Re: Panic when rebooting target server testing srp on 5.0.0-rc2

[Bart Van Assche <bvanassche@acm.org>]

On Wed, 2019-04-03 at 14:47 -0700, Sagi Grimberg wrote:
> > I have not been able to reproduce this issue. But you may want to try
> > the following patch (applies on top of v5.1-rc3):
> > 
> > 
> > Subject: [PATCH] block: Fix blk_mq_try_issue_directly()
> > 
> > If blk_mq_try_issue_directly() returns BLK_STS*_RESOURCE that means that
> > the request has not been queued and that the caller should retry to submit
> > the request. Both blk_mq_request_bypass_insert() and
> > blk_mq_sched_insert_request() guarantee that a request will be processed.
> > Hence return BLK_STS_OK if one of these functions is called. This patch
> > avoids that blk_mq_dispatch_rq_list() crashes when using dm-mpath.
> > 
> > Reported-by: Laurence Oberman <loberman@redhat.com>
> > Fixes: 7f556a44e61d ("blk-mq: refactor the code of issue request directly") # v5.0.
> > Signed-off-by: Bart Van Assche <bvanassche@acm.org>
> > ---
> >   block/blk-mq.c | 9 ++-------
> >   1 file changed, 2 insertions(+), 7 deletions(-)
> > 
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index 652d0c6d5945..b2c20dce8a30 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -1859,16 +1859,11 @@ blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
> >   	case BLK_STS_RESOURCE:
> >   		if (force) {
> >   			blk_mq_request_bypass_insert(rq, run_queue);
> > -			/*
> > -			 * We have to return BLK_STS_OK for the DM
> > -			 * to avoid livelock. Otherwise, we return
> > -			 * the real result to indicate whether the
> > -			 * request is direct-issued successfully.
> > -			 */
> > -			ret = bypass ? BLK_STS_OK : ret;
> > +			ret = BLK_STS_OK;
> >   		} else if (!bypass) {
> >   			blk_mq_sched_insert_request(rq, false,
> >   						    run_queue, false);
> > +			ret = BLK_STS_OK;
> >   		}
> 
> I would move the ret assignment here after the condition statement.

How is that expected to work for the !force && bypass case? I don't
want to alter "ret" in that case.

Thanks,

Bart.