Hi Alexandre,
On Mon, Aug 22, 2022 at 4:03 PM Alexandre Courbot <acourbot@xxxxxxxxxxxx> wrote:
>
> Hi Suwan, apologies for taking so long to come back to this.
>
> On Tue, Aug 2, 2022 at 11:50 PM Kim Suwan <suwan.kim027@xxxxxxxxx> wrote:
> >
> > Hi Alexandre
> >
> > On Tue, Aug 2, 2022 at 11:12 AM Alexandre Courbot <acourbot@xxxxxxxxxxxx> wrote:
> > >
> > > Hi Suwan,
> > >
> > > Thanks for the fast reply!
> > >
> > > On Tue, Aug 2, 2022 at 1:55 AM Kim Suwan <suwan.kim027@xxxxxxxxx> wrote:
> > > >
> > > > Hi Alexandre,
> > > >
> > > > Thanks for reporting the issue.
> > > >
> > > > I think a possible scenario is that request fails at
> > > > virtio_queue_rqs() and it is passed to normal path (virtio_queue_rq).
> > > >
> > > > In this procedure, It is possible that blk_mq_start_request()
> > > > was called twice changing request state from MQ_RQ_IN_FLIGHT to
> > > > MQ_RQ_IN_FLIGHT.
> > >
> > > I have checked whether virtblk_prep_rq_batch() within
> > > virtio_queue_rqs() ever returns 0, and it looks like it never happens.
> > > So as far as I can tell all virtio_queue_rqs() are processed
> > > successfully - but maybe the request can also fail further down the
> > > line? Is there some extra instrumentation I can do to check that?
> > >
> >
> > I'm looking at one more suspicious code.
> > If virtblk_add_req() fails within virtblk_add_req_batch(),
> > virtio_queue_rqs() passes the failed request to the normal path also
> > (virtio_queue_rq). Then, it can call blk_mq_start_request() twice.
> >
> > Because I can't reproduce the issue on my vm, Could you test
> > the below patch?
> > I defer the blk_mq_start_request() call after virtblk_add_req()
> > to ensure that we call blk_mq_start_request() after all the
> > preparations finish.
>
> Your patch seems to solve the problem! I am not seeing the warning
> anymore and the block device looks happy.
Good news! Thanks for the test!
> Let me know if I can do anything else.
Could you test one more patch?
I move blk_mq_start_request(req) before spinlock() to reduce time
holding the lock within virtio_queue_rq().
If it is ok, I will send the patch.
Regards,
Suwan Kim
---
diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
index 30255fcaf181..73a0620a7cff 100644
--- a/drivers/block/virtio_blk.c
+++ b/drivers/block/virtio_blk.c
@@ -322,8 +322,6 @@ static blk_status_t virtblk_prep_rq(struct
blk_mq_hw_ctx *hctx,
if (unlikely(status))
return status;
- blk_mq_start_request(req);
-
vbr->sg_table.nents = virtblk_map_data(hctx, req, vbr);
if (unlikely(vbr->sg_table.nents < 0)) {
virtblk_cleanup_cmd(req);
@@ -349,6 +347,8 @@ static blk_status_t virtio_queue_rq(struct
blk_mq_hw_ctx *hctx,
if (unlikely(status))
return status;
+ blk_mq_start_request(req);
+
spin_lock_irqsave(&vblk->vqs[qid].lock, flags);
err = virtblk_add_req(vblk->vqs[qid].vq, vbr);
if (err) {
@@ -409,6 +409,8 @@ static bool virtblk_add_req_batch(struct virtio_blk_vq *vq,
virtblk_unmap_data(req, vbr);
virtblk_cleanup_cmd(req);
rq_list_add(requeue_list, req);
+ } else {
+ blk_mq_start_request(req);
}
}
