Our current vsc9959_tas_guard_bands_update() algorithm has a limitation
imposed by the hardware design. To avoid packet overruns between one
gate interval and the next (which would add jitter for scheduled traffic
in the next gate), we configure the switch to use guard bands. These are
as large as the largest packet which is possible to be transmitted.
The problem is that at tc-taprio intervals of sizes comparable to a
guard band, there isn't an obvious place in which to split the interval
between the useful portion (for scheduling) and the guard band portion
(where scheduling is blocked).
For example, a 10 us interval at 1Gbps allows 1225 octets to be
transmitted. We currently split the interval between the bare minimum of
33 ns useful time (required to schedule a single packet) and the rest as
guard band.
But 33 ns of useful scheduling time will only allow a single packet to
be sent, be that packet 1200 octets in size, or 60 octets in size. It is
impossible to send 2 60 octets frames in the 10 us window. Except that
if we reduced the guard band (and therefore the maximum allowable SDU
size) to 5 us, the useful time for scheduling is now also 5 us, so more
packets could be scheduled.
The hardware inflexibility of not scheduling according to individual
packet lengths must unfortunately propagate to the user, who needs to
tune the queueMaxSDU values if he wants to fit more small packets into a
10 us interval, rather than one large packet.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Existing felix DSA drivers (vsc9959, vsc9953) are all switches that were
integrated in NXP SoCs, which makes them a bit unusual compared to the
usual Microchip branded Ocelot switches.
To be precise, looking at
Documentation/devicetree/bindings/net/mscc,vsc7514-switch.yaml, one can
see 21 memory regions for the "switch" node, and these correspond to the
"targets" of the switch IP, which are spread throughout the guts of that
SoC's memory space.
In NXP integrations, those targets still exist, but they were condensed
within a single memory region, with no other peripheral in between them,
so it made more sense for the driver to ioremap the entire memory space
of the switch, and then find the targets within that memory space via
some offsets hardcoded in the driver.
The effect of this design decision is that now, the felix driver expects
hardware instantiations to provide their own resource definitions, which
is kind of odd when considering a typical device (those are retrieved
from 'reg' properties in the device tree, using platform_get_resource()
or similar).
Allow other hardware instantiations that share the felix driver to not
provide a hardcoded array of resources in the future. Instead, make the
common denominator based on which regmaps are created be just the
resource "names". Each instantiation comes with its own array of names
that are mandatory for it, and with an optional array of resources.
So we split the resources in 2 arrays, one is what's requested and the
other is what's provided. There is one pool of provided resources, in
felix->info->resources (of length felix->info->num_resources). There are
2 different ways of requesting a resource. One is by enum ocelot_target
(this handles the global regmaps), and one is by int port (this handles
the per-port ones).
For the existing vsc9959 and vsc9953, it would be a bit stupid to
request something that's not provided, given that the 2 arrays are both
defined in the same place.
The advantage is that we can now modify felix_request_regmap_by_name()
to make felix->info->resources[] optional, and if absent, the
implementation can call dev_get_regmap() and this is something that is
compatible with MFD.
Co-developed-by: Colin Foster <colin.foster@in-advantage.com>
Signed-off-by: Colin Foster <colin.foster@in-advantage.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Use less verbose resource definitions in vsc9959 and vsc9953. This also
sets IORESOURCE_MEM in the constant array of resources, so we don't have
to do this from felix_init_structs() - in fact, in the future, we may
even support IORESOURCE_REG resources.
Note that this macro takes start and length as argument, and we had
start and end before. So transform end into length.
While at it, sort the resources according to their offset.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
It turns out that the idea of having a customizable implementation of a
regmap creation from a resource is not exactly useful. The idea was for
the new MFD-based VSC7512 driver to use something that creates a SPI
regmap from a resource. But there are problems in actually getting those
resources (it involves getting them from MFD).
To avoid all that, we'll be getting resources by name, so this custom
init_regmap() method won't be needed. Remove it.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This address is only relevant for the vsc9959, which is a PCIe device
that holds its switch registers in a different PCIe BAR compared to the
registers for the internal MDIO controller.
Hide this aspect from the common felix driver and move the
pci_resource_start() call to the only place that needs it, which is in
vsc9959_mdio_bus_alloc().
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The imdio_res is used only by vsc9959, which references its own
vsc9959_imdio_res through the common felix_info->imdio_res pointer.
Since the common code doesn't care about this resource (and it can't be
part of the common array of resources, either, because it belongs in a
different PCI BAR), just reference it directly.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Remove unnecessary set_drvdata(NULL) function in ->remove(),
the driver_data will be set to NULL in device_unbind_cleanup()
after calling ->remove().
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Changing the DSA master means different things depending on the tagging
protocol in use.
For NPI mode ("ocelot" and "seville"), there is a single port which can
be configured as NPI, but DSA only permits changing the CPU port
affinity of user ports one by one. So changing a user port to a
different NPI port globally changes what the NPI port is, and breaks the
user ports still using the old one.
To address this while still permitting the change of the NPI port,
require that the user ports which are still affine to the old NPI port
are down, and cannot be brought up until they are all affine to the same
NPI port.
