Re: [PATCH] nvme: unquiesce the queue before cleaup it

[jianchao.wang]

Hi keith

Would you please take a look at this patch.

This issue could be reproduced easily with a driver bind/unbind loop,
a reset loop and a IO loop at the same time.

Thanks
Jianchao

On 04/19/2018 04:29 PM, Jianchao Wang wrote:
> There is race between nvme_remove and nvme_reset_work that can
> lead to io hang.
> 
> nvme_remove                    nvme_reset_work
> -> change state to DELETING
>                                -> fail to change state to LIVE
>                                -> nvme_remove_dead_ctrl
>                                  -> nvme_dev_disable
>                                    -> quiesce request_queue
>                                  -> queue remove_work
> -> cancel_work_sync reset_work
> -> nvme_remove_namespaces
>   -> splice ctrl->namespaces
>                                nvme_remove_dead_ctrl_work
>                                -> nvme_kill_queues
>   -> nvme_ns_remove               do nothing
>     -> blk_cleanup_queue
>       -> blk_freeze_queue
> Finally, the request_queue is quiesced state when wait freeze,
> we will get io hang here.
> 
> To fix it, unquiesce the request_queue directly before nvme_ns_remove.
> We have spliced the ctrl->namespaces, so nobody could access them
> and quiesce the queue any more.
> 
> Signed-off-by: Jianchao Wang <jianchao.w.wang@xxxxxxxxxx>
> ---
>  drivers/nvme/host/core.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> index 9df4f71..0e95082 100644
> --- a/drivers/nvme/host/core.c
> +++ b/drivers/nvme/host/core.c
> @@ -3249,8 +3249,15 @@ void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
>  	list_splice_init(&ctrl->namespaces, &ns_list);
>  	up_write(&ctrl->namespaces_rwsem);
>  
> -	list_for_each_entry_safe(ns, next, &ns_list, list)
> +	/*
> +	 * After splice the namespaces list from the ctrl->namespaces,
> +	 * nobody could get them anymore, let's unquiesce the request_queue
> +	 * forcibly to avoid io hang.
> +	 */
> +	list_for_each_entry_safe(ns, next, &ns_list, list) {
> +		blk_mq_unquiesce_queue(ns->queue);
>  		nvme_ns_remove(ns);
> +	}
>  }
>  EXPORT_SYMBOL_GPL(nvme_remove_namespaces);
>  
>