XSK wakeup must use the async ICOSQ (with proper locking), as it is not
guaranteed to run on the same CPU as the channel.
The commit that converted the NAPI trigger path to use the sync ICOSQ
incorrectly applied the same change to XSK, causing XSK wakeups to use
the sync ICOSQ as well. Revert XSK flows to use the async ICOSQ.
XDP program attach/detach triggers channel reopen, while XSK pool
enable/disable can happen on-the-fly via NDOs without reopening
channels. As a result, xsk_pool state cannot be reliably used at
mlx5e_open_channel() time to decide whether an async ICOSQ is needed.
Update the async_icosq_needed logic to depend on the presence of an XDP
program rather than the xsk_pool, ensuring the async ICOSQ is available
when XSK wakeups are enabled.
This fixes multiple issues:
1. Illegal synchronize_rcu() in an RCU read- side critical section via
mlx5e_xsk_wakeup() -> mlx5e_trigger_napi_icosq() ->
synchronize_net(). The stack holds RCU read-lock in xsk_poll().
2. Hitting a NULL pointer dereference in mlx5e_xsk_wakeup():
[] BUG: kernel NULL pointer dereference, address:
0000000000000240
[] #PF: supervisor read access in kernel mode
[] #PF: error_code(0x0000) - not-present page
[] PGD 0 P4D 0
[] Oops: Oops: 0000 [#1] SMP
[] CPU: 0 UID: 0 PID: 2255 Comm: qemu-system-x86 Not tainted 6.19.0-rc5+ #229 PREEMPT(none)
[] Hardware name: [...]
[] RIP: 0010:mlx5e_xsk_wakeup+0x53/0x90 [mlx5_core]
Reported-by: Daniel Borkmann <daniel@iogearbox.net>
Closes: https://lore.kernel.org/all/20260123223916.361295-1-daniel@iogearbox.net/
Fixes: 56aca3e0f730 ("net/mlx5e: Use regular ICOSQ for triggering NAPI")
Tested-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Dragos Tatulea <dtatulea@nvidia.com>
Acked-by: Alice Mikityanska <alice.kernel@fastmail.im>
Link: https://patch.msgid.link/20260217074525.1761454-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
int mlx5e_close_locked(struct net_device *netdev);
void mlx5e_trigger_napi_icosq(struct mlx5e_channel *c);
+void mlx5e_trigger_napi_async_icosq(struct mlx5e_channel *c);
void mlx5e_trigger_napi_sched(struct napi_struct *napi);
int mlx5e_open_channels(struct mlx5e_priv *priv,
goto err_remove_pool;
mlx5e_activate_xsk(c);
- mlx5e_trigger_napi_icosq(c);
+ mlx5e_trigger_napi_async_icosq(c);
/* Don't wait for WQEs, because the newer xdpsock sample doesn't provide
* any Fill Ring entries at the setup stage.
c = priv->channels.c[ix];
mlx5e_activate_rq(&c->rq);
- mlx5e_trigger_napi_icosq(c);
+ mlx5e_trigger_napi_async_icosq(c);
mlx5e_wait_for_min_rx_wqes(&c->rq, MLX5E_RQ_WQES_TIMEOUT);
mlx5e_rx_res_xsk_update(priv->rx_res, &priv->channels, ix, false);
&c->async_icosq->state))
return 0;
- mlx5e_trigger_napi_icosq(c);
+ mlx5e_trigger_napi_async_icosq(c);
}
return 0;
void mlx5e_trigger_napi_icosq(struct mlx5e_channel *c)
{
+ struct mlx5e_icosq *sq = &c->icosq;
bool locked;
- if (!test_and_set_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &c->icosq.state))
- synchronize_net();
+ set_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &sq->state);
+ synchronize_net();
- locked = mlx5e_icosq_sync_lock(&c->icosq);
- mlx5e_trigger_irq(&c->icosq);
- mlx5e_icosq_sync_unlock(&c->icosq, locked);
+ locked = mlx5e_icosq_sync_lock(sq);
+ mlx5e_trigger_irq(sq);
+ mlx5e_icosq_sync_unlock(sq, locked);
- clear_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &c->icosq.state);
+ clear_bit(MLX5E_SQ_STATE_LOCK_NEEDED, &sq->state);
+}
+
+void mlx5e_trigger_napi_async_icosq(struct mlx5e_channel *c)
+{
+ struct mlx5e_icosq *sq = c->async_icosq;
+
+ spin_lock_bh(&sq->lock);
+ mlx5e_trigger_irq(sq);
+ spin_unlock_bh(&sq->lock);
}
void mlx5e_trigger_napi_sched(struct napi_struct *napi)
netif_napi_add_config_locked(netdev, &c->napi, mlx5e_napi_poll, ix);
netif_napi_set_irq_locked(&c->napi, irq);
- async_icosq_needed = !!xsk_pool || priv->ktls_rx_was_enabled;
+ async_icosq_needed = !!params->xdp_prog || priv->ktls_rx_was_enabled;
err = mlx5e_open_queues(c, params, cparam, async_icosq_needed);
if (unlikely(err))
goto err_napi_del;
projects / thirdparty / kernel / linux.git / commitdiff