The tag_8021q mode ("ocelot-8021q") is more flexible, in that each user
port can be freely assigned to one CPU port or to the other. This works
by filtering host addresses towards both tag_8021q CPU ports, and then
restricting the forwarding from a certain user port only to one of the
two tag_8021q CPU ports.
Additionally, the 2 tag_8021q CPU ports can be placed in a LAG. This
works by enabling forwarding via PGID_SRC from a certain user port
towards the logical port ID containing both tag_8021q CPU ports, but
then restricting forwarding per packet, via the LAG hash codes in
PGID_AGGR, to either one or the other.
When we change the DSA master to a LAG device, DSA guarantees us that
the LAG has at least one lower interface as a physical DSA master.
But DSA masters can come and go as lowers of that LAG, and
ds->ops->port_change_master() will not get called, because the DSA
master is still the same (the LAG). So we need to hook into the
ds->ops->port_lag_{join,leave} calls on the CPU ports and update the
logical port ID of the LAG that user ports are assigned to.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Drivers could refuse to offload a LAG configuration for a variety of
reasons, mainly having to do with its TX type. Additionally, since DSA
masters may now also be LAG interfaces, and this will translate into a
call to port_lag_join on the CPU ports, there may be extra restrictions
there. Propagate the netlink extack to this DSA method in order for
drivers to give a meaningful error message back to the user.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
All switch families supported by the ocelot lib (ocelot, felix, seville)
export the same registers so far. But for example felix also has TSN
counters, while the others don't.
To reduce the bloat even further, create an OCELOT_COMMON_STATS() macro
which just lists all stats that are common between switches. The array
elements are still replicated among all of vsc9959_stats_layout,
vsc9953_stats_layout and ocelot_stats_layout.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The current definition of struct ocelot_stat_layout is long-winded (4
lines per entry, and we have hundreds of entries), so we could make an
effort to use the C preprocessor and reduce the line count.
Create an implicit correspondence between enum ocelot_reg, which tells
us the register address (SYS_COUNT_RX_OCTETS etc) and enum ocelot_stat
which allows us to index the ocelot->stats array (OCELOT_STAT_RX_OCTETS
etc), and don't require us to specify both when we define what stats
each switch family has.
Create an OCELOT_STAT() macro that pairs only an enum ocelot_stat to an
enum ocelot_reg, and an OCELOT_STAT_ETHTOOL() macro which also contains
a name exported to the unstructured ethtool -S stringset API. For now,
we define all counters as having the OCELOT_STAT_ETHTOOL() kind, but we
will add more counters in the future which are not exported to the
unstructured ethtool -S.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The hardware counter is called C_TX_AGED, so rename SYS_COUNT_TX_AGING
to SYS_COUNT_TX_AGED. This will become important since we want to
minimize the way in which we declare struct ocelot_stat_layout elements,
using the C preprocessor.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
DSA is integrated with the new standardized ethtool -S --groups option,
but the felix driver only exports unstructured statistics.
Reuse the array of 64-bit statistics collected by ocelot_check_stats_work(),
but just export select values from it.
Since ocelot_check_stats_work() runs periodically to avoid 32-bit
overflow, and the ethtool calling context is sleepable, we update the
64-bit stats one more time, to provide up-to-date values. The locking
scheme with a mutex followed by a spinlock is a bit hard to digest, so
we create and use a ocelot_port_stats_run() helper with a callback that
populates the ethool stats group the caller is interested in.
The exported stats are:
ethtool -S swp0 --groups eth-phy
ethtool -S swp0 --groups eth-mac
ethtool -S swp0 --groups eth-ctrl
ethtool -S swp0 --groups rmon
ethtool --include-statistics --show-pause swp0
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Move the logic from the ocelot switchdev driver's ocelot_get_stats64()
method to the common switch lib and reuse it for the DSA driver.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The Felix PSFP counters suffer from the same problem as the ocelot
ndo_get_stats64 ones - they are 32-bit, so they can easily overflow and
this can easily go undetected.
Add a custom hook in ocelot_check_stats_work() through which driver
specific actions can be taken, and update the stats for the existing
PSFP filters from that hook.
Previously, vsc9959_psfp_filter_add() and vsc9959_psfp_filter_del() were
serialized with respect to each other via rtnl_lock(). However, with the
new entry point into &psfp->sfi_list coming from the periodic worker, we
now need an explicit mutex to serialize access to these lists.
We used to keep a struct felix_stream_filter_counters on stack, through
which vsc9959_psfp_stats_get() - a FLOW_CLS_STATS callback - would
retrieve data from vsc9959_psfp_counters_get(). We need to become
smarter about that in 3 ways:
- we need to keep a persistent set of counters for each stream instead
of keeping them on stack
- we need to promote those counters from u32 to u64, and create a
procedure that properly keeps 64-bit counters. Since we clear the
hardware counters anyway, and we poll every 2 seconds, a simple
increment of a u64 counter with a u32 value will perfectly do the job.
