On Tue, 2018-05-22 at 14:38 -0600, Jens Axboe wrote:
> On 5/22/18 2:33 PM, Bart Van Assche wrote:
> > Thanks for having reported this. How about using the following change to suppress
> > that warning:
> >
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index bb99c03e7a34..84e55ea55baf 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -844,6 +844,7 @@ static void blk_mq_rq_timed_out(struct request *req, bool reserved)
> >
> > switch (ret) {
> > case BLK_EH_HANDLED:
> > + blk_mq_change_rq_state(req, MQ_RQ_TIMED_OUT, MQ_RQ_COMPLETE);
> > __blk_mq_complete_request(req);
> > break;
> > case BLK_EH_RESET_TIMER:
> >
> > I think this will work better than what was proposed in your last e-mail. I'm afraid
> > that with that change that a completion that occurs while the timeout handler is
> > running can be ignored.
>
> What if that races with eg requeue? We get the completion from IRQ, decide we
> need to requeue/restart the request rather than complete it.
Shouldn't block drivers that requeue a request from inside their timeout handler
return BLK_EH_NOT_HANDLED instead of BLK_EH_HANDLED?
Thanks,
Bart.
