Re: [RFC PATCH 1/1] virtio: write back features before verify

[Cornelia Huck <cohuck@redhat.com>]

[cc:qemu-devel]

On Sat, Oct 02 2021, "Michael S. Tsirkin" <mst@redhat.com> wrote:

> On Fri, Oct 01, 2021 at 09:21:25AM +0200, Halil Pasic wrote:
>> On Thu, 30 Sep 2021 07:12:21 -0400
>> "Michael S. Tsirkin" <mst@redhat.com> wrote:
>> 
>> > On Thu, Sep 30, 2021 at 03:20:49AM +0200, Halil Pasic wrote:
>> > > This patch fixes a regression introduced by commit 82e89ea077b9
>> > > ("virtio-blk: Add validation for block size in config space") and
>> > > enables similar checks in verify() on big endian platforms.
>> > > 
>> > > The problem with checking multi-byte config fields in the verify
>> > > callback, on big endian platforms, and with a possibly transitional
>> > > device is the following. The verify() callback is called between
>> > > config->get_features() and virtio_finalize_features(). That we have a
>> > > device that offered F_VERSION_1 then we have the following options
>> > > either the device is transitional, and then it has to present the legacy
>> > > interface, i.e. a big endian config space until F_VERSION_1 is
>> > > negotiated, or we have a non-transitional device, which makes
>> > > F_VERSION_1 mandatory, and only implements the non-legacy interface and
>> > > thus presents a little endian config space. Because at this point we
>> > > can't know if the device is transitional or non-transitional, we can't
>> > > know do we need to byte swap or not.  
>> > 
>> > Hmm which transport does this refer to?
>> 
>> It is the same with virtio-ccw and virtio-pci. I see the same problem
>> with both on s390x. I didn't try with virtio-blk-pci-non-transitional
>> yet (have to figure out how to do that with libvirt) for pci I used
>> virtio-blk-pci.
>> 
>> > Distinguishing between legacy and modern drivers is transport
>> > specific.  PCI presents
>> > legacy and modern at separate addresses so distinguishing
>> > between these two should be no trouble.
>> 
>> You mean the device id? Yes that is bolted down in the spec, but
>> currently we don't exploit that information. Furthermore there
>> is a fat chance that with QEMU even the allegedly non-transitional
>> devices only present a little endian config space after VERSION_1
>> was negotiated. Namely get_config for virtio-blk is implemented in
>> virtio_blk_update_config() which does virtio_stl_p(vdev,
>> &blkcfg.blk_size, blk_size) and in there we don't care
>> about transitional or not:
>> 
>> static inline bool virtio_access_is_big_endian(VirtIODevice *vdev)
>> {
>> #if defined(LEGACY_VIRTIO_IS_BIENDIAN)
>>     return virtio_is_big_endian(vdev);
>> #elif defined(TARGET_WORDS_BIGENDIAN)
>>     if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
>>         /* Devices conforming to VIRTIO 1.0 or later are always LE. */
>>         return false;
>>     }
>>     return true;
>> #else
>>     return false;
>> #endif
>> }
>> 
>
> ok so that's a QEMU bug. Any virtio 1.0 and up
> compatible device must use LE.
> It can also present a legacy config space where the
> endian depends on the guest.

So, how is the virtio core supposed to determine this? A
transport-specific callback?

>
>> > Channel i/o has versioning so same thing?
>> >
>> 
>> Don't think so. Both a transitional and a non-transitional device
>> would have to accept revisions higher than 0 if the driver tried to
>> negotiate those (and we do in our case).
>
> Yes, the modern driver does. And that one is known to be LE.
> legacy driver doesn't.
>
>> > > The virtio spec explicitly states that the driver MAY read config
>> > > between reading and writing the features so saying that first accessing
>> > > the config before feature negotiation is done is not an option. The
>> > > specification ain't clear about setting the features multiple times
>> > > before FEATURES_OK, so I guess that should be fine.
>> > > 
>> > > I don't consider this patch super clean, but frankly I don't think we
>> > > have a ton of options. Another option that may or man not be cleaner,
>> > > but is also IMHO much uglier is to figure out whether the device is
>> > > transitional by rejecting _F_VERSION_1, then resetting it and proceeding
>> > > according tho what we have figured out, hoping that the characteristics
>> > > of the device didn't change.  
>> > 
>> > I am confused here. So is the problem at the device or at the driver level?
>> 
>> We have a driver regression. Since the 82e89ea077b9 ("virtio-blk: Add
>> validation for block size in config space") virtio-blk is broken on
>> s390.
>
> Because of a qemu bug. I agree. It's worth working around in the driver
> since the qemu bug has been around for a very long time.

