Re: [PATCH] virtio_balloon: clear modern features under legacy

[Alexander Duyck]

On Sun, Jul 12, 2020 at 8:10 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
>
> On Fri, Jul 10, 2020 at 09:13:41AM -0700, Alexander Duyck wrote:
> > On Fri, Jul 10, 2020 at 4:31 AM Michael S. Tsirkin <mst@xxxxxxxxxx> wrote:
> > >
> > > Page reporting features were never supported by legacy hypervisors.
> > > Supporting them poses a problem: should we use native endian-ness (like
> > > current code assumes)? Or little endian-ness like the virtio spec says?
> > > Rather than try to figure out, and since results of
> > > incorrect endian-ness are dire, let's just block this configuration.
> > >
> > > Cc: stable@xxxxxxxxxxxxxxx
> > > Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>
> >
> > So I am not sure about the patch description. In the case of page
> > poison and free page reporting I don't think we are defining anything
> > that doesn't already have a definition of how to use in legacy.
> > Specifically the virtio_balloon_config is already defined as having
> > all fields as little endian in legacy mode, and there is a definition
> > for all of the fields in a virtqueue and how they behave in legacy
> > mode.
> >
> > As far as I can see the only item that may be an issue is the command
> > ID being supplied via the virtqueue for free page hinting, which
> > appears to be in native endian-ness. Otherwise it would have fallen
> > into the same category since it is making use of virtio_balloon_config
> > and a virtqueue for supplying the page location and length.
>
>
>
> So as you point out correctly balloon spec says all fields are little
> endian.  Fair enough.
> Problem is when virtio 1 is not negotiated, then this is not what the
> driver assumes for any except a handlful of fields.
>
> But yes it mostly works out.
>
> For example:
>
>
> static void update_balloon_size(struct virtio_balloon *vb)
> {
>         u32 actual = vb->num_pages;
>
>         /* Legacy balloon config space is LE, unlike all other devices. */
>         if (!virtio_has_feature(vb->vdev, VIRTIO_F_VERSION_1))
>                 actual = (__force u32)cpu_to_le32(actual);
>
>         virtio_cwrite(vb->vdev, struct virtio_balloon_config, actual,
>                       &actual);
> }
>
>
> this is LE even without VIRTIO_F_VERSION_1, so matches spec.
>
>                 /* Start with poison val of 0 representing general init */
>                 __u32 poison_val = 0;
>
>                 /*
>                  * Let the hypervisor know that we are expecting a
>                  * specific value to be written back in balloon pages.
>                  */
>                 if (!want_init_on_free())
>                         memset(&poison_val, PAGE_POISON, sizeof(poison_val));
>
>                 virtio_cwrite(vb->vdev, struct virtio_balloon_config,
>                               poison_val, &poison_val);
>
>
> actually this writes a native endian-ness value. All bytes happen to be
> the same though, and host only cares about 0 or non 0 ATM.

So we are safe assuming it is a repeating value, but for correctness
maybe we should make certain to cast this as a le32 value. I can
submit a patch to do that.

> As you say correctly the command id is actually assumed native endian:
>
>
> static u32 virtio_balloon_cmd_id_received(struct virtio_balloon *vb)
> {
>         if (test_and_clear_bit(VIRTIO_BALLOON_CONFIG_READ_CMD_ID,
>                                &vb->config_read_bitmap))
>                 virtio_cread(vb->vdev, struct virtio_balloon_config,
>                              free_page_hint_cmd_id,
>                              &vb->cmd_id_received_cache);
>
>         return vb->cmd_id_received_cache;
> }
>
>
> So guest assumes native, host assumes LE.

This wasn't even the one I was talking about, but now that you point
it out this is definately bug. The command ID I was talking about was
the one being passed via the descriptor ring. That one I believe is
native on both sides.

>
>
>
> > > ---
> > >  drivers/virtio/virtio_balloon.c | 9 +++++++++
> > >  1 file changed, 9 insertions(+)
> > >
> > > diff --git a/drivers/virtio/virtio_balloon.c b/drivers/virtio/virtio_balloon.c
> > > index 5d4b891bf84f..b9bc03345157 100644
> > > --- a/drivers/virtio/virtio_balloon.c
> > > +++ b/drivers/virtio/virtio_balloon.c
> > > @@ -1107,6 +1107,15 @@ static int virtballoon_restore(struct virtio_device *vdev)
> > >
> > >  static int virtballoon_validate(struct virtio_device *vdev)
> > >  {
> > > +       /*
> > > +        * Legacy devices never specified how modern features should behave.
> > > +        * E.g. which endian-ness to use? Better not to assume anything.
> > > +        */
> > > +       if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
> > > +               __virtio_clear_bit(vdev, VIRTIO_BALLOON_F_FREE_PAGE_HINT);
> > > +               __virtio_clear_bit(vdev, VIRTIO_BALLOON_F_PAGE_POISON);
> > > +               __virtio_clear_bit(vdev, VIRTIO_BALLOON_F_REPORTING);
> > > +       }
> > >         /*
> > >          * Inform the hypervisor that our pages are poisoned or
> > >          * initialized. If we cannot do that then we should disable
> >
> > The patch content itself I am fine with since odds are nobody would
> > expect to use these features with a legacy device.
> >
> > Acked-by: Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx>
>
> Hmm so now you pointed out it's just cmd id, maybe I should just fix it
> instead? what do you say?

So the config issues are bugs, but I don't think you saw the one I was
talking about. In the function send_cmd_id_start the cmd_id_active
value which is initialized as a virtio32 is added as a sg entry and
then sent as an outbuf to the device. I'm assuming virtio32 is a host
native byte ordering.