Messages in this thread | | | Date | Tue, 19 Nov 2013 20:12:48 -0700 | From | Jens Axboe <> | Subject | Re: [Regression x2, 3.13-git] virtio block mq hang, iostat busted on virtio devices |
| |
On Wed, Nov 20 2013, Dave Chinner wrote: > On Tue, Nov 19, 2013 at 07:02:30PM -0700, Jens Axboe wrote: > > On Tue, Nov 19 2013, Jens Axboe wrote: > > > > Looks like a race condition, below works for me, please try. > > > > > > > > > > > > Subject: virtio_blk: fix race condition > > > > > > > > virtqueue_kick() isn't multi-thread safe. > > > > > > > > Signed-off-by: Shaohua Li <shli@fusionio.com> > > > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > > index 588479d..f353959 100644 > > > > --- a/drivers/block/virtio_blk.c > > > > +++ b/drivers/block/virtio_blk.c > > > > @@ -204,10 +204,11 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *req) > > > > virtqueue_kick(vblk->vq); > > > > return BLK_MQ_RQ_QUEUE_BUSY; > > > > } > > > > - spin_unlock_irqrestore(&vblk->vq_lock, flags); > > > > > > > > if (last) > > > > virtqueue_kick(vblk->vq); > > > > + spin_unlock_irqrestore(&vblk->vq_lock, flags); > > > > + > > > > return BLK_MQ_RQ_QUEUE_OK; > > > > } > > > > > > Just stumbled on that too. You need one more, btw, for the sg failure > > > case: > > > > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > index 588479d58f52..6a680d4de7f1 100644 > > > --- a/drivers/block/virtio_blk.c > > > +++ b/drivers/block/virtio_blk.c > > > @@ -199,15 +199,16 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx, struct request *req) > > > > > > spin_lock_irqsave(&vblk->vq_lock, flags); > > > if (__virtblk_add_req(vblk->vq, vbr, vbr->sg, num) < 0) { > > > + virtqueue_kick(vblk->vq); > > > spin_unlock_irqrestore(&vblk->vq_lock, flags); > > > blk_mq_stop_hw_queue(hctx); > > > - virtqueue_kick(vblk->vq); > > > return BLK_MQ_RQ_QUEUE_BUSY; > > > } > > > - spin_unlock_irqrestore(&vblk->vq_lock, flags); > > > > > > if (last) > > > virtqueue_kick(vblk->vq); > > > + > > > + spin_unlock_irqrestore(&vblk->vq_lock, flags); > > > return BLK_MQ_RQ_QUEUE_OK; > > > } > > > > Tested successfully here too. > > Ah, so it is exactly the problem I suggested it might be. ;)
It isn't actually, it's not a race between the queue conditions, the stopping/starting etc or inside/outside lock state checking. It's a "simple" race between the virtqueue operations. It is a race, however, but I think that one was given :-)
> > Dave, please give it a go, looks like this > > should fix it up for you. Committed here: > > > > http://git.kernel.dk/?p=linux-block.git;a=commit;h=f02b9ac35a47dff745c7637fbc095f01cc03646e > > Testing it now. might take a little while to confirm given it had > taken a few iterations of xfstests before I tripped over it...
I feel pretty confident in it, fwiw. My test case was boiled down to trigger it in seconds, and it survived a lengthy run afterwards.
-- Jens Axboe
| |