Hi,
After 50b2412b7e7862c5af0cbf4b10d93bc5c712d021 was backported to stable branches (I only tested 5.4), some serious issues started to arrise.
According to linux-rdma, the following two patches that need to go along with 50b2412b7e are missing:
- 1d5558b1f0de net/mlx5: poll cmd EQ in case of command timeout
- 410bd754cd73 net/mlx5: Add retry mechanism to the command entry ...
I managed to apply those mostly cleanly after also applying two dependencies. So the complete list of needed commits for 5.4 is:
1. 3ed879965cc4 net/mlx5: Use async EQ setup cleanup helpers ... 2. 1d5558b1f0de net/mlx5: poll cmd EQ in case of command timeout 3. d43b7007dbd1 net/mlx5: Fix a race when moving command ... 4. 410bd754cd73 net/mlx5: Add retry mechanism to the command entry ...
With those 4 commits applied, the issue is fixed. For reference, that's the output I get with 5.4.77:
Nov 17 01:12:58 store01 kernel: mlx5_ib: Mellanox Connect-IB Infiniband driver v5.0-0 Nov 17 01:12:58 store01 kernel: mlx5_core 0000:01:00.0: cmd_work_handler:887:(pid 383): failed to allocate command entry Nov 17 01:12:58 store01 kernel: infiniband mlx5_0: reg_mr_callback:104:(pid 383): async reg mr failed. status -11 Nov 17 01:12:58 store01 kernel: mlx5_core 0000:01:00.0: cmd_work_handler:887:(pid 383): failed to allocate command entry Nov 17 01:12:58 store01 kernel: mlx5_core 0000:01:00.0: mlx5e_create_mdev_resources:104:(pid 1): alloc td failed, -11 Nov 17 01:12:58 store01 kernel: mlx5_0, 1: ipoib_intf_alloc failed -11