On Wed, Jul 10, 2024 at 12:58:11PM -0700, Daniel Verkamp wrote:
> On Wed, Jul 10, 2024 at 11:39 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> >
> > On Wed, Jul 10, 2024 at 11:12:34AM -0700, Daniel Verkamp wrote:
> > > On Wed, Jul 10, 2024 at 4:43 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > > >
> > > > virtio balloon communicates to the core that in some
> > > > configurations vq #s are non-contiguous by setting name
> > > > pointer to NULL.
> > > >
> > > > Unfortunately, core then turned around and just made them
> > > > contiguous again. Result is that driver is out of spec.
> > >
> > > Thanks for fixing this - I think the overall approach of the patch looks good.
> > >
> > > > Implement what the API was supposed to do
> > > > in the 1st place. Compatibility with buggy hypervisors
> > > > is handled inside virtio-balloon, which is the only driver
> > > > making use of this facility, so far.
> > >
> > > In addition to virtio-balloon, I believe the same problem also affects
> > > the virtio-fs device, since queue 1 is only supposed to be present if
> > > VIRTIO_FS_F_NOTIFICATION is negotiated, and the request queues are
> > > meant to be queue indexes 2 and up. From a look at the Linux driver
> > > (virtio_fs.c), it appears like it never acks VIRTIO_FS_F_NOTIFICATION
> > > and assumes that request queues start at index 1 rather than 2, which
> > > looks out of spec to me, but the current device implementations (that
> > > I am aware of, anyway) are also broken in the same way, so it ends up
> > > working today. Queue numbering in a spec-compliant device and the
> > > current Linux driver would mismatch; what the driver considers to be
> > > the first request queue (index 1) would be ignored by the device since
> > > queue index 1 has no function if F_NOTIFICATION isn't negotiated.
> >
> >
> > Oh, thanks a lot for pointing this out!
> >
> > I see so this patch is no good as is, we need to add a workaround for
> > virtio-fs first.
> >
> > QEMU workaround is simple - just add an extra queue. But I did not
> > reasearch how this would interact with vhost-user.
> >
> > From driver POV, I guess we could just ignore queue # 1 - would that be
> > ok or does it have performance implications?
>
> As a driver workaround for non-compliant devices, I think ignoring the
> first request queue would be a reasonable approach if the device's
> config advertises num_request_queues > 1. Unfortunately, both
> virtiofsd and crosvm's virtio-fs device have hard-coded
> num_request_queues =1, so this won't help with those existing devices.
Do they care what the vq # is though?
We could do some magic to translate VQ #s in qemu.
> Maybe there are other devices that we would need to consider as well;
> commit 529395d2ae64 ("virtio-fs: add multi-queue support") quotes
> benchmarks that seem to be from a different virtio-fs implementation
> that does support multiple request queues, so the workaround could
> possibly be used there.
>
> > Or do what I did for balloon here: try with spec compliant #s first,
> > if that fails then assume it's the spec issue and shift by 1.
>
> If there is a way to "guess and check" without breaking spec-compliant
> devices, that sounds reasonable too; however, I'm not sure how this
> would work out in practice: an existing non-compliant device may fail
> to start if the driver tries to enable queue index 2 when it only
> supports one request queue,
You don't try to enable queue - driver starts by checking queue size.
The way my patch works is that it assumes a non existing queue has
size 0 if not available.
This was actually a documented way to check for PCI and MMIO:
Read the virtqueue size from queue_size. This controls how big the virtqueue is (see 2.6 Virtqueues).
If this field is 0, the virtqueue does not exist.
MMIO:
If the returned value is zero (0x0) the queue is not available.
unfortunately not for CCW, but I guess CCW implementations outside
of QEMU are uncommon enough that we can assume it's the same?
To me the above is also a big hint that drivers are allowed to
query size for queues that do not exist.
> and a spec-compliant device would probably
> balk if the driver tries to enable queue 1 but does not negotiate
> VIRTIO_FS_F_NOTIFICATION. If there's a way to reset and retry the
> whole virtio device initialization process if a device fails like
> this, then maybe it's feasible. (Or can the driver tweak the virtqueue
> configuration and try to set DRIVER_OK repeatedly until it works? It's
> not clear to me if this is allowed by the spec, or what device
> implementations actually do in practice in this scenario.)
>
> Thanks,
> -- Daniel
My patch starts with a spec compliant behaviour. If that fails,
try non-compliant one as a fallback.
--
MST