- FLOW_CLS_STATS also expect incremental counters, so we also need to
zeroize our u64 counters every time sch_flower calls us
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
To support SPI-controlled switches in the future, access to
SYS_STAT_CFG_STAT_VIEW needs to be done outside of any spinlock
protected region, but it still needs to be serialized (by a mutex).
Split the ocelot->stats_lock spinlock into a mutex that serializes
indirect access to hardware registers (ocelot->stat_view_lock) and a
spinlock that serializes access to the u64 ocelot->stats array.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
TSN stream (802.1Qci, 802.1CB) filters are also accessed through
STAT_VIEW, just like the port registers, but these counters are per
stream, rather than per port. So we don't keep them in
ocelot_port_update_stats().
What we can do, however, is we can create register definitions for them
just like we have for the port counters, and delete the last remaining
user of the SYS_CNT register + a group index (read_gix).
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/freescale/fec.h
7d650df99d ("net: fec: add pm_qos support on imx6q platform")
40c79ce13b ("net: fec: add stop mode support for imx8 platform")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The read-modify-write of QSYS_TAG_CONFIG from vsc9959_sched_speed_set()
runs unlocked with respect to the other functions that access it, which
are vsc9959_tas_guard_bands_update(), vsc9959_qos_port_tas_set() and
vsc9959_tas_clock_adjust(). All the others are under ocelot->tas_lock,
so move the vsc9959_sched_speed_set() access under that lock as well, to
resolve the concurrency.
Fixes: 55a515b1f5 ("net: dsa: felix: drop oversized frames with tc-taprio instead of hanging the port")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Experimentally, it looks like when QSYS_QMAXSDU_CFG_7 is set to 605,
frames even way larger than 601 octets are transmitted even though these
should be considered as oversized, according to the documentation, and
dropped.
Since oversized frame dropping depends on frame size, which is only
known at the EOF stage, and therefore not at SOF when cut-through
forwarding begins, it means that the switch cannot take QSYS_QMAXSDU_CFG_*
into consideration for traffic classes that are cut-through.
Since cut-through forwarding has no UAPI to control it, and the driver
enables it based on the mantra "if we can, then why not", the strategy
is to alter vsc9959_cut_through_fwd() to take into consideration which
tc's have oversize frame dropping enabled, and disable cut-through for
them. Then, from vsc9959_tas_guard_bands_update(), we re-trigger the
cut-through determination process.
There are 2 strategies for vsc9959_cut_through_fwd() to determine
whether a tc has oversized dropping enabled or not. One is to keep a bit
mask of traffic classes per port, and the other is to read back from the
hardware registers (a non-zero value of QSYS_QMAXSDU_CFG_* means the
feature is enabled). We choose reading back from registers, because
struct ocelot_port is shared with drivers (ocelot, seville) that don't
support either cut-through nor tc-taprio, and we don't have a felix
specific extension of struct ocelot_port. Furthermore, reading registers
from the Felix hardware is quite cheap, since they are memory-mapped.
Fixes: 55a515b1f5 ("net: dsa: felix: drop oversized frames with tc-taprio instead of hanging the port")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The blamed commit broke tc-taprio schedules such as this one:
tc qdisc replace dev $swp1 root taprio \
num_tc 8 \
map 0 1 2 3 4 5 6 7 \
queues 1@0 1@1 1@2 1@3 1@4 1@5 1@6 1@7 \
base-time 0 \
sched-entry S 0x7f 990000 \
sched-entry S 0x80 10000 \
flags 0x2
because the gate entry for TC 7 (S 0x80 10000 ns) now has a static guard
band added earlier than its 'gate close' event, such that packet
overruns won't occur in the worst case of the largest packet possible.
Since guard bands are statically determined based on the per-tc
QSYS_QMAXSDU_CFG_* with a fallback on the port-based QSYS_PORT_MAX_SDU,
we need to discuss what happens with TC 7 depending on kernel version,
since the driver, prior to commit 55a515b1f5 ("net: dsa: felix: drop
oversized frames with tc-taprio instead of hanging the port"), did not
touch QSYS_QMAXSDU_CFG_*, and therefore relied on QSYS_PORT_MAX_SDU.
1 (before vsc9959_tas_guard_bands_update): QSYS_PORT_MAX_SDU defaults to
1518, and at gigabit this introduces a static guard band (independent
of packet sizes) of 12144 ns, plus QSYS::HSCH_MISC_CFG.FRM_ADJ (bit
time of 20 octets => 160 ns). But this is larger than the time window
itself, of 10000 ns. So, the queue system never considers a frame with
TC 7 as eligible for transmission, since the gate practically never
opens, and these frames are forever stuck in the TX queues and hang
the port.
2 (after vsc9959_tas_guard_bands_update): Under the sole goal of
enabling oversized frame dropping, we make an effort to set
QSYS_QMAXSDU_CFG_7 to 1230 bytes. But QSYS_QMAXSDU_CFG_7 plays
one more role, which we did not take into account: per-tc static guard
band, expressed in L2 byte time (auto-adjusted for FCS and L1 overhead).
