* [PATCH net] net/mlx5: SF, remove probed SF devices when arming vhca events fails
@ 2026-11-17 3:11 Changyul Lee
2026-11-18 3:11 ` sashiko-bot
2026-11-19 18:12 ` netdev-bot+sashiko
0 siblings, 2 replies; 3+ messages in thread
From: Changyul Lee @ 2026-11-17 3:11 UTC (permalink / raw)
To: Saeed Mahameed, Leon Romanovsky, Tariq Toukan, Mark Bloch
Cc: Andrew Lunn, David S. Miller, Eric Dumazet, Jakub Kicinski,
Paolo Abeni, Parav Pandit, Shay Drory, netdev, linux-rdma,
linux-kernel, Changyul Lee
mlx5_sf_dev_table_create() queues work that probes the SFs which are
already active, and then arms the vhca state events. If arming fails,
the arm_err path stops and flushes that work and frees the table, but
the SF devices that the work already added stay on the auxiliary bus
and in table->devices. dev->priv.sf_dev_table is set to NULL, so
mlx5_sf_dev_table_destroy() returns early later and nothing removes
them, while they keep pointing at the parent mlx5_core_dev.
Remove them with mlx5_sf_dev_destroy_all() after the work is flushed,
as mlx5_sf_dev_table_destroy() does. Move the helper up so that it is
defined before the new caller.
Fixes: 4f57332d6a55 ("net/mlx5: SF: Fix probing active SFs during driver probe phase")
Signed-off-by: Changyul Lee <lcy8047@gmail.com>
---
Found with a static analysis tool we are developing.
Compile-tested only, with x86_64_defconfig plus MLX5_CORE, MLX5_CORE_EN,
MLX5_ESWITCH, MLX5_SF and W=1.
.../ethernet/mellanox/mlx5/core/sf/dev/dev.c | 23 ++++++++++---------
1 file changed, 12 insertions(+), 11 deletions(-)
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c b/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
index e75cbf333898..44f6292ad59e 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
@@ -328,6 +328,17 @@ int mlx5_sf_dev_notifier_init(struct mlx5_core_dev *dev)
return mlx5_vhca_event_notifier_register(dev, &dev->priv.sf_dev_nb);
}
+static void mlx5_sf_dev_destroy_all(struct mlx5_sf_dev_table *table)
+{
+ struct mlx5_sf_dev *sf_dev;
+ unsigned long index;
+
+ xa_for_each(&table->devices, index, sf_dev) {
+ xa_erase(&table->devices, index);
+ mlx5_sf_dev_remove_aux(table->dev, sf_dev);
+ }
+}
+
void mlx5_sf_dev_table_create(struct mlx5_core_dev *dev)
{
struct mlx5_sf_dev_table *table;
@@ -360,6 +371,7 @@ void mlx5_sf_dev_table_create(struct mlx5_core_dev *dev)
arm_err:
mlx5_sf_dev_destroy_active_works(table);
mlx5_vhca_event_work_queues_flush(dev);
+ mlx5_sf_dev_destroy_all(table);
add_active_err:
kfree(table);
dev->priv.sf_dev_table = NULL;
@@ -367,17 +379,6 @@ void mlx5_sf_dev_table_create(struct mlx5_core_dev *dev)
mlx5_core_err(dev, "SF DEV table create err = %d\n", err);
}
-static void mlx5_sf_dev_destroy_all(struct mlx5_sf_dev_table *table)
-{
- struct mlx5_sf_dev *sf_dev;
- unsigned long index;
-
- xa_for_each(&table->devices, index, sf_dev) {
- xa_erase(&table->devices, index);
- mlx5_sf_dev_remove_aux(table->dev, sf_dev);
- }
-}
-
void mlx5_sf_dev_notifier_cleanup(struct mlx5_core_dev *dev)
{
if (mlx5_core_is_sf(dev))
base-commit: 7475ea8b58616b33c04afb976cf0fbaad9dd63fb
--
2.43.0
^ permalink raw reply related [flat|nested] 3+ messages in thread
* Re: [PATCH net] net/mlx5: SF, remove probed SF devices when arming vhca events fails
2026-11-17 3:11 [PATCH net] net/mlx5: SF, remove probed SF devices when arming vhca events fails Changyul Lee
@ 2026-11-18 3:11 ` sashiko-bot
2026-11-19 18:12 ` netdev-bot+sashiko
1 sibling, 0 replies; 3+ messages in thread
From: sashiko-bot @ 2026-11-18 3:11 UTC (permalink / raw)
To: Changyul Lee; +Cc: linux-rdma
Thank you for your contribution! Sashiko AI review found 1 potential
issue to consider:
- [Medium] net/mlx5: race and UAF in mlx5_sf_dev_table_create() arm_err
cleanup
--
commit a5d70bb82124aa0963c3bb188af6126f523476a3
Author: Changyul Lee <lcy8047@gmail.com>
net/mlx5: SF, remove probed SF devices when arming vhca events fails
This commit calls mlx5_sf_dev_destroy_all() on the arm_err error path in
mlx5_sf_dev_table_create() so that SF auxiliary devices probed by the
active work are removed if arming VHCA events fails.
