[Qemu-devel] Re: [PATCH +stable] block: don't attempt to merge overlapping requests

[Christoph Hellwig]

On Wed, May 19, 2010 at 10:23:44AM +0100, Stefan Hajnoczi wrote:
> On Wed, May 19, 2010 at 10:06 AM, Avi Kivity <address@hidden> wrote:
> >> In the cache=writeback case the virtio-blk guest driver does:
> >>
> >> blk_queue_ordered(q, QUEUE_ORDERED_DRAIN_FLUSH, ...)
> >>
> >
> > I don't follow. ?What's the implication?
> 
> I was wondering whether the queue is incorrectly set to a mode where
> overlapping write requests aren't being ordered.  Anyway, Christoph
> says overlapping write requests can be issued so my theory is broken.

They can happen, but won't during usual pagecache based writeback.  So
this should not happen for the pagecache based mke2fs workload.  It
could happen for a workload with mkfs that uses O_DIRECT.  And we
need to handle it either way.

And btw, our barrier handling for devices using multiwrite (aka virtio)
is utterly busted right now, patch will follow ASAP.