Re: [Qemu-devel] [PATCH v2 00/15] SCSI s/g + SCSI migration + virtio-scsi

[Hu Tao]

On Tue, Jan 31, 2012 at 01:36:04PM +0100, Paolo Bonzini wrote:
> On 01/30/2012 10:33 AM, Hu Tao wrote:
> >>>  I cannot reproduce this with a 100G qcow2 image (created with
> >>>  "qemu-img create scsi.qcow2 100G" just before launching the host),
> >>>  with a partition starting at sector 2048 and extending to the end of
> >>>  the disk.  mkfs (ext4) takes less than 1 minute and extends the
> >>>  qcow2 file to a little less than 2 gigs.
> >
> >I've re-tested today, with host kernel 2.6.35.6-45.fc14.x86_64,
> >2.6.32-71.el6.x86_64, 3.1.0 and 3.3.0-rc1+, qemu version and guest
> >configuration remain the same, including guest kernel. It apears
> >that the problem appears only when the host kernel version is 3.1.0.

Sorry, forget about the kernel versions. With very heavy IO load, it's
easy to reproduce the problem. In my case, 300 disks were attached to a
virtio-scsi HBA, and 300 dd processes were launched at the same time,
each for a disk.

tests also done for raw disk image, but no such problem.

> 
> There are two parts in the bug:
> 
> 1) I/O cancellation doesn't actually cancel anything on qcow2 (and
> maybe others) since the coroutine transition.  This makes exception
> handling basically unusable in the virtio-scsi guest driver.  I have
> been working on this but I don't have patches ready for testing yet.
> 
> 2) Something unknown in the kernel is causing writes to take a
> longer time than usual.  This appears to be a problem only with
> 3.1.0 guest kernels.
> 
> I think we can wait until (1) causes problems elsewhere.  Thanks for
> testing!
> 
> Paolo

-- 
Thanks,
Hu Tao