On Mon, Dec 07, 2020 at 11:32:54AM +0800, chenxiang (M) wrote:
>
>
> 在 2020/12/3 9:03, Ming Lei 写道:
> > On Wed, Dec 02, 2020 at 10:10:30AM -0800, Bart Van Assche wrote:
> > > On 12/2/20 2:04 AM, Ming Lei wrote:
> > > > When queuing IO request to LLD, STS_RESOURCE may be returned because:
> > > >
> > > > - host in recovery or blocked
> > > > - target queue throttling or blocked
> > > > - LLD rejection
> > > >
> > > > Any one of the above doesn't happen frequently enough.
> > > >
> > > > BLK_STS_DEV_RESOURCE is returned to block layer for avoiding unnecessary
> > > > re-run queue, and it is just one small optimization. However, all
> > > > in-flight requests originated from this scsi device may be completed
> > > > just after reading 'sdev->device_busy', so BLK_STS_DEV_RESOURCE is
> > > > returned to block layer. And the current failed IO won't get chance
> > > > to be queued any more, since it is invisible at that time for either
> > > > scsi_run_queue_async() or blk-mq's RESTART.
> > > >
> > > > Fix the issue by not returning BLK_STS_DEV_RESOURCE in this situation.
> > > >
> > > > Cc: Hannes Reinecke <hare@xxxxxxxx>
> > > > Cc: Sumit Saxena <sumit.saxena@xxxxxxxxxxxx>
> > > > Cc: Kashyap Desai <kashyap.desai@xxxxxxxxxxxx>
> > > > Cc: Bart Van Assche <bvanassche@xxxxxxx>
> > > > Cc: Ewan Milne <emilne@xxxxxxxxxx>
> > > > Cc: Long Li <longli@xxxxxxxxxxxxx>
> > > > Tested-by: "chenxiang (M)" <chenxiang66@xxxxxxxxxxxxx>
> > > > Reported-by: John Garry <john.garry@xxxxxxxxxx>
> > > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
> > > > ---
> > > > drivers/scsi/scsi_lib.c | 3 +--
> > > > 1 file changed, 1 insertion(+), 2 deletions(-)
> > > >
> > > > diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> > > > index 60c7a7d74852..03c6d0620bfd 100644
> > > > --- a/drivers/scsi/scsi_lib.c
> > > > +++ b/drivers/scsi/scsi_lib.c
> > > > @@ -1703,8 +1703,7 @@ static blk_status_t scsi_queue_rq(struct blk_mq_hw_ctx *hctx,
> > > > break;
> > > > case BLK_STS_RESOURCE:
> > > > case BLK_STS_ZONE_RESOURCE:
> > > > - if (atomic_read(&sdev->device_busy) ||
> > > > - scsi_device_blocked(sdev))
> > > > + if (scsi_device_blocked(sdev))
> > > > ret = BLK_STS_DEV_RESOURCE;
> > > > break;
> > > > default:
> > > Since this patch modifies code introduced in commit 86ff7c2a80cd ("blk-mq:
> > > introduce BLK_STS_DEV_RESOURCE"), does this patch perhaps needs a Fixes:
> > > tag?
> > This same race exists before commit 86ff7c2a80cd, so I think the 'Fixes:' tag
> > is misleading.
>
> When reverted the patch "scsi: core: Only re-run queue in scsi_end_request()
> if device queue is busy", it also solves the issue.
> Does the issue is brought by the patch? If so, maybe adding fixes("Fixes:
> ed5dd6a67d5e ("scsi: core: Only re-run queue in scsi_end_request() if device
> queue is busy")") is more accuratte.
The debugfs log shows that the issue isn't related with restart or out
of budget because the single request stays in hctx->dispatch, and the flag of
RQF_DONTPREP is set too.
The scsi's restart(scsi_run_queue_async) from request completion is just
for handling requests in scheduler queue.
Thanks,
Ming
