Re: bio linked list corruption.

[Jens Axboe]

On 10/26/2016 04:51 PM, Linus Torvalds wrote:
> On Wed, Oct 26, 2016 at 3:40 PM, Dave Jones <davej@xxxxxxxxxxxxxxxxx> wrote:
> > I gave it a shot too for shits & giggles.
> > This falls out during boot.
> >
> > [    9.278420] WARNING: CPU: 0 PID: 1 at block/blk-mq.c:1181 blk_sq_make_request+0x465/0x4a0
>
> Hmm. That's the
>
>     WARN_ON_ONCE(rq->mq_ctx != ctx);
>
> that I added to blk_mq_merge_queue_io(), and I really think that
> warning is valid, and the fact that it triggers shows that something
> is wrong with locking.
>
> We just did a
>
>                 spin_lock(&ctx->lock);
>
> and that lock is *supposed* to protect the __blk_mq_insert_request(),
> but that uses rq->mq_ctx.
>
> So if rq->mq_ctx != ctx, then we're locking the wrong context.
>
> Jens - please explain to me why I'm wrong.
>
> Or maybe I actually might have found the problem? In which case please
> send me a patch that fixes it ;)

I think you're pretty close, the two should not be different and I don't
immediately see how. I'll run some testing here, should be easier with
this knowledge.

> Dave: it might be a good idea to split that "WARN_ON_ONCE()" in
> blk_mq_merge_queue_io() into two, since right now it can trigger both
> for the
>
>                 blk_mq_bio_to_request(rq, bio);
>
> path _and_ for the
>
>                 if (!blk_mq_attempt_merge(q, ctx, bio)) {
>                         blk_mq_bio_to_request(rq, bio);
>                         goto insert_rq;

And just in case I can't trigger, would be interesting to add a call to
blk_rq_dump_flags() as well, in case this is some special request.


--
Jens Axboe