On 07/29/2016 11:52 AM, Gabriel Krisman Bertazi wrote:

Jens Axboe [off-list ref] writes:

quoted

I can see that is an issue. Did you consider the case where
blk_mq_timeout_work() is entered, but we don't have any requests
allocated that currently hold a reference? This could happen if
completion races with a timeout.

Hi Jens,

That shouldn't be an issue, I think.  The only functional difference
should be during queue freezes, and in this case, if the request was
already completed by the time we touch it, we`ll only hold the queue
freeze for a little longer, until we release the reference by the end of
the timeout path.  should the final request be released before
blk_mq_timeout_work acquire it's reference, q_usage_counter will reach
zero, and the call to percpu_ref_tryget will fail the same way
percpu_ref_tryget_live would.

Thanks, I'm happy with it, and the detailed comment is a good
improvement. I have added it for this series.

-- 
Jens Axboe