There is a discrepancy between what the driver thinks (that there is
no guard band, and 100% of min_gate_len[tc] is available for egress
scheduling) and what the hardware actually does (crops the equivalent
of QSYS_QMAXSDU_CFG_7 ns out of min_gate_len[tc]). In practice, this
means that the hardware thinks it has exactly 0 ns for scheduling tc 7.
In both cases, even minimum sized Ethernet frames are stuck on egress
rather than being considered for scheduling on TC 7, even if they would
fit given a proper configuration. Considering the current situation,
with vsc9959_tas_guard_bands_update(), frames between 60 octets and 1230
octets in size are not eligible for oversized dropping (because they are
smaller than QSYS_QMAXSDU_CFG_7), but won't be considered as eligible
for scheduling either, because the min_gate_len[7] (10000 ns) minus the
guard band determined by QSYS_QMAXSDU_CFG_7 (1230 octets * 8 ns per
octet == 9840 ns) minus the guard band auto-added for L1 overhead by
QSYS::HSCH_MISC_CFG.FRM_ADJ (20 octets * 8 ns per octet == 160 octets)
leaves 0 ns for scheduling in the queue system proper.
Investigating the hardware behavior, it becomes apparent that the queue
system needs precisely 33 ns of 'gate open' time in order to consider a
frame as eligible for scheduling to a tc. So the solution to this
problem is to amend vsc9959_tas_guard_bands_update(), by giving the
per-tc guard bands less space by exactly 33 ns, just enough for one
frame to be scheduled in that interval. This allows the queue system to
make forward progress for that port-tc, and prevents it from hanging.
Fixes: 297c4de6f7 ("net: dsa: felix: re-enable TAS guard band mode")
Reported-by: Xiaoliang Yang <xiaoliang.yang_1@nxp.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This is a partial revert of commit c295f9831f ("net: mscc: ocelot:
switch from {,un}set to {,un}assign for tag_8021q CPU ports"), because
as it turns out, this isn't how tag_8021q CPU ports under a LAG are
supposed to work.
Under that scenario, all user ports are "assigned" to the single
tag_8021q CPU port represented by the logical port corresponding to the
bonding interface. So one CPU port in a LAG would have is_dsa_8021q_cpu
set to true (the one whose physical port ID is equal to the logical port
ID), and the other one to false.
In turn, this makes 2 undesirable things happen:
(1) PGID_CPU contains only the first physical CPU port, rather than both
(2) only the first CPU port will be added to the private VLANs used by
ocelot for VLAN-unaware bridging
To make the driver behave in the same way for both bonded CPU ports, we
need to bring back the old concept of setting up a port as a tag_8021q
CPU port, and this is what deals with VLAN membership and PGID_CPU
updating. But we also need the CPU port "assignment" (the user to CPU
port affinity), and this is what updates the PGID_SRC forwarding rules.
All DSA CPU ports are statically configured for tag_8021q mode when the
tagging protocol is changed to ocelot-8021q. User ports are "assigned"
to one CPU port or the other dynamically (this will be handled by a
future change).
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
With so many counter addresses recently discovered as being wrong, it is
desirable to at least have a central database of information, rather
than two: one through the SYS_COUNT_* registers (used for
ndo_get_stats64), and the other through the offset field of struct
ocelot_stat_layout elements (used for ethtool -S).
The strategy will be to keep the SYS_COUNT_* definitions as the single
source of truth, but for that we need to expand our current definitions
to cover all registers. Then we need to convert the ocelot region
creation logic, and stats worker, to the read semantics imposed by going
through SYS_COUNT_* absolute register addresses, rather than offsets
of 32-bit words relative to SYS_COUNT_RX_OCTETS (which should have been
SYS_CNT, by the way).
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The ocelot counters are 32-bit and require periodic reading, every 2
seconds, by ocelot_port_update_stats(), so that wraparounds are
detected.
Currently, the counters reported by ocelot_get_stats64() come from the
32-bit hardware counters directly, rather than from the 64-bit
accumulated ocelot->stats, and this is a problem for their integrity.
The strategy is to make ocelot_get_stats64() able to cherry-pick
individual stats from ocelot->stats the way in which it currently reads
them out from SYS_COUNT_* registers. But currently it can't, because
ocelot->stats is an opaque u64 array that's used only to feed data into
ethtool -S.
To solve that problem, we need to make ocelot->stats indexable, and
associate each element with an element of struct ocelot_stat_layout used
by ethtool -S.
This makes ocelot_stat_layout a fat (and possibly sparse) array, so we
need to change the way in which we access it. We no longer need
OCELOT_STAT_END as a sentinel, because we know the array's size
(OCELOT_NUM_STATS). We just need to skip the array elements that were
left unpopulated for the switch revision (ocelot, felix, seville).
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
ocelot_get_stats64() currently runs unlocked and therefore may collide
with ocelot_port_update_stats() which indirectly accesses the same
counters. However, ocelot_get_stats64() runs in atomic context, and we
cannot simply take the sleepable ocelot->stats_lock mutex. We need to
convert it to an atomic spinlock first. Do that as a preparatory change.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Reading stats using the SYS_COUNT_* register definitions is only used by
ocelot_get_stats64() from the ocelot switchdev driver, however,
currently the bucket definitions are incorrect.
