[dpdk-dev] [PATCH v2] net/mlx5: fix memory regions release deadlock

[Michael Baum]

The mpx5 PMD maintains the list of devices for those the memory
operation callback routines must be invoked to keep the device MRs (MR
is the entity backing the hardware DMA transactions) consistent with the
mapped memory.
Each device context in the list is protected with dedicated lock on per
device basis, which might be taken inside the callback routine.

When device is closing the PMD frees all MRs by calling
mlx5_mr_release(), that might call rte_free() under the taken device
lock.  If this rte_free call triggers the entire memory segment freeing
it, in its turn, invokes the callback routine and attempt to take the
lock inside this one causes the deadlock.

The patch proposes the remove the device from the callback list first
and then call mlx5_mr_release() and free the remaining device MRs
explicitely.

Fixes: 0e3d0525b2f2 ("net/mlx5: fix memory event callback list")
Cc: stable at dpdk.org

Signed-off-by: Michael Baum <michaelba at mellanox.com>
Acked-by: Viacheslav Ovsiienko <viacheslavo at mellanox.com>
Acked-by: Matan Azrad <matan at mellanox.com>
---

v2:
rephrase commit masage.


 drivers/net/mlx5/mlx5.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index f80e403..759491f 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -679,12 +679,12 @@ struct mlx5_flow_id_pool *
 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
 	if (--sh->refcnt)
 		goto exit;
-	/* Release created Memory Regions. */
-	mlx5_mr_release(sh);
 	/* Remove from memory callback device list. */
 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
 	LIST_REMOVE(sh, mem_event_cb);
 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
+	/* Release created Memory Regions. */
+	mlx5_mr_release(sh);
 	/* Remove context from the global device list. */
 	LIST_REMOVE(sh, next);
 	/*
-- 
1.8.3.1