Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Wed, 27 Jul 2022 10:11:57 +0800
From:   Ming Lei <ming.lei@redhat.com>
To:     Yu Kuai <yukuai1@huaweicloud.com>
Cc:     axboe@kernel.dk, osandov@fb.com, linux-block@vger.kernel.org,
        linux-kernel@vger.kernel.org, yukuai3@huawei.com,
        yi.zhang@huawei.com
Subject: Re: [PATCH v3] blk-mq: fix io hung due to missing commit_rqs
Message-ID: <YuCe7e5KapIqzO0S@T590>
References: <20220726122224.1790882-1-yukuai1@huaweicloud.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220726122224.1790882-1-yukuai1@huaweicloud.com>
Precedence: bulk

On Tue, Jul 26, 2022 at 08:22:24PM +0800, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Currently, in virtio_scsi, if 'bd->last' is not set to true while
> dispatching request, such io will stay in driver's queue, and driver
> will wait for block layer to dispatch more rqs. However, if block
> layer failed to dispatch more rq, it should trigger commit_rqs to
> inform driver.
> 
> There is a problem in blk_mq_try_issue_list_directly() that commit_rqs
> won't be called:
> 
> // assume that queue_depth is set to 1, list contains two rq
> blk_mq_try_issue_list_directly
>  blk_mq_request_issue_directly
>  // dispatch first rq
>  // last is false
>   __blk_mq_try_issue_directly
>    blk_mq_get_dispatch_budget
>    // succeed to get first budget
>    __blk_mq_issue_directly
>     scsi_queue_rq
>      cmd->flags |= SCMD_LAST
>       virtscsi_queuecommand
>        kick = (sc->flags & SCMD_LAST) != 0
>        // kick is false, first rq won't issue to disk
>  queued++
> 
>  blk_mq_request_issue_directly
>  // dispatch second rq
>   __blk_mq_try_issue_directly
>    blk_mq_get_dispatch_budget
>    // failed to get second budget
>  ret == BLK_STS_RESOURCE
>   blk_mq_request_bypass_insert
>  // errors is still 0
> 
>  if (!list_empty(list) || errors && ...)
>   // won't pass, commit_rqs won't be called
> 
> In this situation, first rq relied on second rq to dispatch, while
> second rq relied on first rq to complete, thus they will both hung.
> 
> Fix the problem by also treat 'BLK_STS_*RESOURCE' as 'errors' since
> it means that request is not queued successfully.
> 
> Same problem exists in blk_mq_dispatch_rq_list(), 'BLK_STS_*RESOURCE'
> can't be treated as 'errors' here, fix the problem by calling
> commit_rqs if queue_rq return 'BLK_STS_*RESOURCE'.
> 
> Fixes: d666ba98f849 ("blk-mq: add mq_ops->commit_rqs()")
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> ---
> Changes in v3:
>  - as poinited out by Ming, v2 will break return value by treating
>  BLK_STS_*RESOURCE as errors.
> Changes in v2:
>  - suggested by Ming, handle blk_mq_dispatch_rq_list() as well.
>  - change title and modify commit message.
> 
>  block/blk-mq.c | 5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 70177ee74295..7d26b222cbc9 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -1931,7 +1931,8 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *list,
>  	/* If we didn't flush the entire list, we could have told the driver
>  	 * there was more coming, but that turned out to be a lie.
>  	 */
> -	if ((!list_empty(list) || errors) && q->mq_ops->commit_rqs && queued)
> +	if ((!list_empty(list) || errors || needs_resource ||
> +	     ret == BLK_STS_DEV_RESOURCE) && q->mq_ops->commit_rqs && queued)

'ret == BLK_STS_DEV_RESOURCE' may happen before, but 'ret' could be
overridden later.

Thanks,
Ming