Separately, on both RX and TX, we have the following problems:
- a 256-1023 bucket which actually tracks the 256-511 packets
- the 1024-1526 bucket actually tracks the 512-1023 packets
- the 1527-max bucket actually tracks the 1024-1526 packets
=> nobody tracks the packets from the real 1527-max bucket
Additionally, the RX_PAUSE, RX_CONTROL, RX_LONGS and RX_CLASSIFIED_DROPS
all track the wrong thing. However this doesn't seem to have any
consequence, since ocelot_get_stats64() doesn't use these.
Even though this problem only manifests itself for the switchdev driver,
we cannot split the fix for ocelot and for DSA, since it requires fixing
the bucket definitions from enum ocelot_reg, which makes us necessarily
adapt the structures from felix and seville as well.
Fixes: 84705fc165 ("net: dsa: felix: introduce support for Seville VSC9953 switch")
Fixes: 5605194877 ("net: dsa: ocelot: add driver for Felix switch family")
Fixes: a556c76adc ("net: mscc: Add initial Ocelot switch support")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
What the driver actually reports as 256-511 is in fact 512-1023, and the
TX packets in the 256-511 bucket are not reported. Fix that.
Fixes: 5605194877 ("net: dsa: ocelot: add driver for Felix switch family")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The way in which dsa_tree_change_tag_proto() works is that when
dsa_tree_notify() fails, it doesn't know whether the operation failed
mid way in a multi-switch tree, or it failed for a single-switch tree.
So even though drivers need to fail cleanly in
ds->ops->change_tag_protocol(), DSA will still call dsa_tree_notify()
again, to restore the old tag protocol for potential switches in the
tree where the change did succeeed (before failing for others).
This means for the felix driver that if we report an error in
felix_change_tag_protocol(), we'll get another call where proto_ops ==
old_proto_ops. If we proceed to act upon that, we may do unexpected
things. For example, we will call dsa_tag_8021q_register() twice in a
row, without any dsa_tag_8021q_unregister() in between. Then we will
actually call dsa_tag_8021q_unregister() via old_proto_ops->teardown,
which (if it manages to run at all, after walking through corrupted data
structures) will leave the ports inoperational anyway.
The bug can be readily reproduced if we force an error while in
tag_8021q mode; this crashes the kernel.
echo ocelot-8021q > /sys/class/net/eno2/dsa/tagging
echo edsa > /sys/class/net/eno2/dsa/tagging # -EPROTONOSUPPORT
Unable to handle kernel NULL pointer dereference at virtual address 0000000000000014
Call trace:
vcap_entry_get+0x24/0x124
ocelot_vcap_filter_del+0x198/0x270
felix_tag_8021q_vlan_del+0xd4/0x21c
dsa_switch_tag_8021q_vlan_del+0x168/0x2cc
dsa_switch_event+0x68/0x1170
dsa_tree_notify+0x14/0x34
dsa_port_tag_8021q_vlan_del+0x84/0x110
dsa_tag_8021q_unregister+0x15c/0x1c0
felix_tag_8021q_teardown+0x16c/0x180
felix_change_tag_protocol+0x1bc/0x230
dsa_switch_event+0x14c/0x1170
dsa_tree_change_tag_proto+0x118/0x1c0
Fixes: 7a29d220f4 ("net: dsa: felix: reimplement tagging protocol change with function pointers")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Link: https://lore.kernel.org/r/20220808125127.3344094-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
min_gate_len[tc] is supposed to track the shortest interval of
continuously open gates for a traffic class. For example, in the
following case:
TC 76543210
t0 00000001b 200000 ns
t1 00000010b 200000 ns
min_gate_len[0] and min_gate_len[1] should be 200000, while
min_gate_len[2-7] should be 0.
However what happens is that min_gate_len[0] is 200000, but
min_gate_len[1] ends up being 0 (despite gate_len[1] being 200000 at the
point where the logic detects the gate close event for TC 1).
The problem is that the code considers a "gate close" event whenever it
sees that there is a 0 for that TC (essentially it's level rather than
edge triggered). By doing that, any time a gate is seen as closed
without having been open prior, gate_len, which is 0, will be written
into min_gate_len. Once min_gate_len becomes 0, it's impossible for it
to track anything higher than that (the length of actually open
intervals).
To fix this, we make the writing to min_gate_len[tc] be edge-triggered,
which avoids writes for gates that are closed in consecutive intervals.
However what this does is it makes us need to special-case the
permanently closed gates at the end.
Fixes: 55a515b1f5 ("net: dsa: felix: drop oversized frames with tc-taprio instead of hanging the port")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://lore.kernel.org/r/20220804202817.1677572-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
felix_vsc9959.c calls taprio_offload_get() and taprio_offload_free(),
symbols exported by net/sched/sch_taprio.c. As such, we must disallow
building the Felix driver as built-in when the symbol exported by
tc-taprio isn't present in the kernel image.