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c b/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
> index e75cbf333898d..44f6292ad59ec 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
[ ... ]
> @@ -360,6 +371,7 @@ void mlx5_sf_dev_table_create(struct mlx5_core_dev *dev)
> arm_err:
> mlx5_sf_dev_destroy_active_works(table);
> mlx5_vhca_event_work_queues_flush(dev);
> + mlx5_sf_dev_destroy_all(table);
> add_active_err:
> kfree(table);
> dev->priv.sf_dev_table = NULL;
[Severity: Medium]
Can mlx5_sf_dev_state_change_handler() race with mlx5_sf_dev_destroy_all()
and kfree(table) on the arm_err path?
If mlx5_sf_dev_vhca_arm_all() fails after arming one or more SF functions,
those functions remain armed in firmware while dev->priv.sf_dev_nb is still
registered and dev->priv.sf_dev_table still points to table. Unlike
mlx5_sf_dev_table_destroy(), which runs after mlx5_sf_dev_notifier_cleanup()
has unregistered the notifier, arm_err runs with the notifier active.
When mlx5_sf_dev_destroy_all() removes an already-probed SF, tearing down
its HCA transitions the function from MLX5_VHCA_STATE_IN_USE back to
MLX5_VHCA_STATE_ACTIVE, which can queue a new VHCA state event after
mlx5_vhca_event_work_queues_flush() has already returned:
CPU0 CPU1 (vhca_events wq)
mlx5_sf_dev_table_create()
mlx5_sf_dev_vhca_arm_all()
/* arms SF 0, fails on SF 1 */
mlx5_sf_dev_destroy_active_works()
mlx5_vhca_event_work_queues_flush()
mlx5_sf_dev_destroy_all(table)
xa_erase(&table->devices, 0)
mlx5_sf_dev_remove_aux()
/* SF 0 becomes ACTIVE,
fires VHCA event */
mlx5_sf_dev_state_change_handler()
table = dev->priv.sf_dev_table;
kfree(table)
xa_load(&table->devices, 0)
<-- use-after-free
dev->priv.sf_dev_table = NULL;
Even if the handler runs before kfree(table), seeing MLX5_VHCA_STATE_ACTIVE
after xa_erase() causes mlx5_sf_dev_add() to re-add the SF auxiliary device
that mlx5_sf_dev_destroy_all() just removed, and a state change to
ALLOCATED or TEARDOWN_REQUEST can call mlx5_sf_dev_del() concurrently with
mlx5_sf_dev_destroy_all().
Clearing dev->priv.sf_dev_table before mlx5_vhca_event_work_queues_flush()
(with a barrier or WRITE_ONCE/READ_ONCE pair) would ensure no handler runs
against table while mlx5_sf_dev_destroy_all() and kfree(table) execute.
--
Sashiko AI review · https://sashiko.dev/#/patchset/20261117031141.1720619-1-lcy8047@gmail.com?part=1
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH net] net/mlx5: SF, remove probed SF devices when arming vhca events fails
2026-11-17 3:11 [PATCH net] net/mlx5: SF, remove probed SF devices when arming vhca events fails Changyul Lee
2026-11-18 3:11 ` sashiko-bot
@ 2026-11-19 18:12 ` netdev-bot+sashiko
1 sibling, 0 replies; 3+ messages in thread
From: netdev-bot+sashiko @ 2026-11-19 18:12 UTC (permalink / raw)
To: lcy8047
Cc: saeedm, leon, tariqt, mbloch, andrew+netdev, davem, edumazet,
kuba, pabeni, parav, shayd, netdev, linux-rdma, linux-kernel
Thank you for your contribution! Sashiko AI review found 1 potential
issue(s) to consider.
Critical: 0 · High: 1 (2) · Medium: 0 · Low: 0
If you believe none of the issues are legitimate, please reply to each
issue of Medium severity or higher and explain why the AI is wrong.
If any of them are legitimate and you plan to respin, please let patchwork
know by including "pw-bot: cr" as a separate line at the end of your reply
(one such reply per series is enough).
- [High] The patch adds a mlx5_sf_dev_destroy_all() call to the arm_err
path of mlx5_sf_dev_table_create().
Pre-existing issues:
- [High] This problem wasn't introduced by this patch, but it sits on the
path the patch changes.
- [High] This problem wasn't introduced by this patch, but it is the same
synchronization gap the patch relies on.
--
net/mlx5: SF, remove probed SF devices when arming vhca events fails
This patch adds a mlx5_sf_dev_destroy_all() call to the arm_err path of
mlx5_sf_dev_table_create(). The goal is to remove SF auxiliary devices
that the active-SF probe work already added. It also moves the helper
above its new caller.
