net/mlx5e: Fix possible deadlock on mlx5e_tx_timeout_work
authorMoshe Shemesh <moshe@nvidia.com>
Wed, 21 Sep 2022 15:45:11 +0000 (18:45 +0300)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 20 Dec 2023 16:01:44 +0000 (17:01 +0100)
[ Upstream commit eab0da38912ebdad922ed0388209f7eb0a5163cd ]

Due to the cited patch, devlink health commands take devlink lock and
this may result in deadlock for mlx5e_tx_reporter as it takes local
state_lock before calling devlink health report and on the other hand
devlink health commands such as diagnose for same reporter take local
state_lock after taking devlink lock (see kernel log below).

To fix it, remove local state_lock from mlx5e_tx_timeout_work() before
calling devlink_health_report() and take care to cancel the work before
any call to close channels, which may free the SQs that should be
handled by the work. Before cancel_work_sync(), use current_work() to
check we are not calling it from within the work, as
mlx5e_tx_timeout_work() itself may close the channels and reopen as part
of recovery flow.

While removing state_lock from mlx5e_tx_timeout_work() keep rtnl_lock to
ensure no change in netdev->real_num_tx_queues, but use rtnl_trylock()
and a flag to avoid deadlock by calling cancel_work_sync() before
closing the channels while holding rtnl_lock too.

Kernel log:
======================================================
WARNING: possible circular locking dependency detected
6.0.0-rc3_for_upstream_debug_2022_08_30_13_10 #1 Not tainted
------------------------------------------------------
kworker/u16:2/65 is trying to acquire lock:
ffff888122f6c2f8 (&devlink->lock_key#2){+.+.}-{3:3}, at: devlink_health_report+0x2f1/0x7e0

but task is already holding lock:
ffff888121d20be0 (&priv->state_lock){+.+.}-{3:3}, at: mlx5e_tx_timeout_work+0x70/0x280 [mlx5_core]

which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #1 (&priv->state_lock){+.+.}-{3:3}:
       __mutex_lock+0x12c/0x14b0
       mlx5e_rx_reporter_diagnose+0x71/0x700 [mlx5_core]
       devlink_nl_cmd_health_reporter_diagnose_doit+0x212/0xa50
       genl_family_rcv_msg_doit+0x1e9/0x2f0
       genl_rcv_msg+0x2e9/0x530
       netlink_rcv_skb+0x11d/0x340
       genl_rcv+0x24/0x40
       netlink_unicast+0x438/0x710
       netlink_sendmsg+0x788/0xc40
       sock_sendmsg+0xb0/0xe0
       __sys_sendto+0x1c1/0x290
       __x64_sys_sendto+0xdd/0x1b0
       do_syscall_64+0x3d/0x90
       entry_SYSCALL_64_after_hwframe+0x46/0xb0

-> #0 (&devlink->lock_key#2){+.+.}-{3:3}:
       __lock_acquire+0x2c8a/0x6200
       lock_acquire+0x1c1/0x550
       __mutex_lock+0x12c/0x14b0
       devlink_health_report+0x2f1/0x7e0
       mlx5e_health_report+0xc9/0xd7 [mlx5_core]
       mlx5e_reporter_tx_timeout+0x2ab/0x3d0 [mlx5_core]
       mlx5e_tx_timeout_work+0x1c1/0x280 [mlx5_core]
       process_one_work+0x7c2/0x1340
       worker_thread+0x59d/0xec0
       kthread+0x28f/0x330
       ret_from_fork+0x1f/0x30

other info that might help us debug this:

 Possible unsafe locking scenario:

       CPU0                    CPU1
       ----                    ----
  lock(&priv->state_lock);
                               lock(&devlink->lock_key#2);
                               lock(&priv->state_lock);
  lock(&devlink->lock_key#2);

 *** DEADLOCK ***

4 locks held by kworker/u16:2/65:
 #0: ffff88811a55b138 ((wq_completion)mlx5e#2){+.+.}-{0:0}, at: process_one_work+0x6e2/0x1340
 #1: ffff888101de7db8 ((work_completion)(&priv->tx_timeout_work)){+.+.}-{0:0}, at: process_one_work+0x70f/0x1340
 #2: ffffffff84ce8328 (rtnl_mutex){+.+.}-{3:3}, at: mlx5e_tx_timeout_work+0x53/0x280 [mlx5_core]
 #3: ffff888121d20be0 (&priv->state_lock){+.+.}-{3:3}, at: mlx5e_tx_timeout_work+0x70/0x280 [mlx5_core]