Fixes: 1c9017e44a ("net: dsa: felix: keep reference on entire tc-taprio config")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://lore.kernel.org/r/20220704190241.1288847-2-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Time-sensitive networking code needs to work with PTP times expressed in
nanoseconds, and with packet transmission times expressed in
picoseconds, since those would be fractional at higher than gigabit
speed when expressed in nanoseconds.
Convert the existing uses in tc-taprio and the ocelot/felix DSA driver
to a PSEC_PER_NSEC macro. This macro is placed in include/linux/time64.h
as opposed to its relatives (PSEC_PER_SEC etc) from include/vdso/time64.h
because the vDSO library does not (yet) need/use it.
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: Vincenzo Frascino <vincenzo.frascino@arm.com> # for the vDSO parts
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, sending a packet into a time gate too small for it (or always
closed) causes the queue system to hold the frame forever. Even worse,
this frame isn't subject to aging either, because for that to happen, it
needs to be scheduled for transmission in the first place. But the frame
will consume buffer memory and frame references while it is forever held
in the queue system.
Before commit a4ae997adc ("net: mscc: ocelot: initialize watermarks to
sane defaults"), this behavior was somewhat subtle, as the switch had a
more intricately tuned default watermark configuration out of reset,
which did not allow any single port and tc to consume the entire switch
buffer space. Nonetheless, the held frames are still there, and they
reduce the total backplane capacity of the switch.
However, after the aforementioned commit, the behavior can be very
clearly seen, since we deliberately allow each {port, tc} to consume the
entire shared buffer of the switch minus the reservations (and we
disable all reservations by default). That is to say, we allow a
permanently closed tc-taprio gate to hang the entire switch.
A careful inspection of the documentation shows that the QSYS:Q_MAX_SDU
per-port-tc registers serve 2 purposes: one is for guard band calculation
(when zero, this falls back to QSYS:PORT_MAX_SDU), and the other is to
enable oversized frame dropping (when non-zero).
Currently the QSYS:Q_MAX_SDU registers are all zero, so oversized frame
dropping is disabled. The goal of the change is to enable it seamlessly.
For that, we need to hook into the MTU change, tc-taprio change, and
port link speed change procedures, since we depend on these variables.
Frames are not dropped on egress due to a queue system oversize
condition, instead that egress port is simply excluded from the mask of
valid destination ports for the packet. If there are no destination
ports at all, the ingress counter that increments is the generic
"drop_tail" in ethtool -S.
The issue exists in various forms since the tc-taprio offload was introduced.
Fixes: de143c0e27 ("net: dsa: felix: Configure Time-Aware Scheduler via taprio offload")
Reported-by: Richie Pearn <richard.pearn@nxp.com>
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In vsc9959_tas_clock_adjust(), the INIT_GATE_STATE field is not changed,
only the ENABLE field. Similarly for the disabling of the time-aware
shaper in vsc9959_qos_port_tas_set().
To reflect this, keep the QSYS_TAG_CONFIG_INIT_GATE_STATE_M mask out of
the read-modify-write procedure to make it clearer what is the intention
of the code.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In a future change we will need to remember the entire tc-taprio config
on all ports rather than just the base time, so use the
taprio_offload_get() helper function to replace ocelot_port->base_time
with ocelot_port->taprio.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
drivers/net/ethernet/microchip/sparx5/sparx5_switchdev.c
9c5de246c1 ("net: sparx5: mdb add/del handle non-sparx5 devices")
fbb89d02e3 ("net: sparx5: Allow mdb entries to both CPU and ports")
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Both PSFP stats and the port stats read by ocelot_check_stats_work() are
indirectly read through the same mechanism - write to STAT_CFG:STAT_VIEW,
read from SYS:STAT:CNT[n].
It's just that for port stats, we write STAT_VIEW with the index of the
port, and for PSFP stats, we write STAT_VIEW with the filter index.
So if we allow them to run concurrently, ocelot_check_stats_work() may
change the view from vsc9959_psfp_counters_get(), and vice versa.
Fixes: 7d4b564d6a ("net: dsa: felix: support psfp filter on vsc9959")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://lore.kernel.org/r/20220629183007.3808130-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When adjusting the PTP clock, the base time of the TAS configuration
will become unreliable. We need reset the TAS configuration by using a
new base time.
For example, if the driver gets a base time 0 of Qbv configuration from
user, and current time is 20000. The driver will set the TAS base time
to be 20000. After the PTP clock adjustment, the current time becomes
10000. If the TAS base time is still 20000, it will be a future time,
and TAS entry list will stop running. Another example, if the current
time becomes to be 10000000 after PTP clock adjust, a large time offset
can cause the hardware to hang.
This patch introduces a tas_clock_adjust() function to reset the TAS
module by using a new base time after the PTP clock adjustment. This can
avoid issues above.
Due to PTP clock adjustment can occur at any time, it may conflict with
the TAS configuration. We introduce a new TAS lock to serialize the
access to the TAS registers.
Signed-off-by: Xiaoliang Yang <xiaoliang.yang_1@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Update the VCAP filters to support multiple tag_8021q CPU ports.
TX works using a filter for VLAN ID on the ingress of the CPU port, with
a redirect and a VLAN pop action. This can be updated trivially by
amending the ingress port mask of this rule to match on all tag_8021q
CPU ports.