Yes, since we introduced virtio 1 support, I guess...

>
>
>> The deeper problem is in the spec. We stated that the driver may read
>> config space before the feature negotiation is finalized, but we didn't
>> think enough about what happens when native endiannes is not little
>> endian in the different cases.
>
> Because the spec is very clear that endian-ness is LE.
> I don't see a spec issue yet here, just an implementation issue.

Maybe not really a bug in the spec, but probably an issue, as this seems
to have been unclear to most people so far.

>
>> I believe, for non-transitional devices we have a problem in the host as
>> well (i.e. in QEMU).
>
> Because QEMU ignores the spec and instead relies on the feature
> negotiation.
>
>> 
>> > I suspect it's actually the host that has the issue, not
>> > the guest?
>> 
>> I tend to say we have a problem both in the host and in the guest. I'm
>> more concerned about the problem in the guest, because that is a really
>> nasty regression.
>
> The problem is in the guest. The bug is in the host ;)
>
>> For the host. I think for legacy we don't have a
>> problem, because both sides would operate on the assumption no
>> _F_VERSION_1, IMHO the implementation for the transitional devices is
>> correct.
>
> Well no, the point of transitional is really to be 1.0 compliant
> *and* also expose a legacy interface.

Worth noting that PCI and CCW are a tad different here: PCI exposes an
additional interface, while CCW uses a revision negotiation mechanism
(for CCW, legacy and standard-compliant are much closer on the transport
side as for PCI.) MMIO does not do transitional, if I'm not wrong.

>
>> For non-transitional flavor, it depends on the device. For
>> example virtio-net and virtio-blk is broken, because we use primitives
>> like virtio_stl_p() and those don't do the right thing before feature
>> negotiation is completed. On the other hand virtio-crypto.c as a truly
>> non-transitional device uses stl_le_p() and IMHO does the right thing.
>> 
>> Thanks for your comments! I hope I managed to answer your questions. I
>> need some guidance on how do we want to move forward on this.
>> 
>> Regards,
>> Halil
>
> OK so. I don't have a problem with the patch itself,
> assuming it's enough to work around all buggy hosts.
> I am especially worried about things like vhost/vhost-user,
> I suspect they might have a bug like this too, and
> I am not sure whether your work around is enough for these.
> Can you check please?
>
> If not we'll have to move all validate code to after FEATURES_OK
> is set.

What is supposed to happen for validate after FEATURES_OK? The driver
cannot change any features at that point in time, it can only fail to
use the device.

>
> We do however want to document that this API can be called
> multiple times since that was not the case
> previously.
>
> Also, I would limit this to when
> - the validate callback exists
> - the guest endian-ness is not LE
>
> We also want to document the QEMU bug in a comment here,
> e.g. 
>
> /*
>  * QEMU before version 6.2 incorrectly uses driver features with guest
>  * endian-ness to set endian-ness for config space instead of just using
>  * LE for the modern interface as per spec.
>  * This breaks reading config in the validate callback.
>  * To work around that, when device is 1.0 (so supposed to be LE)
>  * but guest is not LE, then send the features to device one extra
>  * time before validation.
>  */

Do we need to consider migration, or do we not need to be bug-compatible
in this case?

>
> Finally I'd like to see the QEMU bug fix before I merge this one,
> since it will be harder to test with a fix.
>
>
>
>
>> > 
>> > 
>> > > Signed-off-by: Halil Pasic <pasic@linux.ibm.com>
>> > > Fixes: 82e89ea077b9 ("virtio-blk: Add validation for block size in config space")
>> > > Reported-by: markver@us.ibm.com
>> > > ---
>> > >  drivers/virtio/virtio.c | 4 ++++
>> > >  1 file changed, 4 insertions(+)
>> > > 
>> > > diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
>> > > index 0a5b54034d4b..9dc3cfa17b1c 100644
>> > > --- a/drivers/virtio/virtio.c
>> > > +++ b/drivers/virtio/virtio.c
>> > > @@ -249,6 +249,10 @@ static int virtio_dev_probe(struct device *_d)
>> > >  		if (device_features & (1ULL << i))
>> > >  			__virtio_set_bit(dev, i);
>> > >  
>> > > +	/* Write back features before validate to know endianness */
>> > > +	if (device_features & (1ULL << VIRTIO_F_VERSION_1))
>> > > +		dev->config->finalize_features(dev);
>> > > +
>> > >  	if (drv->validate) {
>> > >  		err = drv->validate(dev);
>> > >  		if (err)
>> > > 
>> > > base-commit: 02d5e016800d082058b3d3b7c3ede136cdc6ddcb
>> > > -- 
>> > > 2.25.1  
>> >