Re: [PATCH] virtio_net: Revert "virtio_net: set the default max ring size by find_vqs()"

From: Michael S. Tsirkin
Date: Tue Aug 16 2022 - 01:22:16 EST


On Mon, Aug 15, 2022 at 02:46:04PM -0700, Andres Freund wrote:
> Hi,
>
> On 2022-08-15 17:39:08 -0400, Michael S. Tsirkin wrote:
> > On Mon, Aug 15, 2022 at 02:28:39PM -0700, Andres Freund wrote:
> > > On 2022-08-15 17:04:10 -0400, Michael S. Tsirkin wrote:
> > > > So virtio has a queue_size register. When read, it will give you
> > > > originally the maximum queue size. Normally we just read it and
> > > > use it as queue size.
> > > >
> > > > However, when queue memory allocation fails, and unconditionally with a
> > > > network device with the problematic patch, driver is asking the
> > > > hypervisor to make the ring smaller by writing a smaller value into this
> > > > register.
> > > >
> > > > I suspect that what happens is hypervisor still uses the original value
> > > > somewhere.
> > >
> > > It looks more like the host is never told about the changed size for legacy
> > > devices...
> > >
> > > Indeed, adding a vp_legacy_set_queue_size() & call to it to setup_vq(), makes
> > > 5.19 + restricting queue sizes to 1024 boot again.
> >
> > Interesting, the register is RO in the legacy interface.
> > And to be frank I can't find where is vp_legacy_set_queue_size
> > even implemented. It's midnight here too ...
>
> Yea, I meant that added both vp_legacy_set_queue_size() and a call to it. I
> was just quickly experimenting around.

interesting that it's writeable on GCP. It's RO on QEMU.

>
> > Yes I figured this out too. And I was able to reproduce on qemu now.
>
> Cool.
>
>
> > I'm posting a new patchset reverting all the handing of resize
> > restrictions, I think we should rethink it for the next release.
>
> Makes sense.
>
> Greetings,
>
> Andres Freund