Linux Virtualization - Nov 2014 - [PATCH V3 1/2] virtio: introduce methods of sanitizing device features

On Mon, 17 Nov 2014 11:37:01 +0200
"Michael S. Tsirkin" <mst at redhat.com> wrote:
> On Mon, Nov 17, 2014 at 05:17:17PM +0800, Jason Wang wrote:
> > Buggy host may advertised buggy host features (a usual case is that
host
> > advertise a feature whose dependencies were missed). In this case,
driver
> > should detect and disable the buggy features by itself.
> > 
> > This patch introduces driver specific sanitize_features() method which
is
> > called just before features finalizing to detect and disable buggy
features
> > advertised by host.
> > 
> > Virtio-net will be the first user.
> > 
> > Cc: Rusty Russell <rusty at rustcorp.com.au>
> > Cc: Michael S. Tsirkin <mst at redhat.com>
> > Cc: Cornelia Huck <cornelia.huck at de.ibm.com>
> > Cc: Wanlong Gao <gaowanlong at cn.fujitsu.com>
> > Signed-off-by: Jason Wang <jasowang at redhat.com>
> 
> Hmm this conflicts with virtio 1.0 work: we drop
> features as bitmap there.
But that's an implementation detail, no? We'll still need a way for the
driver to sanitize features, and I think this interface works just fine.

On Mon, Nov 17, 2014 at 10:44:30AM +0100, Cornelia Huck
wrote:
> On Mon, 17 Nov 2014 11:37:01 +0200
> "Michael S. Tsirkin" <mst at redhat.com> wrote:
> 
> > On Mon, Nov 17, 2014 at 05:17:17PM +0800, Jason Wang wrote:
> > > Buggy host may advertised buggy host features (a usual case is
that host
> > > advertise a feature whose dependencies were missed). In this
case, driver
> > > should detect and disable the buggy features by itself.
> > > 
> > > This patch introduces driver specific sanitize_features() method
which is
> > > called just before features finalizing to detect and disable
buggy features
> > > advertised by host.
> > > 
> > > Virtio-net will be the first user.
> > > 
> > > Cc: Rusty Russell <rusty at rustcorp.com.au>
> > > Cc: Michael S. Tsirkin <mst at redhat.com>
> > > Cc: Cornelia Huck <cornelia.huck at de.ibm.com>
> > > Cc: Wanlong Gao <gaowanlong at cn.fujitsu.com>
> > > Signed-off-by: Jason Wang <jasowang at redhat.com>
> > 
> > Hmm this conflicts with virtio 1.0 work: we drop
> > features as bitmap there.
> 
> But that's an implementation detail, no? We'll still need a way for
the
> driver to sanitize features, and I think this interface works just fine.
Now that you mention it, I don't think we do.

The spec is quite explicit that devices must not expose invalid
combinations of features.

Admittedly, BUG_ON isn't very friendly to hypervisors.

But e.g. failing probe seems better than trying to work around
hypervisor bugs - otherwise we'll be stuck maintaining compatibility
with hypervisors forever.

-- 
MST

On Mon, 17 Nov 2014 12:11:39 +0200
"Michael S. Tsirkin" <mst at redhat.com> wrote:
> On Mon, Nov 17, 2014 at 10:44:30AM +0100, Cornelia Huck wrote:
> > On Mon, 17 Nov 2014 11:37:01 +0200
> > "Michael S. Tsirkin" <mst at redhat.com> wrote:
> > 
> > > On Mon, Nov 17, 2014 at 05:17:17PM +0800, Jason Wang wrote:
> > > > Buggy host may advertised buggy host features (a usual case
is that host
> > > > advertise a feature whose dependencies were missed). In this
case, driver
> > > > should detect and disable the buggy features by itself.
> > > > 
> > > > This patch introduces driver specific sanitize_features()
method which is
> > > > called just before features finalizing to detect and disable
buggy features
> > > > advertised by host.
> > > > 
> > > > Virtio-net will be the first user.
> > > > 
> > > > Cc: Rusty Russell <rusty at rustcorp.com.au>
> > > > Cc: Michael S. Tsirkin <mst at redhat.com>
> > > > Cc: Cornelia Huck <cornelia.huck at de.ibm.com>
> > > > Cc: Wanlong Gao <gaowanlong at cn.fujitsu.com>
> > > > Signed-off-by: Jason Wang <jasowang at redhat.com>
> > > 
> > > Hmm this conflicts with virtio 1.0 work: we drop
> > > features as bitmap there.
> > 
> > But that's an implementation detail, no? We'll still need a
way for the
> > driver to sanitize features, and I think this interface works just
fine.
> 
> Now that you mention it, I don't think we do.
> 
> The spec is quite explicit that devices must not expose invalid
> combinations of features.
Unfortunately, this does not ensure that there won't be buggy
hypervisors out there, just as there's buggy hardware floating around.
> 
> Admittedly, BUG_ON isn't very friendly to hypervisors.
> 
> But e.g. failing probe seems better than trying to work around
> hypervisor bugs - otherwise we'll be stuck maintaining compatibility
> with hypervisors forever.
Good point. Failing probe is still much better than hitting BUG_ONs.

We'll still need a driver callback, though, that can return an error on
bogus feature bit combinations.

On 11/17/2014 06:11 PM, Michael S. Tsirkin wrote:
> On Mon, Nov 17, 2014 at 10:44:30AM +0100, Cornelia Huck wrote:
>> On Mon, 17 Nov 2014 11:37:01 +0200
>> "Michael S. Tsirkin" <mst at redhat.com> wrote:
>>
>>> On Mon, Nov 17, 2014 at 05:17:17PM +0800, Jason Wang wrote:
>>>> Buggy host may advertised buggy host features (a usual case is
that host
>>>> advertise a feature whose dependencies were missed). In this
case, driver
>>>> should detect and disable the buggy features by itself.
>>>>
>>>> This patch introduces driver specific sanitize_features()
method which is
>>>> called just before features finalizing to detect and disable
buggy features
>>>> advertised by host.
>>>>
>>>> Virtio-net will be the first user.
>>>>
>>>> Cc: Rusty Russell <rusty at rustcorp.com.au>
>>>> Cc: Michael S. Tsirkin <mst at redhat.com>
>>>> Cc: Cornelia Huck <cornelia.huck at de.ibm.com>
>>>> Cc: Wanlong Gao <gaowanlong at cn.fujitsu.com>
>>>> Signed-off-by: Jason Wang <jasowang at redhat.com>
>>> Hmm this conflicts with virtio 1.0 work: we drop
>>> features as bitmap there.
>> But that's an implementation detail, no? We'll still need a way
for the
>> driver to sanitize features, and I think this interface works just
fine.
> Now that you mention it, I don't think we do.
>
> The spec is quite explicit that devices must not expose invalid
> combinations of features.
>
> Admittedly, BUG_ON isn't very friendly to hypervisors.
>
> But e.g. failing probe seems better than trying to work around
> hypervisor bugs - otherwise we'll be stuck maintaining compatibility
> with hypervisors forever.
>
I'm ok with failing the probe.

But it won't cost big effort to workaround only features dependencies
issue. I don't see how this block any further features implementation.
Looking at virtio-net, it also depends on network core to fix NETIF_F_*
dependencies.

There seems no way to get rid of maintaining compatibility, e.g  the
workarounds for the buggy hypervisor without VIRTIO_F_ANY_LAYOUT support.