Re: [PATCH 0/5] scsi: fnic: use blk_mq_tagset_busy_iter() to walk scsi commands

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

On Thu, Apr 22, 2021 at 08:08:41AM +0800, Ming Lei wrote:
> On Wed, Apr 21, 2021 at 10:14:56PM +0200, Hannes Reinecke wrote:
> > On 4/21/21 9:55 AM, Ming Lei wrote:
> > > Hello Guys,
> > > 
> > > fnic uses the following way to walk scsi commands in failure handling,
> > > which is obvious wrong, because caller of scsi_host_find_tag has to
> > > guarantee that the tag is active.
> > > 
> > >          for (tag = 0; tag < fnic->fnic_max_tag_id; tag++) {
> > > 				...
> > >                  sc = scsi_host_find_tag(fnic->lport->host, tag);
> > > 				...
> > > 		}
> > > 
> > > Fix the issue by using blk_mq_tagset_busy_iter() to walk
> > > request/scsi_command.
> > > 
> > > thanks,
> > > Ming
> > > 
> > > 
> > > Ming Lei (5):
> > >    scsi: fnic: use blk_mq_tagset_busy_iter() to walk scsi commands in
> > >      fnic_terminate_rport_io
> > >    scsi: fnic: use blk_mq_tagset_busy_iter() to walk scsi commands in
> > >      fnic_clean_pending_aborts
> > >    scsi: fnic: use blk_mq_tagset_busy_iter() to walk scsi commands in
> > >      fnic_cleanup_io
> > >    scsi: fnic: use blk_mq_tagset_busy_iter() to walk scsi commands in
> > >      fnic_rport_exch_reset
> > >    scsi: fnic: use blk_mq_tagset_busy_iter() to walk scsi commands in
> > >      fnic_is_abts_pending
> > > 
> > >   drivers/scsi/fnic/fnic_scsi.c | 933 ++++++++++++++++++----------------
> > >   1 file changed, 493 insertions(+), 440 deletions(-)
> > > 
> > > Cc: Satish Kharat <satishkh@cisco.com>
> > > Cc: Karan Tilak Kumar <kartilak@cisco.com>
> > > Cc: David Jeffery <djeffery@redhat.com>
> > > 
> > Well, this is actually not that easy for fnic.
> > Problem is the reset hack hch put in some time ago (cf
> > fnic_host_start_tag()), which will cause any TMF to use a tag which is _not_
> > visible to the busy iter.
> 
> 'git grep -n fnic_host_start_tag ./' shows nothing.
> 
> > That will cause the iter to miss any TMF, with unpredictable results if a
> > TMF is running at the same time than, say, a link bounce.
> 
> Wrt. linus tree or next tree, I don't see any issue wrt. your concern.
> 
> > 
> > I have folded this as part of my patchset for reserved commands in SCSI;
> > that way fnic can use 'normal' tags for TMFs, which are then visible to the
> > busy iter and life's good.
> 
> No, this fix is one bug fix, which can't depend on your reserved
> command in SCSI, and they need to be backported to stable tree too.

Hi Hannes,

We have customers report on this issue, could you please let us know
if you will post out one bug-fix only version?

Thanks,
Ming