Re: [PATCH net-next V2 09/10] net/mlx5: LAG, drop MLX5_MAX_PORTS scratch array from device rescan
From: Shay Drory
Date: Thu Nov 19 2026 - 03:42:42 EST
On Wed, 18 Nov 2026 10:31:59 +0000, netdev-bot+sashiko@xxxxxxxxxx wrote:
> Thank you for your contribution! Sashiko AI review found 1 potential
> issue(s) to consider.
>
> Critical: 0 · High: 1 · 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] In mlx5_lag_rescan_devices_locked_filter()
> (drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c), this patch…
>
> --
>
> net/mlx5: LAG, drop MLX5_MAX_PORTS scratch array from device rescan
>
> This commit removes the devs[MLX5_MAX_PORTS] array from
> mlx5_lag_rescan_devices_locked_filter(). The post-unlock rescan loop now
> walks the LAG members again with mlx5_lag_for_each(). The commit message
> describes this as having no functional change.
>
> > diff --git a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
> > index a7b12353b7225..3e88b3f51612e 100644
> > --- a/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
> > +++ b/drivers/net/ethernet/mellanox/mlx5/core/lag/lag.c
>
> [ ... ]
>
> > @@ -1356,12 +1354,16 @@ static void mlx5_lag_rescan_devices_locked_filter(struct mlx5_lag *ldev,
> > fn->dev->priv.flags &= ~MLX5_PRIV_FLAGS_DISABLE_IB_ADEV;
> > else
> > fn->dev->priv.flags |= MLX5_PRIV_FLAGS_DISABLE_IB_ADEV;
> > - devs[num_devs++] = fn->dev;
> > }
> >
> > mlx5_lag_drop_lock_for_reps(ldev, filter);
> > - for (i = 0; i < num_devs; i++)
> > - mlx5_rescan_drivers_locked(devs[i]);
> > + mlx5_lag_for_each(i, 0, ldev, filter) {
> > + fn = mlx5_lag_fn(ldev, i);
> > + if (fn->dev->priv.flags & MLX5_PRIV_FLAGS_DISABLE_ALL_ADEV)
> > + continue;
> > +
> > + mlx5_rescan_drivers_locked(fn->dev);
> > + }
>
> [Severity: High]
> This loop now walks ldev->fns after mlx5_lag_drop_lock_for_reps() has
> released ldev->lock. Can a concurrent SD member removal free a lag_fn
> while the walk is in progress?
Valid. The race between SD join/leave and a LAG transition that drops
ldev->lock is pre-existing; SD join/leave doesn't observe
mode_changes_in_progress. It will be fixed in a follow-up series.
>
> On each step, mlx5_get_next_lag_func() reads xarray entries and fn->dev.
> mlx5_lag_fn() then does an xa_load(), and the loop dereferences
> fn->dev->priv.flags. The only thing guarding membership while the lock is
> dropped is ldev->mode_changes_in_progress. The comment in
> mlx5_lag_drop_lock_for_reps() says:
>
> /* Keep PF membership stable while ldev->lock is dropped. Device add
> * and remove paths observe mode_changes_in_progress and retry.
> */
>
> The SD paths in lib/sd.c don't appear to check that counter, though.
> sd_lag_cleanup() takes the lock directly:
>
> mutex_lock(&ldev->lock);
> mlx5_sd_for_each_secondary(i, primary, pos)
> mlx5_ldev_remove_mdev(ldev, pos);
>
> mlx5_ldev_remove_mdev() then frees the entry right away, with no RCU
> deferral:
>
> fn->dev = NULL;
> dev->priv.lag = NULL;
> xa_erase(&ldev->fns, fn->idx);
> kfree(fn);
>
> sd_lag_cleanup() is serialized only by sd->devcom. For
> MLX5_LAG_FILTER_ALL, this function runs under the HCA devcom instead.
> mlx5_do_bond_work() takes only the HCA devcom trylock plus ldev->lock, and
> not mlx5_mpesw_sd_devcoms_lock(). It reaches this function as follows:
>
> mlx5_do_bond_work()
> mlx5_do_bond()
> mlx5_lag_shared_fdb_create(ldev, ..., MLX5_LAG_MODE_SRIOV, 0)
> mlx5_lag_remove_devices_filter(ldev, MLX5_LAG_FILTER_ALL)
> mlx5_lag_rescan_devices_locked_filter()
>
> mlx5_disable_lag() -> mlx5_lag_shared_fdb_destroy(ldev, 0) reaches it the
> same way. Group id 0 maps to MLX5_LAG_FILTER_ALL, which matches every
> member, SD secondaries included.
>
> At the same time, removing an SD member's eth aux driver can run:
>
> mlx5e_remove() / mlx5e_suspend()
> mlx5_sd_cleanup()
> sd_lag_cleanup()
> mlx5_ldev_remove_mdev()
>
> This removal can come from a sysfs unbind of mlx5_core.eth.N. It can also
> come from a reload of an SD secondary whose hca_devcom_comp is NULL or in a
> different component. Neither case holds the HCA devcom that the bond work
> holds.
>
> If that happens, mlx5_get_next_lag_func() could read fn->dev from a freed
> lag_fn. mlx5_lag_fn() could also return NULL after xa_erase(), and the
> loop would then dereference it.
>
> The old loop touched only the devs[] core_dev pointers once the lock was
> dropped. Those pointers stay valid across an aux unbind, so lag_fn and
> xarray state was never accessed without the lock.
>
> A similar question applies to sd_lag_init(), which also takes ldev->lock
> without checking mode_changes_in_progress and calls mlx5_ldev_add_mdev().
> Could the second loop then rescan a newly added member whose
> MLX5_PRIV_FLAGS_DISABLE_IB_ADEV flag the first loop never updated?
>
> Given the above, is the "no functional change" statement in the commit
> message accurate?
>
> Would it be safer to keep a snapshot taken under the lock? For example, a
> dynamically sized array, or a reference on each core_dev collected before
> mlx5_lag_drop_lock_for_reps().
>
> As far as I can tell, the last patch in the series doesn't touch this
> function or the SD paths, so this would still be present at the end of the
> series.
> > mlx5_lag_retake_lock_after_reps(ldev);
> > }
> >
>
> --
> Sashiko AI review · https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261116103010.1129666-1-tariqt%40nvidia.com