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c b/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
> index e75cbf333898d..44f6292ad59ec 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/sf/dev/dev.c
[ ... ]
> @@ -360,6 +371,7 @@ void mlx5_sf_dev_table_create(struct mlx5_core_dev *dev)
> arm_err:
> mlx5_sf_dev_destroy_active_works(table);
> mlx5_vhca_event_work_queues_flush(dev);
> + mlx5_sf_dev_destroy_all(table);
[Severity: High]
Is it safe to call mlx5_sf_dev_destroy_all() here while VHCA state change
events are still being delivered?
mlx5_load() calls mlx5_vhca_event_start(), which registers the VHCA EQ
notifier, before it calls mlx5_sf_dev_table_create(). sf_dev_nb was
registered even earlier. The arm_err path unregisters neither of them.
mlx5_vhca_event_work_queues_flush() only drains work that is already
queued.
mlx5_sf_dev_vhca_arm_all() also returns on the first failure, so the
functions armed before it stay armed. mlx5_vhca_event_notify() re-arms a
function every time it delivers an event:
mlx5_vhca_event_notify()
mlx5_vhca_event_arm(dev, event->function_id);
...
blocking_notifier_call_chain(&dev->priv.vhca_state_n_head, 0, event);
dev->priv.sf_dev_table is still set at this point, so
mlx5_sf_dev_state_change_handler() gets past its only check:
if (!table)
return 0;
...
sf_dev = xa_load(&table->devices, sf_index);
Could the handler's xa_load() and the xa_for_each() in
mlx5_sf_dev_destroy_all() both return the same sf_dev before either
xa_erase() runs? If so, both would call mlx5_sf_dev_remove_aux(). That
runs auxiliary_device_delete() and auxiliary_device_uninit() twice, and
the second caller uses sf_dev after mlx5_sf_dev_release() has freed it.
An MLX5_VHCA_STATE_ACTIVE event could also be handled while the loop runs,
or after it ends but before kfree(table). In that case mlx5_sf_dev_add()
can xa_insert() at an index the loop has already passed. Wouldn't that SF
stay on the auxiliary bus with parent_mdev pointing at the parent device?
That is the same leak the commit message says this patch fixes. If the
insert happens after kfree(table), it writes into a freed xarray.
The commit message says this matches mlx5_sf_dev_table_destroy(), but that
function depends on this:
/* Now that event handler is not running, it is safe to destroy
* the sf device without race.
*/
That holds in mlx5_unload() because mlx5_vhca_event_stop() runs first and
does both the unregister and the flush. This path only does the flush.
Should the handler be quiesced before the table is walked and freed, for
example by stopping event delivery or adding real locking? Setting
dev->priv.sf_dev_table to NULL earlier would not be enough on its own,
because an mlx5_sf_dev_add() that is already running still uses the table.
> add_active_err:
> kfree(table);
> dev->priv.sf_dev_table = NULL;
[Severity: High]
This is a pre-existing issue that was not introduced by this patch. Can
mlx5_sf_dev_state_change_handler() still use the table after this
kfree()?
The handler reads dev->priv.sf_dev_table once, with no lock and no RCU,
and only checks it for NULL. If a handler loads the pointer before the
NULL store, it can then call xa_load(), xa_erase() or xa_insert() on the
freed table->devices. mlx5_sf_dev_add() also reads freed table fields:
sf_dev->bar_base_addr = table->base_address + (sf_index * table->sf_bar_length);
Swapping the kfree() and the NULL store would not close this, because a
handler may already hold the old pointer.
This patch also widens the window: the slow mlx5_sf_dev_destroy_all() now
runs between the flush and kfree(table).
[Severity: High]
This isn't a bug introduced by this patch, but the mlx5_load() error
unwind has the same ordering problem. The err_traps_reg label destroys the
table before VHCA events are stopped:
drivers/net/ethernet/mellanox/mlx5/core/main.c:mlx5_load() {
...
err_traps_reg:
mlx5_sf_dev_table_destroy(dev);
mlx5_sriov_detach(dev);
err_sriov:
mlx5_lag_remove_mdev(dev);
mlx5_ec_cleanup(dev);
err_ec:
mlx5_vhca_event_stop(dev);
...
}
mlx5_unload() uses the opposite order:
mlx5_vhca_event_stop(dev);
mlx5_sf_dev_table_destroy(dev);
Suppose mlx5_devlink_traps_register() or mlx5_data_direct_init() fails
after mlx5_sf_dev_table_create() succeeded. Won't
mlx5_sf_dev_table_destroy() then run mlx5_sf_dev_destroy_all() and
kfree(table) while the EQ notifier is still registered and the vhca event
workqueues have not been flushed?
That breaks the "event handler is not running" precondition in the comment
in mlx5_sf_dev_table_destroy(). It opens the same races described above:
double removal of an sf_dev, a leaked SF, and use of the freed table.
A mlx5_sf_dev_add_active_work() item that already passed its
stop_active_wq check could also still be inside mlx5_sf_dev_add() while
mlx5_sf_dev_destroy_all() runs. Nothing on this path flushes the vhca
event workqueues first.
--
Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261117031141.1720619-1-lcy8047%40gmail.com
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2026-11-19 18:12 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-17 3:11 [PATCH net] net/mlx5: SF, remove probed SF devices when arming vhca events fails Changyul Lee
2026-11-18 3:11 ` sashiko-bot
2026-11-19 18:12 ` netdev-bot+sashiko
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox