On Thu, Nov 08, 2018 at 09:10:09PM +0200, Leon Romanovsky wrote:
> From: Moni Shoua <monis@xxxxxxxxxxxx>
>
> Fix 2 flows that may cause a process to hang on wait_for_completion():
>
> 1. When callback for create MKEY command returns with bad status
> 2. When callback for create MKEY command is called before executer of
> command calls wait_for_completion()
>
> The following call trace might be triggered in the above flows:
>
> INFO: task echo_server:1655 blocked for more than 120 seconds.
> "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> echo_server D ffff880813fb6898 0 1655 1 0x00000004
> ffff880423f5b880 0000000000000086 ffff880402290fd0 ffff880423f5bfd8
> ffff880423f5bfd8 ffff880423f5bfd8 ffff880402290fd0 ffff880813fb68a0
> 7fffffffffffffff ffff880813fb6898 ffff880402290fd0 ffff880813fb6898
> Call Trace:
> [<ffffffff816a94c9>] schedule+0x29/0x70
> [<ffffffff816a6fd9>] schedule_timeout+0x239/0x2c0
> [<ffffffffc07309e2>] ? mlx5_cmd_exec_cb+0x22/0x30 [mlx5_core]
> [<ffffffffc073e697>] ? mlx5_core_create_mkey_cb+0xb7/0x220 [mlx5_core]
> [<ffffffff811b94b7>] ? mm_drop_all_locks+0xd7/0x110
> [<ffffffff816a987d>] wait_for_completion+0xfd/0x140
> [<ffffffff810c4810>] ? wake_up_state+0x20/0x20
> [<ffffffffc08fd308>] mlx5_mr_cache_alloc+0xa8/0x170 [mlx5_ib]
> [<ffffffffc0909626>] implicit_mr_alloc+0x36/0x190 [mlx5_ib]
> [<ffffffffc090a26e>] mlx5_ib_alloc_implicit_mr+0x4e/0xa0 [mlx5_ib]
> [<ffffffffc08ff2f3>] mlx5_ib_reg_user_mr+0x93/0x6a0 [mlx5_ib]
> [<ffffffffc0907410>] ? mlx5_ib_exp_query_device+0xab0/0xbc0 [mlx5_ib]
> [<ffffffffc04998be>] ib_uverbs_exp_reg_mr+0x2fe/0x550 [ib_uverbs]
> [<ffffffff811edaff>] ? do_huge_pmd_anonymous_page+0x2bf/0x530
> [<ffffffffc048f6cc>] ib_uverbs_write+0x3ec/0x490 [ib_uverbs]
> [<ffffffff81200d2d>] vfs_write+0xbd/0x1e0
> [<ffffffff81201b3f>] SyS_write+0x7f/0xe0
> [<ffffffff816b4fc9>] system_call_fastpath+0x16/0x1b
>
> Fixes: 49780d42dfc9 ("IB/mlx5: Expose MR cache for mlx5_ib")
> Signed-off-by: Moni Shoua <monis@xxxxxxxxxxxx>
> Reviewed-by: Majd Dibbiny <majd@xxxxxxxxxxxx>
> Signed-off-by: Leon Romanovsky <leonro@xxxxxxxxxxxx>
> drivers/infiniband/hw/mlx5/mlx5_ib.h | 1 +
> drivers/infiniband/hw/mlx5/mr.c | 15 ++++++++++++---
> 2 files changed, 13 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
> index b651a7a6fde9..cd9335e368bd 100644
> +++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
> @@ -644,6 +644,7 @@ struct mlx5_cache_ent {
> struct delayed_work dwork;
> int pending;
> struct completion compl;
> + atomic_t do_complete;
> };
>
> struct mlx5_mr_cache {
> diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
> index 9b195d65a13e..259dd49c6874 100644
> +++ b/drivers/infiniband/hw/mlx5/mr.c
> @@ -143,7 +143,7 @@ static void reg_mr_callback(int status, void *context)
> kfree(mr);
> dev->fill_delay = 1;
> mod_timer(&dev->delay_timer, jiffies + HZ);
> - return;
> + goto do_complete;
> }
>
> mr->mmkey.type = MLX5_MKEY_MR;
> @@ -167,8 +167,13 @@ static void reg_mr_callback(int status, void *context)
> pr_err("Error inserting to mkey tree. 0x%x\n", -err);
> write_unlock_irqrestore(&table->lock, flags);
>
> - if (!completion_done(&ent->compl))
> +do_complete:
> + spin_lock_irqsave(&ent->lock, flags);
> + if (atomic_read(&ent->do_complete)) {
> complete(&ent->compl);
> + atomic_dec(&ent->do_complete);
> + }
> + spin_unlock_irqrestore(&ent->lock, flags);
Oh, this is quite an ugly way to use completions, I think this has
veered into misusing completion territory.. The completion_done was
never right...
add_keys should accept a flag indicating that this MR has a completor
waiting and should trigger complete() on CB finishing... Can probably
store the flag someplace in the MR.
Jason