RX works using a filter for ingress port on the egress of the CPU port,
with a VLAN push action. Here we need to replicate these filters for
each tag_8021q CPU port, and let them all have the same action.
This means that the OCELOT_VCAP_ES0_TAG_8021Q_RXVLAN() cookie needs to
encode a unique value for every {user port, CPU port} pair it's given.
Do this by encoding the CPU port in the upper 16 bits of the cookie, and
the user port in the lower 16 bits.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is a desire for the felix driver to gain support for multiple
tag_8021q CPU ports, but the current model prevents it.
This is because ocelot_apply_bridge_fwd_mask() only takes into
consideration whether a port is a tag_8021q CPU port, but not whose CPU
port it is.
We need a model where we can have a direct affinity between an ocelot
port and a tag_8021q CPU port. This serves as the basis for multiple CPU
ports.
Declare a "dsa_8021q_cpu" backpointer in struct ocelot_port which
encodes that affinity. Repurpose the "ocelot_set_dsa_8021q_cpu" API to
"ocelot_assign_dsa_8021q_cpu" to express the change of paradigm.
Note that this change makes the first practical use of the new
ocelot_port->index field in ocelot_port_unassign_dsa_8021q_cpu(), where
we need to remove the old tag_8021q CPU port from the reserved VLAN range.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Absorb the final details of calling ocelot_port_{,un}set_dsa_8021q_cpu(),
i.e. the need to lock &ocelot->fwd_domain_lock, into the callee, to
simplify the caller and permit easier code reuse later.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add more logic to ocelot_port_{,un}set_dsa_8021q_cpu() from the ocelot
switch lib by encapsulating the ocelot_apply_bridge_fwd_mask() call that
felix used to have.
This is necessary because the CPU port change procedure will also need
to do this, and it's good to reduce code duplication by having an entry
point in the ocelot switch lib that does all that is needed.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
PGID_CPU must be updated every time a port is configured or unconfigured
as a tag_8021q CPU port. The ocelot switch lib already has a hook for
that operation, so move the updating of PGID_CPU to those hooks.
These bits are pretty specific to DSA, so normally I would keep them out
of the common switch lib, but when tag_8021q is in use, this has
implications upon the forwarding mask determined by
ocelot_apply_bridge_fwd_mask() and called extensively by the switch lib.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
PGID_BC is configured statically by ocelot_init() to flood towards the
CPU port module, and dynamically by ocelot_port_set_bcast_flood()
towards all user ports.
When the tagging protocol changes, the intention is to turn off flooding
towards the old pipe towards the host, and to turn it on towards the new
pipe.
Due to a recent change which removed the adjustment of PGID_BC from
felix_set_host_flood(), 3 things happen.
- when we change from NPI to tag_8021q mode: in this mode, the CPU port
module is accessed via registers, and used to read PTP packets with
timestamps. We fail to disable broadcast flooding towards the CPU port
module, and to enable broadcast flooding towards the physical port
that serves as a DSA tag_8021q CPU port.
- from tag_8021q to NPI mode: in this mode, the CPU port module is
redirected to a physical port. We fail to disable broadcast flooding
towards the physical tag_8021q CPU port, and to enable it towards the
CPU port module at ocelot->num_phys_ports.
- when the ports are put in promiscuous mode, we also fail to update
PGID_BC towards the host pipe of the current protocol.
First issue means that felix_check_xtr_pkt() has to do extra work,
because it will not see only PTP packets, but also broadcasts. It needs
to dequeue these packets just to drop them.
Third issue is inconsequential, since PGID_BC is allocated from the
nonreserved multicast PGID space, and these PGIDs are conveniently
initialized to 0x7f (i.e. flood towards all ports except the CPU port
module). Broadcasts reach the NPI port via ocelot_init(), and reach the
tag_8021q CPU port via the hardware defaults.
Second issue is also inconsequential, because we fail both at disabling
and at enabling broadcast flooding on a port, so the defaults mentioned
above are preserved, and they are fine except for the performance impact.
Fixes: 7a29d220f4 ("net: dsa: felix: reimplement tagging protocol change with function pointers")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently the ocelot switch lib is unaware of the index of a struct
ocelot_port, since that is kept in the encapsulating structures of outer
drivers (struct dsa_port :: index, struct ocelot_port_private :: chip_port).
With the upcoming increase in complexity associated with assigning DSA
tag_8021q CPU ports to certain user ports, it becomes necessary for the
switch lib to be able to retrieve the index of a certain ocelot_port.
Therefore, introduce a new u8 to ocelot_port (same size as the chip_port
used by the ocelot switchdev driver) and rework the existing code to
populate and use it.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The error handling for the current tagging protocol change procedure is
a bit brittle (we dismantle the previous tagging protocol entirely
before setting up the new one). By identifying which parts of a tagging
protocol are unique to itself and which parts are shared with the other,
we can implement a protocol change procedure where error handling is a
bit more robust, because we start setting up the new protocol first, and
tear down the old one only after the setup of the specific and shared
parts succeeded.
