumerates
the PCI bus. It also hooks in PCI ops where appropriate. It also destroys the
vPHB when the physical card is removed.
Secondly, it add an in kernel API for AFU to use CXL. AFUs must present a
driver that firstly binds as a PCI device. This PCI device can then be using
to do CXL specific operations (that can't sit in the PCI ops) using this API.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Rename cxl_afu_reset() to __cxl_afu_reset() to we can reuse this function name
in the API.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Now that we parse the AFU Configuration record, dump some info on it when in
debug mode.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When probing we call pci_enable_device() but don't call pci_disable_device() on
fail. This causes refcounting issues in the PCI subsystem if a second driver
tries to bind to the same device.
This patch adds the pci_disable_device() to the probe error path. This error
path is hit when this cxl driver tries to bind to AFUs (on the vPHB) rather
than the physical device.
Signed-off-by: Michael Neuling <mikey@neuling.org>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When we expose AFUs as virtual PCI devices, they may look like the physical
CAPI PCI card. ie they may have the same vendor/device IDs.
We want to avoid these AFUs binding to this driver and any init this driver may
do.
Re-order card init to check the VSEC earlier before assigning BARs or
activating CXL. Also change the dev used in early prints as the adapter struct
may not be inited at this earlier stage.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Export the "AFU Error Buffer" via sysfs attribute (afu_err_buf). AFU
error buffer is used by the AFU to report application specific
errors. The contents of this buffer are AFU specific and are intended to
be interpreted by the application interacting with the afu.
Suggested-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Vaibhav Jain <vaibhav@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
If an AFU claims to have a configuration record but doesn't actually
contain a vendor and device ID, fail the AFU initialisation. Right now
this is just a way of politely letting AFU developers know that they
need to fix their config space, but later on we may expose the AFUs as
actual PCI devices in their own right and don't want to inadvertendly
expose an AFU with a bad config space.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
An AFU may optionally contain one or more PCIe like configuration
records, which can be used to identify the AFU.
This patch adds support for exposing the raw config space and the
vendor, device and class code under sysfs. These will appear in a
subdirectory of the AFU device corresponding with the configuration
record number, e.g.
cat /sys/class/cxl/afu0.0/cr0/vendor
0x1014
cat /sys/class/cxl/afu0.0/cr0/device
0x4350
cat /sys/class/cxl/afu0.0/cr0/class
0x120000
hexdump -C /sys/class/cxl/afu0.0/cr0/config
00000000 14 10 50 43 00 00 00 00 06 00 00 12 00 00 00 00 |..PC............|
00000010 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................|
*
00000100
These files behave in much the same way as the equivalent files for PCI
devices, with one exception being that the config file is currently
read-only and restricted to the root user. It is not necessarily
required to be this strict, but we currently do not have a compelling
use-case to make it writable and/or world-readable, so I erred on the
side of being restrictive.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
When unbinding and rebinding the driver on a system with a card in PHB0, this
error condition is reached after a few attempts:
ERROR: Bad of_node_put() on /pciex@3fffe40000000
CPU: 0 PID: 3040 Comm: bash Not tainted 3.18.0-rc3-12545-g3627ffe #152
Call Trace:
[c000000721acb5c0] [c00000000086ef94] .dump_stack+0x84/0xb0 (unreliable)
[c000000721acb640] [c00000000073a0a8] .of_node_release+0xd8/0xe0
[c000000721acb6d0] [c00000000044bc44] .kobject_release+0x74/0xe0
[c000000721acb760] [c0000000007394fc] .of_node_put+0x1c/0x30
[c000000721acb7d0] [c000000000545cd8] .cxl_probe+0x1a98/0x1d50
[c000000721acb900] [c0000000004845a0] .local_pci_probe+0x40/0xc0
[c000000721acb980] [c000000000484998] .pci_device_probe+0x128/0x170
[c000000721acba30] [c00000000052400c] .driver_probe_device+0xac/0x2a0
[c000000721acbad0] [c000000000522468] .bind_store+0x108/0x160
[c000000721acbb70] [c000000000521448] .drv_attr_store+0x38/0x60
[c000000721acbbe0] [c000000000293840] .sysfs_kf_write+0x60/0xa0
[c000000721acbc50] [c000000000292500] .kernfs_fop_write+0x140/0x1d0
[c000000721acbcf0] [c000000000208648] .vfs_write+0xd8/0x260
[c000000721acbd90] [c000000000208b18] .SyS_write+0x58/0x100
[c000000721acbe30] [c000000000009258] syscall_exit+0x0/0x98
We are missing a call to of_node_get(). pnv_pci_to_phb_node() should
call of_node_get() otherwise np's reference count isn't incremented and
it might go away. Rename pnv_pci_to_phb_node() to pnv_pci_get_phb_node()
so it's clear it calls of_node_get().
