Re: kernel BUG at drivers/scsi/scsi_lib.c:1096!

[Mike Snitzer <snitzer@redhat.com>]

On Wed, Nov 25 2015 at  4:04am -0500,
Hannes Reinecke <hare@suse.de> wrote:

> On 11/20/2015 04:28 PM, Ewan Milne wrote:
> > On Fri, 2015-11-20 at 15:55 +0100, Hannes Reinecke wrote:
> >> Can't we have a joint effort here?
> >> I've been spending a _LOT_ of time trying to debug things here, but
> >> none of the ideas I've come up with have been able to fix anything.
> > 
> > Yes.  I'm not the one primarily looking at it, and we don't have a
> > reproducer in-house.  We just have the one dump right now.
> > 
> >>
> >> I'm almost tempted to increase the count from scsi_alloc_sgtable()
> >> by one and be done with ...
> >>
> > 
> > That might not fix it if it is a problem with the merge code, though.
> > 
> And indeed, it doesn't.

How did you arrive at that?  Do you have a reproducer now?

> Seems I finally found the culprit.
> 
> What happens is this:
> We have two paths, with these seg_boundary_masks:
> 
> path-1:    seg_boundary_mask = 65535,
> path-2:    seg_boundary_mask = 4294967295,
> 
> consequently the DM request queue has this:
> 
> md-1:    seg_boundary_mask = 65535,
> 
> What happens now is that a request is being formatted, and sent
> to path 2. During submission req->nr_phys_segments is formatted
> with the limits of path 2, arriving at a count of 3.
> Now the request gets retried on path 1, but as the NOMERGE request
> flag is set req->nr_phys_segments is never updated.
> But blk_rq_map_sg() ignores all counters, and just uses the
> bi_vec directly, resulting in a count of 4 -> boom.
> 
> So the culprit here is the NOMERGE flag,

NOMERGE is always set in __blk_rq_prep_clone() for cloned requests.

> which is evaluated via
> ->dm_dispatch_request()
>   ->blk_insert_cloned_request()
>     ->blk_rq_check_limits()

blk_insert_cloned_request() is the only caller of blk_rq_check_limits();
anyway after reading your mail I'm still left wondering if your proposed
patch is correct.

> If the above assessment is correct, the following patch should
> fix it:
> 
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 801ced7..12cccd6 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -1928,7 +1928,7 @@ EXPORT_SYMBOL(submit_bio);
>   */
>  int blk_rq_check_limits(struct request_queue *q, struct request *rq)
>  {
> -       if (!rq_mergeable(rq))
> +       if (rq->cmd_type != REQ_TYPE_FS)
>                 return 0;
> 
>         if (blk_rq_sectors(rq) > blk_queue_get_max_sectors(q,
> rq->cmd_flags)) {
> 
> 
> Mike? Jens?
> Can you comment on it?

You're not explaining the actual change in the patch very well; I think
you're correct but you're leaving the justification as an exercise to 
the reviewer:

blk_rq_check_limits() will call blk_recalc_rq_segments() after the
!rq_mergeable() check but you're saying for this case in question we
never get there -- due to the cloned request having NOMERGE set.

So in blk_rq_check_limits() you've unrolled rq_mergeable() and
open-coded the lone remaining check (rq->cmd_type != REQ_TYPE_FS)

I agree that the (rq->cmd_flags & REQ_NOMERGE_FLAGS) check in
the blk_insert_cloned_request() call-chain (via rq_mergeable()) makes no
sense for cloned requests that always have NOMERGE set.

So you're saying that by having blk_rq_check_limits() go on to call
blk_recalc_rq_segments() this bug will be fixed?

BTW, I think blk_rq_check_limits()'s export should be removed and the
function made static and renamed to blk_clone_rq_check_limits(), again:
blk_insert_cloned_request() is the only caller of blk_rq_check_limits()

Seems prudent to make that change now to be clear that this code is only
used by cloned requests.

Thanks,
Mike