[dm-devel] Block regression since 3.1-rc3

[Shaohua Li]

2011/10/7 Jeff Moyer <jmoyer at redhat.com>:
> Christophe Saout <christophe at saout.de> writes:
>
>> Hi Jeff,
>>
>>> Anyway, it would help a great deal if you could retrigger the failure
>>> and provide the full failure output.  You can get that by issuing the
>>> 'dmesg' command and redirecting it to a file.
>>
>> Oh, sorry, yes, there's a line missing.
>>
>> Line 323 is this one: BUG_ON(!rq->bio || rq->bio != rq->biotail);
>
> OK, it turns out my testing was incomplete.  I only tested targets that
> had a write-through cache, so I didn't hit this problem.  It reproduces
> pretty easily with just multipath involved (no linear target on top) when
> running against the right storage.
>
> So, here's a patch, but I don't have a full explanation for it just yet.
> What I observed was that, on fsync, blkdev_issue_flush was called.
> Eventually, the flush request gets cloned, and blk_insert_cloned_request
> is called.  This cloned request never actually gets issued to the
> q->requst_fn (scsi_request_fn in my case).  So, it may be that there is
> no plug list for this, so the queue isn't goosed?  I'll try to come up
> with a better explanation, or Tejun may just know off the top of his
> head what's going on.
>
> So, the patch works for me, but is very much just an RFC.
>
> Cheers,
> Jeff
>
> Signed-off-by: Jeff Moyer <jmoyer at redhat.com>
>
> diff --git a/block/blk-flush.c b/block/blk-flush.c
> index 491eb30..7aa4736 100644
> --- a/block/blk-flush.c
> +++ b/block/blk-flush.c
> @@ -320,7 +320,7 @@ void blk_insert_flush(struct request *rq)
>                return;
>        }
>
> -       BUG_ON(!rq->bio || rq->bio != rq->biotail);
> +       BUG_ON(rq->bio && rq->bio != rq->biotail);
>
>        /*
>         * If there's data but flush is not necessary, the request can be
> @@ -345,6 +345,12 @@ void blk_insert_flush(struct request *rq)
>        rq->end_io = flush_data_end_io;
>
>        blk_flush_complete_seq(rq, REQ_FSEQ_ACTIONS & ~policy, 0);
> +
> +       /*
> +        * A cloned empty flush needs a queue kick to make progress.
> +        */
> +       if (!rq->bio)
> +               blk_run_queue_async(q);
>  }
Looks the dm request based flush logic is broken.

saved_make_request_fn
  __make_request
    blk_insert_flush
but blk_insert_flush doesn't put the original request to list, instead, the
q->flush_rq is in list.
then
dm_request_fn
  blk_peek_request
    dm_prep_fn
      clone_rq
  map_request
    blk_insert_cloned_request
so q->flush_rq is cloned, and get dispatched. but we can't clone q->flush_rq
and use it to do flush. map_request even could assign a different blockdev to
the cloned request.

Clone q->flush_rq is absolutely wrong. we can clear ogirinal request's flush bit
and set cloned request flush bit, so we can skip the blk_insert_flush logic for
original request.

can you please try this patch. Only compilation tested.

Thanks,
Shaohua
-------------- next part --------------
A non-text attachment was scrubbed...
Name: block-cloned-request-flush.patch
Type: text/x-patch
Size: 2240 bytes
Desc: not available
URL: <http://listman.redhat.com/archives/dm-devel/attachments/20111008/0d6d43af/attachment.bin>