stack backtrace:
CPU: 1 PID: 65 Comm: kworker/u16:2 Not tainted 6.0.0-rc3_for_upstream_debug_2022_08_30_13_10 #1
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014
Workqueue: mlx5e mlx5e_tx_timeout_work [mlx5_core]
Call Trace:
 <TASK>
 dump_stack_lvl+0x57/0x7d
 check_noncircular+0x278/0x300
 ? print_circular_bug+0x460/0x460
 ? find_held_lock+0x2d/0x110
 ? __stack_depot_save+0x24c/0x520
 ? alloc_chain_hlocks+0x228/0x700
 __lock_acquire+0x2c8a/0x6200
 ? register_lock_class+0x1860/0x1860
 ? kasan_save_stack+0x1e/0x40
 ? kasan_set_free_info+0x20/0x30
 ? ____kasan_slab_free+0x11d/0x1b0
 ? kfree+0x1ba/0x520
 ? devlink_health_do_dump.part.0+0x171/0x3a0
 ? devlink_health_report+0x3d5/0x7e0
 lock_acquire+0x1c1/0x550
 ? devlink_health_report+0x2f1/0x7e0
 ? lockdep_hardirqs_on_prepare+0x400/0x400
 ? find_held_lock+0x2d/0x110
 __mutex_lock+0x12c/0x14b0
 ? devlink_health_report+0x2f1/0x7e0
 ? devlink_health_report+0x2f1/0x7e0
 ? mutex_lock_io_nested+0x1320/0x1320
 ? trace_hardirqs_on+0x2d/0x100
 ? bit_wait_io_timeout+0x170/0x170
 ? devlink_health_do_dump.part.0+0x171/0x3a0
 ? kfree+0x1ba/0x520
 ? devlink_health_do_dump.part.0+0x171/0x3a0
 devlink_health_report+0x2f1/0x7e0
 mlx5e_health_report+0xc9/0xd7 [mlx5_core]
 mlx5e_reporter_tx_timeout+0x2ab/0x3d0 [mlx5_core]
 ? lockdep_hardirqs_on_prepare+0x400/0x400
 ? mlx5e_reporter_tx_err_cqe+0x1b0/0x1b0 [mlx5_core]
 ? mlx5e_tx_reporter_timeout_dump+0x70/0x70 [mlx5_core]
 ? mlx5e_tx_reporter_dump_sq+0x320/0x320 [mlx5_core]
 ? mlx5e_tx_timeout_work+0x70/0x280 [mlx5_core]
 ? mutex_lock_io_nested+0x1320/0x1320
 ? process_one_work+0x70f/0x1340
 ? lockdep_hardirqs_on_prepare+0x400/0x400
 ? lock_downgrade+0x6e0/0x6e0
 mlx5e_tx_timeout_work+0x1c1/0x280 [mlx5_core]
 process_one_work+0x7c2/0x1340
 ? lockdep_hardirqs_on_prepare+0x400/0x400
 ? pwq_dec_nr_in_flight+0x230/0x230
 ? rwlock_bug.part.0+0x90/0x90
 worker_thread+0x59d/0xec0
 ? process_one_work+0x1340/0x1340
 kthread+0x28f/0x330
 ? kthread_complete_and_exit+0x20/0x20
 ret_from_fork+0x1f/0x30
 </TASK>

Fixes: c90005b5f75c ("devlink: Hold the instance lock in health callbacks")
Signed-off-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/net/ethernet/mellanox/mlx5/core/en.h
drivers/net/ethernet/mellanox/mlx5/core/en_main.c

index 86f2690..20a6bc1 100644 (file)
@@ -818,6 +818,7 @@ enum {
        MLX5E_STATE_DESTROYING,
        MLX5E_STATE_XDP_TX_ENABLED,
        MLX5E_STATE_XDP_ACTIVE,
+       MLX5E_STATE_CHANNELS_ACTIVE,
 };
 
 struct mlx5e_modify_sq_param {
index acb4077..c3961c2 100644 (file)
@@ -2668,6 +2668,7 @@ void mlx5e_close_channels(struct mlx5e_channels *chs)
 {
        int i;
 
+       ASSERT_RTNL();
        if (chs->ptp) {
                mlx5e_ptp_close(chs->ptp);
                chs->ptp = NULL;
@@ -2945,17 +2946,29 @@ void mlx5e_activate_priv_channels(struct mlx5e_priv *priv)
        if (mlx5e_is_vport_rep(priv))
                mlx5e_rep_activate_channels(priv);
 
+       set_bit(MLX5E_STATE_CHANNELS_ACTIVE, &priv->state);
+
        mlx5e_wait_channels_min_rx_wqes(&priv->channels);
 
        if (priv->rx_res)
                mlx5e_rx_res_channels_activate(priv->rx_res, &priv->channels);
 }
 
+static void mlx5e_cancel_tx_timeout_work(struct mlx5e_priv *priv)
+{
+       WARN_ON_ONCE(test_bit(MLX5E_STATE_CHANNELS_ACTIVE, &priv->state));
+       if (current_work() != &priv->tx_timeout_work)
+               cancel_work_sync(&priv->tx_timeout_work);
+}
+
 void mlx5e_deactivate_priv_channels(struct mlx5e_priv *priv)
 {
        if (priv->rx_res)
                mlx5e_rx_res_channels_deactivate(priv->rx_res);
 
+       clear_bit(MLX5E_STATE_CHANNELS_ACTIVE, &priv->state);
+       mlx5e_cancel_tx_timeout_work(priv);
+
        if (mlx5e_is_vport_rep(priv))
                mlx5e_rep_deactivate_channels(priv);
 
@@ -4734,8 +4747,17 @@ static void mlx5e_tx_timeout_work(struct work_struct *work)
        struct net_device *netdev = priv->netdev;
        int i;
 
-       rtnl_lock();
-       mutex_lock(&priv->state_lock);
+       /* Take rtnl_lock to ensure no change in netdev->real_num_tx_queues
+        * through this flow. However, channel closing flows have to wait for
+        * this work to finish while holding rtnl lock too. So either get the
+        * lock or find that channels are being closed for other reason and
+        * this work is not relevant anymore.
+        */
+       while (!rtnl_trylock()) {
+               if (!test_bit(MLX5E_STATE_CHANNELS_ACTIVE, &priv->state))
+                       return;
+               msleep(20);
+       }
 
        if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
                goto unlock;
@@ -4754,7 +4776,6 @@ static void mlx5e_tx_timeout_work(struct work_struct *work)
        }
 
 unlock:
-       mutex_unlock(&priv->state_lock);
        rtnl_unlock();
 }