Allow SD devices to transition to switchdev before the SD group is
fully up. Metadata allocation requires the SD group to be ready, so
defer it from esw_offloads_enable() until SD shared-FDB activation.
Add mlx5_esw_offloads_init_deferred_metadata() which allocates per-vport
metadata and refreshes the ingress ACLs that were previously programmed
with metadata=0. The helper is idempotent and can be called multiple
times.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-15-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
On an SD device, vport representors are not functional until the SD
group is combined and shared FDB is active. Skip the initial load and
the reload paths in that window; reps are loaded as part of the SD LAG
activation flow once it becomes active.
In addition, explicitly unload representors when SD LAG is destroyed.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-14-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Enable MPESW LAG creation over SD LAG members, forming a composite LAG
hierarchy. This allows bonding multiple SD groups together under a
single MPESW configuration with shared FDB.
When enabling composite MPESW, the individual SD LAG shared FDB
configurations are temporarily torn down and recreated when the
composite LAG is disabled.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-13-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
SD LAG is a virtual LAG without hardware LAG support, so it cannot use
the firmware vport LAG commands. Implement a software-based vport LAG
using egress ACL bounce rules.
Add esw_set_slave_egress_rule() to create an egress ACL rule on the
slave's manager vport that bounces traffic to the master's manager
vport. This achieves the same traffic steering as hardware vport LAG.
Redirect mlx5_cmd_create_vport_lag() and mlx5_cmd_destroy_vport_lag()
to the software implementation when operating in SD LAG mode.
In addition, adjust lag_demux creation to check SD LAG mode as well.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-12-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Extend mlx5_lag_disable_change() to properly disable both regular LAG
and SD LAG when requested. Each LAG type uses its own devcom component
for locking.
Use mlx5_sd_get_devcom() helper to retrieve the SD devcom component,
needed for proper locking when disabling SD LAG.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-11-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The lag demux resources (flow table, flow group, and rules xarray)
are stored on the shared ldev. With Socket Direct, multiple SD groups
each create their own demux FT/FG during their master's IB device
initialization. Since they all write to the same ldev fields, the
second group's init overwrites the first group's pointers, leaking
the first group's FT/FG.
During teardown, the cleanup uses the overwritten pointers, destroying
the wrong group's resources and leaving leaked flow tables in the LAG
namespace. These leaked tables can interfere with subsequently created
demux tables.
Move the demux resources from the shared ldev to per-master lag_func
instances. Each master device now owns its own independent demux
state. The rule_add and rule_del helpers look up the appropriate
master's lag_func via the existing filter/group infrastructure.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-10-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
When the eswitch transitions, propagate the change to SD: secondaries
get their TX flow table root reconfigured for the new mode, and when
all group devices move to switchdev, the per-group shared FDB is
activated.
Shared FDB activation is best-effort - failure does not block the
eswitch transition; the next transition retries.
Note: the existing mlx5_get_sd() guard that blocks switchdev for SD
devices is intentionally retained. It will be removed once all
supporting patches are in place.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-8-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
In Socket Direct configurations the primary and secondary PFs share the
same native_port_num. The eswitch vport metadata encodes pf_num in its
upper bits to distinguish vports across PFs. Without SD-awareness, both
PFs generate identical metadata, causing FDB rules to steer traffic to
the wrong representor.
Add mlx5_sd_pf_num_get() which remaps the pf_num for SD devices.
Use it so each PF in an SD group produces unique vport metadata.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-7-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add mlx5_fs_cmd_query_l2table_silent() to query the current silent mode
state from firmware. This allows detecting if firmware has already put
secondary devices into silent mode.
During SD group registration, query the silent mode of each device. If
a device is already in silent mode (set by firmware), record this in
the fw_silents_secondaries flag and use it to help determine the
primary/secondary roles.
When fw_silents_secondaries is set, skip the driver-initiated silent
mode set/unset operations since firmware manages this state. This
handles configurations where firmware persistently silences secondary
devices.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-6-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Refactor SD group registration to use devcom event-driven role
determination to ensure SD is marked as ready only after roles are fully
assigned and the group state is consistent, making outside accessors,
which will be added in downstream patches, safe to use without races.
