Re: Hang in 9p/virtio

[Cornelia Huck]

On Tue, 2 Aug 2016 15:35:34 +0200
Vegard Nossum <vegard.nossum@xxxxxxxxxx> wrote:

> On 08/02/2016 11:13 AM, Vegard Nossum wrote:
> > On 08/02/2016 11:03 AM, Cornelia Huck wrote:
> >> On Sat, 30 Jul 2016 23:42:18 +0200
> >> Vegard Nossum <vegard.nossum@xxxxxxxxxx> wrote:
> >>
> >>> Hi,
> >>>
> >>> With fault injection triggering an allocation failure for the
> >>> alloc_indirect() call in virtqueue_add() I'm seeing a hang in
> >>> p9_virtio_zc_request() -- it seems to be waiting here indefinitely
> >>> (i.e. at least 120 seconds):
> >>>
> > [...]
> >
> >> What happens is that the code falls back to direct virtio addressing
> >> (after indirect addressing failed) - and this should work.
> >>
> >> I'm more inclined to suspect a qemu instead of a kernel bug, as your
> >> qemu version is quite old and there have been fixes in the virtio
> >> buffer handling and virtio-9p in the meantime. (I'm suspecting
> >> "virtio-9p: fix any_layout".)
> >>
> >> Could you retry with a more recent qemu (at least version 2.4)?
> >
> > I think maybe the version number in the stack trace is a bit misleading,
> > this is the full/actual version:
> >
> > $ kvm --version
> > QEMU emulator version 2.5.0 (Debian 1:2.5+dfsg-5ubuntu10.1), Copyright
> > (c) 2003-2008 Fabrice Bellard
> >
> > I'll still try to get qemu from git and see if it makes a difference.
> > Thanks,
> 
> I still seem to get it:
> 
> $ qemu-system-x86_64 --version
> QEMU emulator version 2.6.91 (v2.7.0-rc1-2-gcc0100f-dirty), Copyright 
> (c) 2003-2008 Fabrice Bellard

:(

Sorry, no good immediate idea.

One thing would be to check whether you get notified by qemu after the
request was queued (i.e., whether vring_interrupt() ever gets called
with 9p's req_done() after the alloc failure was injected). This would
help to suggest whether to continue debugging here or in qemu.

I still think the root of this error is some failure of the virtio 9p
code to deal with non-indirect buffers, either in the driver or in qemu.