Re: [Bug] double ->queue_rq() because of timeout in ->queue_rq()

From: Ming Lei <ming.lei@redhat.com>
To: Bart Van Assche <bvanassche@acm.org>
Cc: Jens Axboe <axboe@kernel.dk>, Christoph Hellwig <hch@lst.de>,
	djeffery@redhat.com, stefanha@redhat.com,
	linux-block@vger.kernel.org, linux-scsi@vger.kernel.org,
	virtualization@lists.linux-foundation.org
Subject: Re: [Bug] double ->queue_rq() because of timeout in ->queue_rq()
Date: Fri, 21 Oct 2022 08:57:44 +0800	[thread overview]
Message-ID: <Y1HuiFXyQ1k+OH92@T590> (raw)
In-Reply-To: <7d5eae39-3a56-df7d-eb72-3cb910c2b802@acm.org>

On Thu, Oct 20, 2022 at 01:26:48PM -0700, Bart Van Assche wrote:
> On 10/20/22 02:10, Ming Lei wrote:
> > [ ... ]
> 
> Hi Ming,
> 
> Fixing this in the block layer seems fine to me. A few comments:
> 
> > +	/* Before walking tags, we must ensure any submit started before the
> > +	 * current time has finished. Since the submit uses srcu or rcu, wait
> > +	 * for a synchronization point to ensure all running submits have
> > +	 * finished
> > +	 */
> 
> Should the above comment follow the style of other comments in the block
> layer?

OK.

> 
> > +	blk_mq_wait_quiesce_done(q);
> > +
> > +	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &expired);
> 
> The above doesn't look sufficient to me since .queue_rq() may be called
> while blk_mq_queue_tag_busy_iter() is in progress. How about moving the
> blk_mq_wait_quiesce_done() call into blk_mq_check_expired() and preventing
> new .queue_rq() calls before the timeout handler is called?

blk_mq_timeout_work() records the time before calling
blk_mq_wait_quiesce_done(), and only handle requests which is timed out
before the recorded jiffies, so new queued request won't be covered
in this time.

Thanks,
Ming