The devcom events:
- SD_PRIMARY_SET event: each device compares bus numbers with peers
to determine which should be primary
- SD_SECONDARIES_SET event: secondaries register themselves with the
elected primary device
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-5-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Some devcom events are not expected to fail. Rather than attempting
a rollback that may not be meaningful, allow callers to pass
DEVCOM_CANT_FAIL as the rollback_event to indicate that the event
handler should not fail. If it does, emit a warning and stop
propagating to further peers, but skip the rollback path.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-4-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Factor mlx5_devcom_send_event() into two functions:
- mlx5_devcom_locked_send_event(): performs the dispatch (and
rollback) with comp->sem already held by the caller.
- mlx5_devcom_send_event(): unchanged wrapper that takes comp->sem,
calls the locked variant, and releases it.
This lets callers bracket multiple event broadcasts under a single
held write lock, eliminating the gap between consecutive dispatches
where peer state could change.
Will be used by a downstream patch.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Mark Bloch <mbloch@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260612113904.537595-3-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
GRO_LEGACY_MAX_SIZE = 65536; total_len being 65536 is too big to fit
into a u16. As can be seen in skb_gro_receive, packets bigger or equal
to gro_max_size (or GRO_LEGACY_MAX_SIZE) are dropped with -E2BIG. Apply
the same boundary to geneve_post_decap_hint to avoid writing 65536 to a
16-bit iph->tot_len field with an overflow.
Fixes: fd0dd79657 ("geneve: use GRO hint option in the RX path")
Signed-off-by: Alice Mikityanska <alice@isovalent.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260611192955.604661-3-alice.kernel@fastmail.im
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently, the tc tool only supports adding and deleting rules from
the driver but does not support querying rules from the driver.
This patch adds a rule dump file in debugfs to check whether the driver's
configuration matches the configuration issued by tc flow.
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260610060618.834987-6-shaojijie@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, the driver does not support FLOW_DISSECTOR_KEY_IP and
FLOW_DISSECTOR_KEY_ENC_KEYID. But the hardware supports
ip_tos (FLOW_DISSECTOR_KEY_IP) and
outer_tun_vni (FLOW_DISSECTOR_KEY_ENC_KEYID).
This patch adds support for FLOW_DISSECTOR_KEY_IP and
FLOW_DISSECTOR_KEY_ENC_KEYID.
Additionally, since tc flow cannot effectively support
l2_user_def, l3_user_def, and l4_user_def,
this patch explicitly sets them to not be used.
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260610060618.834987-5-shaojijie@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, the driver supports only one action:HCLGE_FD_ACTION_SELECT_TC.
This patch adds support for HCLGE_FD_ACTION_SELECT_QUEUE and
HCLGE_FD_ACTION_DROP_PACKET.
A rule can have only one action. Therefore, the driver intercepts rules
that have multiple actions or no action.
Note: The driver considers cls_flower->classid as an action:
HCLGE_FD_ACTION_SELECT_TC.
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260610060618.834987-4-shaojijie@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently, when the tc tool is used to set flow table rules, the IP address
and MAC address can be configured separately, for example, src_xx or dst_xx
can be configured separately.
Therefore, the driver needs to check whether the mask is all zero in
keys, such as FLOW_DISSECTOR_KEY_IPV4_ADDRS, FLOW_DISSECTOR_KEY_IPV6_ADDRS,
and FLOW_DISSECTOR_KEY_ETH_ADDRS.
If the mask is all zero, the tuple is not configured.
In this case, the driver adds the tuple to unused_tuple.
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260610060618.834987-3-shaojijie@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Remove the tc parameter from the add_cls_flower() ops callback and
refactor action parsing to support future extensions for SELECT_QUEUE
and DROP_PACKET actions.
Changes:
* Remove the tc parameter from the add_cls_flower() callback signature.
* Extract TC-based action parsing into hclge_get_tc_flower_action().
* Move the dissector->used_keys check from hclge_parse_cls_flower() to
hclge_check_cls_flower(), and restrict ETH_ADDRS to
HCLGE_FD_MODE_DEPTH_2K_WIDTH_400B_STAGE_1 mode since hardware only
supports MAC matching there.
* Migrate error reporting from dev_err() to netlink extended ACK (extack).
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260610060618.834987-2-shaojijie@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
For both the join and leave paths, the logic goes through the following
steps: determines which FDB should be used on a port after the current
changeupper change, populate the private port structures with the new
FDB and, if necessary, make as not used the old FDB.
Instead of having two distinct paths inside the
dpaa2_switch_port_set_fdb() for linking=true and linking=false, unify
them. This will hopefully help in making this function easier to read.
