Re: [Qemu-devel] [PATCH 0/2] virtio: revert virtio_queue_set_notification() nesting

[Laszlo Ersek <lersek@redhat.com>]

On 01/12/17 12:46, Stefan Hajnoczi wrote:
> The virtio_queue_set_notification() nesting introduced for AioContext polling
> raised an assertion with virtio-net (even in non-polling mode).  Converting
> virtio-net and virtio-crypto to use virtio_queue_set_notification() in a
> nesting fashion would be invasive and isn't worth it.
> 
> Patch 1 contains the revert to resolve the bug that Doug noticed.
> 
> Patch 2 is a less efficient but safe alternative.
> 
> Stefan Hajnoczi (2):
>   Revert "virtio: turn vq->notification into a nested counter"
>   virtio: disable notifications again after poll succeeded
> 
>  hw/virtio/virtio.c | 21 +++++++++------------
>  1 file changed, 9 insertions(+), 12 deletions(-)
> 

I confirm that with current master
(2ccede18bd24fce5db83fef3674563a1f256717b), my TCG aarch64 guest,
running ArmVirtQemu UEFI firmware, crashes with the following assertion
failure:

qemu-system-aarch64: .../hw/virtio/virtio.c:215:
virtio_queue_set_notification: Assertion `vq->notification_disabled > 0'
failed.

This guest does not use iPXE's UEFI SNP driver for virtio-net, instead
it uses OVMF's own, built-in VirtioNetDxe driver.

With both patches applied, everything works fine. The assertion failure
is gone, and I could ping a public host from the UEFI shell command line.

Tested-by: Laszlo Ersek <lersek@redhat.com>

I'm unsure if my use case covers polling mode, so it might be prudent to
add the tag to patch #1 only. I set the breakpoint that you gave Doug
(using "virsh start --paused" + attaching GDB to the running QEMU
process, before the firmware got any chance to execute), and the
breakpoint (virtio_queue_host_notifier_aio_poll_begin) didn't fire
during the test.

Thanks!
Laszlo