Signed-off-by: Ryan Grimm <grimm@linux.vnet.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Adds reset to sysfs which will PERST the card. If load_image_on_perst is set
to "user" or "factory", the PERST will cause that image to be loaded.
load_image_on_perst is set to "user" for production.
"none" could be used for debugging. The PSL trace arrays are preserved which
then can be read through debugfs.
PERST also triggers CAPP recovery. An HMI comes in, which is handled by EEH.
EEH unbinds the driver, calls into Sapphire to reinitialize the PHB, then
rebinds the driver.
Signed-off-by: Ryan Grimm <grimm@linux.vnet.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Turning snoops on is the last step in CAPP recovery. Sapphire is expected to
have reinitialized the PHB and done the previous recovery steps.
Add mode argument to opal call to do this. Driver can turn snoops off although
it does not currently.
Signed-off-by: Ryan Grimm <grimm@linux.vnet.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Select defaults such that a PERST causes flash image reload. Select which
image based on what the card is set up to load.
CXL_VSEC_PERST_LOADS_IMAGE selects whether PERST assertion causes flash image
load.
CXL_VSEC_PERST_SELECT_USER selects which image is loaded on the next PERST.
cxl_update_image_control writes these bits into the VSEC.
Signed-off-by: Ryan Grimm <grimm@linux.vnet.ibm.com>
Acked-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Upon inspection of the implementation specific registers, it was
discovered that the high bit of the implementation specific RXCTL
register was enabled, which enables the DEADB00F debug feature.
The debug feature causes MMIO reads to a disabled AFU to respond with
0xDEADB00F instead of all Fs. In general this should not be visible as
the kernel will only allow MMIO access to enabled AFUs, but there may be
some circumstances where an AFU may become disabled while it is use.
One such case would be an AFU designed to only be used in the dedicated
process mode and to disable itself after it has completed it's work
(however even in that case the effects of this debug flag would be
limited as the userspace application must have completed any required
MMIO accesses before the AFU disables itself with or without the flag).
This patch removes the debug flag and replaces the magic value
programmed into this register with a preprocessor define so it is
clearer what the rest of this initialisation does.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
We had a known sleep while atomic bug if a CXL device was forcefully
unbound while it was in use. This could occur as a result of EEH, or
manually induced with something like this while the device was in use:
echo 0000:01:00.0 > /sys/bus/pci/drivers/cxl-pci/unbind
The issue was that in this code path we iterated over each context and
forcefully detached it with the contexts_lock spin lock held, however
the detach also needed to take the spu_mutex, and call schedule.
This patch changes the contexts_lock to a mutex so that we are not in
atomic context while doing the detach, thereby avoiding the sleep while
atomic.
Also delete the related TODO comment, which suggested an alternate
solution which turned out to not be workable.
Cc: stable@vger.kernel.org
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
This is the core of the cxl driver.
It adds support for using cxl cards in the powernv environment only (ie POWER8
bare metal). It allows access to cxl accelerators by userspace using the
/dev/cxl/afuM.N char devices.
The kernel driver has no knowledge of the function implemented by the
accelerator. It provides services to userspace via the /dev/cxl/afuM.N
devices. When a program opens this device and runs the start work IOCTL, the
accelerator will have coherent access to that processes memory using the same
virtual addresses. That process may mmap the device to access any MMIO space
the accelerator provides. Also, reads on the device will allow interrupts to
be received. These services are further documented in a later patch in
Documentation/powerpc/cxl.txt.
Documentation of the cxl hardware architecture and userspace API is provided in
subsequent patches.
Signed-off-by: Ian Munsie <imunsie@au1.ibm.com>
Signed-off-by: Michael Neuling <mikey@neuling.org>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Daniel Axtens