No behavior changes are expected.
Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/20260610150912.1788482-6-ioana.ciornei@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Move the FDB selection for when a port leaves bridge into a new helper -
dpaa2_switch_fdb_for_leave(). This will hopefully make the
dpaa2_switch_port_set_fdb() function easier to read and follow. The new
helper only determines the FDB to be used, any updates into the private
port structure still gets done in the set_fdb() function.
No changes in the actual behavior are intended.
Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/20260610150912.1788482-5-ioana.ciornei@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The dpaa2_switch_port_set_fdb() function handles the setup of the FDB
for both changeupper cases: join and leave. Move the code block which
handles the join path into a new helper - dpaa2_switch_fdb_for_join() -
with the hope that the entire function will become easier to read and
extend with other use cases in the future.
This new helper just determines and returns what FDB should be used for
a specific port, the cleanup of the old FDB and the actual setup in the
per port structure remains in the dpaa2_switch_port_set_fdb() function.
No changes in the actual behavior are intended.
Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/20260610150912.1788482-4-ioana.ciornei@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since there dpaa2_switch_port_set_fdb() never fails and its return value
was never checked, change its prototype to return void.
Also, instead of determining if the DPAA2 port is joining or leaving an
upper based on the value of the 'bridge_dev' parameter, add the
'linking' parameter to explicitly specify the action.
Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
Link: https://patch.msgid.link/20260610150912.1788482-2-ioana.ciornei@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
[Why]
In e1000_set_eeprom(), the eeprom_buff is allocated to hold a range of
words. However, only the boundary words (the first and the last) are
populated from the EEPROM if the write request is not word-aligned.
The words in the middle of the buffer remain uninitialized because they
are intended to be completely overwritten by the new data via memcpy().
The previous implementation had a loop that performed le16_to_cpus()
on the entire buffer. This resulted in endianness conversion being
performed on uninitialized memory for all interior words.
Fix this by converting the endianness only for the boundary words
immediately after they are successfully read from the EEPROM.
Found by Linux Verification Center (linuxtesting.org) with SVACE.
Co-developed-by: Iskhakov Daniil <dish@amicon.ru>
Signed-off-by: Iskhakov Daniil <dish@amicon.ru>
Signed-off-by: Agalakov Daniil <ade@amicon.ru>
Tested-by: Avigail Dahan <avigailx.dahan@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260609213559.178657-14-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
[Why]
In e1000_set_eeprom(), the eeprom_buff is allocated to hold a range of
words. However, only the boundary words (the first and the last) are
populated from the EEPROM if the write request is not word-aligned.
The words in the middle of the buffer remain uninitialized because they
are intended to be completely overwritten by the new data via memcpy().
The previous implementation had a loop that performed le16_to_cpus()
on the entire buffer. This resulted in endianness conversion being
performed on uninitialized memory for all interior words.
Fix this by converting the endianness only for the boundary words
immediately after they are successfully read from the EEPROM.
Found by Linux Verification Center (linuxtesting.org) with SVACE.
Co-developed-by: Iskhakov Daniil <dish@amicon.ru>
Signed-off-by: Iskhakov Daniil <dish@amicon.ru>
Signed-off-by: Agalakov Daniil <ade@amicon.ru>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260609213559.178657-13-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The __napi_schedule_irqoff() macro is intended to bypass saving and
restoring IRQ state when scheduling is requested from an IRQ handler,
where hard interrupts are already disabled. Use this macro in all three
interrupt handlers.
This was tested on a system with an I218-V and MSI interrupts. Because
this is an optimization, I was interested in measuring the impact, so I
added ktime_get() time measurement to e1000_intr_msi and a print of the
last sample in the watchdog task. For each test case I ran a
bi-directional iperf3 to saturate the line. With some help from awk,
here are the statistics.
49 samples each, all units ns
previous: min 678 max 1265 mean 879.429 median 806 stddev 137.188
noirq: min 707 max 1165 mean 811.857 median 790 stddev 89.486
According to this informal comparison, the mean time to handle an
interrupt from start to finish is improved by about 8% under load.
