Re: [PATCH] blk-mq: avoid to iterate over stale request

[Ming Lei <ming.lei@redhat.com>]

On Thu, Dec 16, 2021 at 05:25:41PM +0530, Kashyap Desai wrote:
> > >
> > > I mean, shared bitmap in my whole discussion. Megaraid_sas driver use
> > > shared bitmap, so it is exposed and It is confirmed from this
> discussion.
> > > Do we still have exposure (if "blk-mq: avoid to iterate over stale
> > > request" is not part of kernel)  to mpi3mr type driver which does not
> > > use shared bitmap but has nr_hw_queues > 1. ?
> >
> > Not sure I understand your poing, but patch "blk-mq: avoid to iterate
> over
> > stale request" can cover both shared tags or not.
> 
> I agree with all above reply.
> 
> My query is for mpi3mr driver which is not calling "host->host_tagset =
> 1;", but nr_hw_queues are more than 1.
> Current <mpi3mr> driver is nvme style interface. nr_hw_queues  > 1 and
> host->host_tagset = 0.
> 
> Is this patch  "blk-mq: avoid to iterate over stale request" is applicable
> for <mpi3mr> driver ?

Yeah, this change covers both ->host_tagset and !->host_tagset, same
with the following words:

	scsi_host_check_in_flight() is used to counting scsi in-flight requests
	after scsi host is blocked, so no new scsi command can be marked as
	SCMD_STATE_INFLIGHT. However, driver tag allocation still can be run at
	that time by blk-mq core.
	
	The issue is in blk_mq_tagset_busy_iter(). One request is in-flight, but
	this request may be kept in another slot of ->rqs[], meantime the slot
	can be allocated out but ->rqs[] isn't updated yet. Then the in-flight
	request is counted twice as SCMD_STATE_INFLIGHT.



Thanks,
Ming