On Wed, Jan 17, 2018 at 10:49:58PM -0500, Mike Snitzer wrote:
> On Wed, Jan 17 2018 at 10:39pm -0500,
> Ming Lei <ming.lei@xxxxxxxxxx> wrote:
>
> > On Wed, Jan 17, 2018 at 10:33:35PM -0500, Mike Snitzer wrote:
> > > On Wed, Jan 17 2018 at 7:54P -0500,
> > > Mike Snitzer <snitzer@xxxxxxxxxx> wrote:
> > >
> > > > But sure, I suppose there is something I missed when refactoring Ming's
> > > > change to get it acceptable for upstream. I went over the mechanical
> > > > nature of what I did many times (comparing Ming's v4 to my v5).
> > >
> > > And yes there is one subtlety that I missed.
> > >
> > > > The call to blk_mq_request_bypass_insert will only occur via
> > > > __blk_mq_fallback_to_insert. Which as the name implies this is not the
> > > > fast path. This will occur if the underlying blk-mq device cannot get
> > > > resources it needs in order to issue the request. Specifically: if/when
> > > > in __blk_mq_try_issue_directly() the hctx is stopped, or queue is
> > > > quiesced, or it cannot get the driver tag or dispatch_budget (in the
> > > > case of scsi-mq).
> > > >
> > > > The same fallback, via call to blk_mq_request_bypass_insert, occured
> > > > with Ming's v4 though.
> > >
> > > Turns out Ming's v4 doesn't fallback to insert for the "or it cannot get
> > > the driver tag or dispatch_budget" case.
> > >
> > > This patch should fix it (Laurence, please report back on if this fixes
> > > your list_add corruption, pretty sure it will):
> > >
> > > From: Mike Snitzer <snitzer@xxxxxxxxxx>
> > > Date: Wed, 17 Jan 2018 22:02:07 -0500
> > > Subject: [PATCH] blk mq: don't blk_mq_request_bypass_insert _and_ return BLK_STS_RESOURCE
> > >
> > > It isn't ever valid to call blk_mq_request_bypass_insert() and return
> > > BLK_STS_RESOURCE.
> > >
> > > Unfortunately after commit 396eaf21ee ("blk-mq: improve DM's blk-mq IO
> > > merging via blk_insert_cloned_request feedback") we do just that if
> > > blk_mq_request_direct_issue() cannot get the resources (driver_tag or
> > > dispatch_budget) needed to directly issue a request. This will lead to
> > > "list_add corruption" because blk-mq submits the IO but then reports
> > > that it didn't (BLK_STS_RESOURCE in this case).
> > >
> > > Fix this by simply returning BLK_STS_RESOURCE for this case.
> > >
> > > Fixes: 396eaf21ee ("blk-mq: improve DM's blk-mq IO merging via blk_insert_cloned_request feedback")
> > > Reported-by: Laurence Oberman <loberman@xxxxxxxxxx>
> > > Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx>
> > > ---
> > > block/blk-mq.c | 12 +++++-------
> > > 1 file changed, 5 insertions(+), 7 deletions(-)
> > >
> > > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > > index c418858a60ef..8bee37239255 100644
> > > --- a/block/blk-mq.c
> > > +++ b/block/blk-mq.c
> > > @@ -1799,20 +1799,18 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
> > > if (q->elevator && !bypass_insert)
> > > goto insert;
> > >
> > > - if (!blk_mq_get_driver_tag(rq, NULL, false))
> > > - goto insert;
> > > -
> > > - if (!blk_mq_get_dispatch_budget(hctx)) {
> > > + if (!blk_mq_get_driver_tag(rq, NULL, false) ||
> > > + !blk_mq_get_dispatch_budget(hctx)) {
> > > + /* blk_mq_put_driver_tag() is idempotent */
> > > blk_mq_put_driver_tag(rq);
> > > + if (bypass_insert)
> > > + return BLK_STS_RESOURCE;
> > > goto insert;
> > > }
> > >
> > > return __blk_mq_issue_directly(hctx, rq, cookie);
> > > insert:
> > > __blk_mq_fallback_to_insert(rq, run_queue, bypass_insert);
> > > - if (bypass_insert)
> > > - return BLK_STS_RESOURCE;
> > > -
> > > return BLK_STS_OK;
> > > }
> >
> > Hi Mike,
> >
> > I'd suggest to use the following one, which is simple and clean:
> >
> >
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index 4d4af8d712da..816ff5d6bc88 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -1856,15 +1856,6 @@ static blk_status_t __blk_mq_issue_directly(struct blk_mq_hw_ctx *hctx,
> > return ret;
> > }
> >
> > -static void __blk_mq_fallback_to_insert(struct request *rq,
> > - bool run_queue, bool bypass_insert)
> > -{
> > - if (!bypass_insert)
> > - blk_mq_sched_insert_request(rq, false, run_queue, false);
> > - else
> > - blk_mq_request_bypass_insert(rq, run_queue);
> > -}
> > -
> > static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
> > struct request *rq,
> > blk_qc_t *cookie,
> > @@ -1892,10 +1883,10 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
> >
> > return __blk_mq_issue_directly(hctx, rq, cookie);
> > insert:
> > - __blk_mq_fallback_to_insert(rq, run_queue, bypass_insert);
> > if (bypass_insert)
> > return BLK_STS_RESOURCE;
> >
> > + blk_mq_sched_insert_request(rq, false, run_queue, false);
> > return BLK_STS_OK;
> > }
> >
> > @@ -1911,7 +1902,7 @@ static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
> >
> > ret = __blk_mq_try_issue_directly(hctx, rq, cookie, false);
> > if (ret == BLK_STS_RESOURCE)
> > - __blk_mq_fallback_to_insert(rq, true, false);
> > + blk_mq_sched_insert_request(rq, false, true, false);
> > else if (ret != BLK_STS_OK)
> > blk_mq_end_request(rq, ret);
> >
>
> That'd be another way to skin the cat.. BUT it is different than your
> original approach (due to the case I detailed in earlier mail).
Yeah, as I mentioned, it is a bug in this patch.
>
> But I like the simplicity of always returning BLK_STS_RESOURCE.
>
> Reviewed-by: Mike Snitzer <snitzer@xxxxxxxxxx>
>
> Laurance, please test Ming's patch instead. Ming, if Laurance finds
> this fix works (which is should) please put a formal header on the patch
> and submit for Jens to pick up. Sorry about screwing this up.
>
> (feel free to re-use portions of my above patch header to help explain
> the problem in your header)
Hi Mike & Laurance,
Please hold a moment, this one has one issue, and I will post a formal
one.
Thanks,
Ming
