Re: [PATCH] blk-mq: fix blk_mq_tagset_busy_iter

From: Ming Lei
Date: Thu Aug 02 2018 - 13:24:09 EST


On Thu, Aug 02, 2018 at 10:18:38AM -0700, James Bottomley wrote:
> On Thu, 2018-08-02 at 11:08 -0600, Jens Axboe wrote:
> > On 8/2/18 11:06 AM, Ming Lei wrote:
> > > On Thu, Aug 02, 2018 at 09:54:06AM -0700, James Bottomley wrote:
> > > > On Fri, 2018-08-03 at 00:43 +0800, Ming Lei wrote:
> > > > > Commit d250bf4e776ff09d5("blk-mq: only iterate over inflight
> > > > > requests
> > > > > in blk_mq_tagset_busy_iter") uses 'blk_mq_rq_state(rq) ==
> > > > > MQ_RQ_IN_FLIGHT' to replace 'blk_mq_request_started(req)', this
> > > > > way is wrong, and causes lots of test system hang during
> > > > > booting.
> > > > >
> > > > > Fix the issue by using blk_mq_request_started(req) inside
> > > > > bt_tags_iter().
> > > > >
> > > > > Fixes: d250bf4e776ff09d5 ("blk-mq: only iterate over inflight
> > > > > requests in blk_mq_tagset_busy_iter")
> > > > > Cc: Josef Bacik <josef@xxxxxxxxxxxxxx>
> > > > > Cc: Christoph Hellwig <hch@xxxxxx>
> > > > > Cc: Guenter Roeck <linux@xxxxxxxxxxxx>
> > > > > Cc: Mark Brown <broonie@xxxxxxxxxx>
> > > > > Cc: Matt Hart <matthew.hart@xxxxxxxxxx>
> > > > > Cc: Johannes Thumshirn <jthumshirn@xxxxxxx>
> > > > > Cc: John Garry <john.garry@xxxxxxxxxx>
> > > > > Cc: Hannes Reinecke <hare@xxxxxxxx>,
> > > > > Cc: "Martin K. Petersen" <martin.petersen@xxxxxxxxxx>,
> > > > > Cc: James Bottomley <James.Bottomley@xxxxxxxxxxxxxxxxxxxxx>
> > > > > Cc: linux-scsi@xxxxxxxxxxxxxxx
> > > > > Cc: linux-kernel@xxxxxxxxxxxxxxx
> > > > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
> > > > > ---
> > > > >  block/blk-mq-tag.c | 2 +-
> > > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > >
> > > > > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
> > > > > index 09b2ee6694fb..3de0836163c2 100644
> > > > > --- a/block/blk-mq-tag.c
> > > > > +++ b/block/blk-mq-tag.c
> > > > > @@ -271,7 +271,7 @@ static bool bt_tags_iter(struct sbitmap
> > > > > *bitmap,
> > > > > unsigned int bitnr, void *data)
> > > > >    * test and set the bit before assining ->rqs[].
> > > > >    */
> > > > >   rq = tags->rqs[bitnr];
> > > > > - if (rq && blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT)
> > > > > + if (rq && blk_mq_request_started(rq))
> > > >
> > > > So now we have dueling versions of this patch:
> > > >
> > > > https://marc.info/?l=linux-scsi&m=153322802207688
> > > >
> > > > Can we at least make sure we've root caused the problem and
> > > > confirmed we've got it fixed before we start the formal patch
> > > > process?  When we
> > >
> > > EH uses scsi_host_busy to check if the error handler needs to be
> > > waken up. And blk_mq_tagset_busy_iter() is used for implementing
> > > scsi_host_busy(), so causes EH not waken up, then this timed-out
> > > request can't be handled.
>
> Yes, I know what the problem is and why this patch is necessary and
> that it is very likely the root cause. However, can we confirm that it
> fixes the boot hang completely before we declare victory?

Frankly speaking, I can reproduce & verify it, but still suggest to wait
for ack from our report guys.

>
> > > > do start the formal patch process, please give appropriate credit
> > > > to the reporter(s) since this has been a royal pain for them to
> > > > help us track down.
> > >
> > > Sure.
> > >
> > > Jens, could you add reported-by if you are fine with this version?
> > > Or please just let me know if new version is needed, then I can add
> > > it.
> >
> > I'll add that, would also love a tested-by from the reporter. The
> > patch looks good to me, however.
>
> Is there a reason why blk_mq_request_started() isn't a static inline?
> It looks to be somewhat in the hot path.

Looks good idea, which should have been in header file, will do it in
V2.

Thanks,
Ming