Signed-off-by: Matt Vollrath <tactii@gmail.com>
Reviewed-by: Aleksandr Loktionov <aleksandr.loktionov@intel.com>
Tested-by: Michal Cohen <michalx.cohen@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260609213559.178657-12-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This patch continues the effort to refactor workqueue APIs, which has begun
with the changes introducing new workqueues and a new alloc_workqueue flag:
commit 128ea9f6cc ("workqueue: Add system_percpu_wq and system_dfl_wq")
commit 930c2ea566 ("workqueue: Add new WQ_PERCPU flag")
The point of the refactoring is to eventually alter the default behavior of
workqueues to become unbound by default so that their workload placement is
optimized by the scheduler.
Before that to happen, workqueue users must be converted to the better named
new workqueues with no intended behaviour changes:
system_wq -> system_percpu_wq
system_unbound_wq -> system_dfl_wq
This way the old obsolete workqueues (system_wq, system_unbound_wq) can be
removed in the future.
Link: https://lore.kernel.org/all/20250221112003.1dSuoGyc@linutronix.de/
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Tested-by: Samuel Salin <Samuel.salin@intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
Link: https://patch.msgid.link/20260609213559.178657-2-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace the file-scope static npc_priv with a kcalloc'd struct filled
from hardware bank/subbank geometry at init (num_banks is no longer a
const compile-time constant; drop init_done and use a non-NULL
npc_priv pointer for liveness). Thread npc_priv_get() / pointer access
through the CN20K NPC code paths, extend teardown to kfree the root
struct on failure and in npc_cn20k_deinit, and adjust MCAM section
setup to use the discovered subbank count.
Allocate MCAM debugfs dstats via devm_kzalloc instead of a static matrix,
and use the allocated backing store consistently when computing deltas
(including the counter rollover compare).
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
Link: https://patch.msgid.link/20260609040453.711932-10-rkannoth@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Default CN20K NPC rule allocation now keys off the active MCAM keyword
width: use X4 with a bank-masked reference index when the silicon uses
X4 keys, and X2 with the raw index otherwise (replacing the previous
always-X2 / eidx + 1 behaviour).
In the AF flow-install path, flows that need more than 256 key bits
query the NPC profile; if the platform is fixed to X2 entries, fail
with -EOPNOTSUPP instead of requesting X4. Otherwise select X4 for the
MCAM alloc.
On the PF, cache and pass the profile kw_type from npc_get_pfl_info
through otx2_mcam_pfl_info_get(), and use it when allocating MCAM
entries for RSS/defaults and when installing ethtool flows on CN20K,
including masking the reference index for X4 slot layout.
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
Link: https://patch.msgid.link/20260609040453.711932-9-rkannoth@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Flashing updated firmware on deployed devices is cumbersome. Provide a
mechanism to load a custom KPU (Key Parse Unit) profile directly from
the filesystem at module load time.
When the rvu_af module is loaded with the kpu_profile parameter, the
specified profile is read from /lib/firmware/kpu and programmed into
the KPU registers. Add npc_kpu_profile_cam2 for the extended cam format
used by filesystem-loaded profiles and support ptype/ptype_mask in
npc_config_kpucam when profile->from_fs is set.
Usage:
1. Copy the KPU profile file to /lib/firmware/kpu.
2. Build OCTEONTX2_AF as a module.
3. Load: insmod rvu_af.ko kpu_profile=<profile_name>
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
Link: https://patch.msgid.link/20260609040453.711932-8-rkannoth@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Add NIX_LF_DONT_FREE_DFT_IDXS so the PF can send NIX LF free during hw
reinit or teardown without the AF freeing CN20K default NPC rule indexes
while the driver still owns that state (otx2_init_hw_resources and
otx2_free_hw_resources).
On CN20K, allocate default NPC rules from NIX LF alloc before
nix_interface_init, roll back with npc_cn20k_dft_rules_free on failure,
and free from NIX LF free when the new flag is not set. Tighten
rvu_mbox_handler_nix_lf_alloc error handling: use a single rc, propagate
qmem_alloc and other errors, and set -ENOMEM only when kcalloc fails
(remove the blanket -ENOMEM at the free_mem path).
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
Link: https://patch.msgid.link/20260609040453.711932-7-rkannoth@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
CN20K NPC MCAM is split into 32 subbanks that are searched in a
predefined order during allocation. Lower-numbered subbanks have
higher priority than higher-numbered ones.
Add a runtime "srch_order" to control the order in which
subbanks are searched during MCAM allocation.
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
Link: https://patch.msgid.link/20260609040453.711932-6-rkannoth@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
