Re: [PATCH] vdpa: set old virtio status at cvq isolation probing end

[Eugenio Perez Martin]

On Mon, Jul 31, 2023 at 8:36 AM Jason Wang <jasowang@redhat.com> wrote:
>
> On Wed, Jul 26, 2023 at 2:27 PM Eugenio Perez Martin
> <eperezma@redhat.com> wrote:
> >
> > On Wed, Jul 26, 2023 at 4:07 AM Jason Wang <jasowang@redhat.com> wrote:
> > >
> > > On Wed, Jul 26, 2023 at 2:21 AM Eugenio Pérez <eperezma@redhat.com> wrote:
> > > >
> > > > The device already has a virtio status set by vhost_vdpa_init by the
> > > > time vhost_vdpa_probe_cvq_isolation is called. vhost_vdpa_init set
> > > > S_ACKNOWLEDGE and S_DRIVER, so it is invalid to just reset it.
> > > >
> > > > It is invalid to start the device after it, but all devices seems to be
> > > > fine with it.  Fixing qemu so it follows virtio start procedure.
> > > >
> > > > Fixes: 152128d64697 ("vdpa: move CVQ isolation check to 
> > > > net_init_vhost_vdpa")
> > > > Reported-by: Dragos Tatulea <dtatulea@nvidia.com>
> > > > Signed-off-by: Eugenio Pérez <eperezma@redhat.com>
> > > > ---
> > > >  net/vhost-vdpa.c | 2 ++
> > > >  1 file changed, 2 insertions(+)
> > > >
> > > > diff --git a/net/vhost-vdpa.c b/net/vhost-vdpa.c
> > > > index 9795306742..d7e2b714b4 100644
> > > > --- a/net/vhost-vdpa.c
> > > > +++ b/net/vhost-vdpa.c
> > > > @@ -1333,6 +1333,8 @@ static int vhost_vdpa_probe_cvq_isolation(int 
> > > > device_fd, uint64_t features,
> > > >  out:
> > > >      status = 0;
> > > >      ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
> > > > +    status = VIRTIO_CONFIG_S_ACKNOWLEDGE | VIRTIO_CONFIG_S_DRIVER;
> > > > +    ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
> > >
> > > So if we fail after FEATURES_OK, this basically clears that bit. Spec
> > > doesn't say it can or not, I wonder if a reset is better?
> > >
> >
> > I don't follow this, the reset is just above the added code, isn't it?
>
> I meant for error path:
>
> E.g:
>     uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE |
>                      VIRTIO_CONFIG_S_DRIVER |
>                      VIRTIO_CONFIG_S_FEATURES_OK;
> ...
>     r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
> ....
>         if (cvq_group != -ENOTSUP) {
>             r = cvq_group;
>             goto out;
>         }
>
> out:
>     status = VIRTIO_CONFIG_S_ACKNOWLEDGE | VIRTIO_CONFIG_S_DRIVER;
>     ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
>
> We're basically clearing FEATURES_OK?
>

Yes, it is the state that previous functions (vhost_vdpa_init) set. We
need to leave it that way, either if the backend supports cvq
isolation or not, or in the case of an error. Not doing that way makes
vhost_dev_start (and vhost_vdpa_set_features) set the features before
setting VIRTIO_CONFIG_S_ACKNOWLEDGE | VIRTIO_CONFIG_S_DRIVER.
Otherwise, the guest can (and do) access to config space before
_S_ACKNOWLEDGE | _S_DRIVER.


> >
> > > Btw, spec requires a read of status after setting FEATURES_OK, this
> > > seems to be missed in the current code.
> > >
> >
> > I'm ok with that, but this patch does not touch that part.
> >
> > To fix this properly we should:
> > - Expose vhost_vdpa_set_dev_features_fd as we did in previous versions
> > of the series that added vhost_vdpa_probe_cvq_isolation [1].
> > - Get status after vhost_vdpa_add_status, so both vhost start code and
> > this follows the standard properly.
> >
> > Is it ok to do these on top of this patch?
>
> Fine.
>
> Thanks
>
> >
> > Thanks!
> >
> > [1] 
> > https://lore.kernel.org/qemu-devel/20230509154435.1410162-4-eperezma@redhat.com/
> >
> >
> > > Thanks
> > >
> > > >      return r;
> > > >  }
> > > >
> > > > --
> > > > 2.39.3
> > > >
> > >
> >
>