The protocol change is a bit too open-coded too, in the area of
migrating host flood settings and MDBs. By identifying what differs
between tagging protocols (the forwarding masks for host flooding) we
can implement a more straightforward migration procedure which is
handled in the shared portion of the protocol change, rather than
individually by each protocol.
Therefore, a more structured approach calls for the introduction of a
structure of function pointers per tagging protocol. This covers setup,
teardown and the host forwarding mask. In the future it will also cover
how to prepare for a new DSA master.
The initial tagging protocol setup (at driver probe time) and the final
teardown (at driver removal time) are also adapted to call into the
structured methods of the specific protocol in current use. This is
especially relevant for teardown, where we previously called
felix_del_tag_protocol() only for the first CPU port. But by not
specifying which CPU port this is for, we gain more flexibility to
support multiple CPU ports in the future.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Ocelot switches support a single active CPU port at a time (at least as
a trapping destination, i.e. for control traffic). This is true
regardless of whether we are using the native copy-to-CPU-port-module
functionality, or a redirect action towards the software-defined
tag_8021q CPU port.
Currently we assume that the trapping destination in tag_8021q mode is
the first CPU port, yet in the future we may want to migrate the user
ports to the second CPU port.
For that to work, we need to make sure that the tag_8021q trapping
destination is a CPU port that is active, i.e. is used by at least some
user port on which the trap was added. Otherwise, we may end up
redirecting the traffic to a CPU port which isn't even up.
Note that due to the current design where we simply choose the CPU port
of the first port from the trap's ingress port mask, it may be that a
CPU port absorbes control traffic from user ports which aren't affine to
it as per user space's request. This isn't ideal, but is the lesser of
two evils. Following the user-configured affinity for traps would mean
that we can no longer reuse a single TCAM entry for multiple traps,
which is what we actually do for e.g. PTP. Either we duplicate and
deduplicate TCAM entries on the fly when user-to-CPU-port mappings
change (which is unnecessarily complicated), or we redirect trapped
traffic to all tag_8021q CPU ports if multiple such ports are in use.
The latter would have actually been nice, if it actually worked, but it
doesn't, since a OCELOT_MASK_MODE_REDIRECT action towards multiple ports
would not take PGID_SRC into consideration, and it would just duplicate
the packet towards each (CPU) port, leading to duplicates in software.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
DSA has not supported (and probably will not support in the future
either) independent tagging protocols per CPU port.
Different switch drivers have different requirements, some may need to
replicate some settings for each CPU port, some may need to apply some
settings on a single CPU port, while some may have to configure some
global settings and then some per-CPU-port settings.
In any case, the current model where DSA calls ->change_tag_protocol for
each CPU port turns out to be impractical for drivers where there are
global things to be done. For example, felix calls dsa_tag_8021q_register(),
which makes no sense per CPU port, so it suppresses the second call.
Let drivers deal with replication towards all CPU ports, and remove the
CPU port argument from the function prototype.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Acked-by: Luiz Angelo Daros de Luca <luizluca@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
At the time - commit 7569459a52 ("net: dsa: manage flooding on the CPU
ports") - not introducing a dedicated switch callback for host flooding
made sense, because for the only user, the felix driver, there was
nothing different to do for the CPU port than set the flood flags on the
CPU port just like on any other bridge port.
There are 2 reasons why this approach is not good enough, however.
(1) Other drivers, like sja1105, support configuring flooding as a
function of {ingress port, egress port}, whereas the DSA
->port_bridge_flags() function only operates on an egress port.
So with that driver we'd have useless host flooding from user ports
which don't need it.
(2) Even with the felix driver, support for multiple CPU ports makes it
difficult to piggyback on ->port_bridge_flags(). The way in which
the felix driver is going to support host-filtered addresses with
multiple CPU ports is that it will direct these addresses towards
both CPU ports (in a sort of multicast fashion), then restrict the
forwarding to only one of the two using the forwarding masks.
Consequently, flooding will also be enabled towards both CPU ports.
However, ->port_bridge_flags() gets passed the index of a single CPU
port, and that leaves the flood settings out of sync between the 2
CPU ports.
This is to say, it's better to have a specific driver method for host
flooding, which takes the user port as argument. This solves problem (1)
by allowing the driver to do different things for different user ports,
and problem (2) by abstracting the operation and letting the driver do
whatever, rather than explicitly making the DSA core point to the CPU
port it thinks needs to be touched.
This new method also creates a problem, which is that cross-chip setups
are not handled. However I don't have hardware right now where I can
test what is the proper thing to do, and there isn't hardware compatible
with multi-switch trees that supports host flooding. So it remains a
problem to be tackled in the future.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For symmetry with host FDBs and MDBs where the indirection is now
handled outside the ocelot switch lib, do the same for bridge port
flags (unicast/multicast/broadcast flooding).
The only caller of the ocelot switch lib which uses the NPI port is the
Felix DSA driver.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For symmetry with host FDBs where the indirection is now handled outside
the ocelot switch lib, do the same for host MDB entries. The only caller
of the ocelot switch lib which uses the NPI port is the Felix DSA driver.
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Vladimir Oltean