Re: [Qemu-devel] [PATCH V3 2/3] virtio-blk: fail get_features when both scsi and 1.0 were set

[Cornelia Huck]

On Wed, 22 Jul 2015 19:34:31 +0300
"Michael S. Tsirkin" <address@hidden> wrote:

> On Wed, Jul 22, 2015 at 06:11:16PM +0200, Cornelia Huck wrote:
> > On Wed, 22 Jul 2015 17:53:47 +0300
> > "Michael S. Tsirkin" <address@hidden> wrote:
> > 
> > > On Wed, Jul 22, 2015 at 12:55:22PM +0200, Cornelia Huck wrote:
> > > > On Wed, 22 Jul 2015 13:44:14 +0300
> > > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > > 
> > > > > On Wed, Jul 22, 2015 at 12:38:40PM +0200, Cornelia Huck wrote:
> > > > > > On Wed, 22 Jul 2015 13:32:17 +0300
> > > > > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > > > > 
> > > > > > > On Wed, Jul 22, 2015 at 12:25:31PM +0200, Cornelia Huck wrote:
> > > > > > > > On Wed, 22 Jul 2015 12:21:32 +0300
> > > > > > > > "Michael S. Tsirkin" <address@hidden> wrote:
> > > > > > > > 
> > > > > > > > > On Wed, Jul 22, 2015 at 10:58:43AM +0200, Cornelia Huck wrote:
> > > > > > > > > > On Wed, 22 Jul 2015 13:59:51 +0800
> > > > > > > > > > Jason Wang <address@hidden> wrote:
> > > > > > > > > > 
> > > > > > > > > > > SCSI passthrough was no longer supported in virtio 1.0, 
> > > > > > > > > > > so this patch
> > > > > > > > > > > fail the get_features() when both 1.0 and scsi is set. 
> > > > > > > > > > > And also only
> > > > > > > > > > > advertise VIRTIO_BLK_F_SCSI for legacy virtio-blk device.
> > > > > > > > > > > 
> > > > > > > > > > > Signed-off-by: Jason Wang <address@hidden>
> > > > > > > > > > > ---
> > > > > > > > > > >  hw/block/virtio-blk.c | 9 ++++++++-
> > > > > > > > > > >  1 file changed, 8 insertions(+), 1 deletion(-)
> > > > > > > > > > > 
> > > > > > > > > > > diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c
> > > > > > > > > > > index 4c27974..4716c3e 100644
> > > > > > > > > > > --- a/hw/block/virtio-blk.c
> > > > > > > > > > > +++ b/hw/block/virtio-blk.c
> > > > > > > > > > > @@ -731,7 +731,14 @@ static uint64_t 
> > > > > > > > > > > virtio_blk_get_features(VirtIODevice *vdev, uint64_t 
> > > > > > > > > > > features,
> > > > > > > > > > >      virtio_add_feature(&features, VIRTIO_BLK_F_GEOMETRY);
> > > > > > > > > > >      virtio_add_feature(&features, VIRTIO_BLK_F_TOPOLOGY);
> > > > > > > > > > >      virtio_add_feature(&features, VIRTIO_BLK_F_BLK_SIZE);
> > > > > > > > > > > -    virtio_add_feature(&features, VIRTIO_BLK_F_SCSI);
> > > > > > > > > > > +    if (__virtio_has_feature(features, 
> > > > > > > > > > > VIRTIO_F_VERSION_1)) {
> > > > > > > > > > > +        if (s->conf.scsi) {
> > > > > > > > > > > +            error_setg(errp, "Virtio 1.0 does not 
> > > > > > > > > > > support scsi passthrough!");
> > > > > > > > > > > +            return 0;
> > > > > > > > > > > +        }
> > > > > > > > > > > +    } else {
> > > > > > > > > > > +        virtio_add_feature(&features, VIRTIO_BLK_F_SCSI);
> > > > > > > > > > > +    }
> > > > > > > > > > > 
> > > > > > > > > > >      if (s->conf.config_wce) {
> > > > > > > > > > >          virtio_add_feature(&features, 
> > > > > > > > > > > VIRTIO_BLK_F_CONFIG_WCE);
> > > > > > > > > > 
> > > > > > > > > > Do we advertise F_SCSI even if scsi is not configured in 
> > > > > > > > > > order to keep
> > > > > > > > > > the same bits as before? I'm afraid I don't remember, that 
> > > > > > > > > > thread was
> > > > > > > > > > long :/
> > > > > > > > > > 
> > > > > > > > > > I'm asking because I'd like to depend on that bit to decide 
> > > > > > > > > > whether I
> > > > > > > > > > can negotiate revision 1 for ccw and subsequently offer 
> > > > > > > > > > VERSION_1. It
> > > > > > > > > > would be an easy thing to do, and I'd like to avoid mucking 
> > > > > > > > > > around with
> > > > > > > > > > device-specific configuration from the transport.
> > > > > > > > > > 
> > > > > > > > > > To illustrate what I'm talking about, my current patchset 
> > > > > > > > > > for virtio-1
> > > > > > > > > > on ccw is here:
> > > > > > > > > > 
> > > > > > > > > > git://github.com/cohuck/qemu virtio-1-ccw-2.5
> > > > > > > > > 
> > > > > > > > > 
> > > > > > > > > I still think you are over-engineering it.
> > > > > > > > > 
> > > > > > > > > Just add a property to disable the modern interface.
> > > > > > > > > Anyone using scsi passthrough will have to set that,
> > > > > > > > > if not - above patch will make initialization fail.
> > > > > > > > 
> > > > > > > > And I still think requiring user action and not having this work
> > > > > > > > transparently is a bad idea...
> > > > > > > 
> > > > > > > Having what work transparently? SCSI passthrough?
> > > > > > > Look, either you agree with Paolo or not.
> > > > > > > Paolo thinks it's a deprecated hack not really worth supporting 
> > > > > > > long term.
> > > > > > > If you agree, I don't see why is asking for an extra property
> > > > > > > such a bit deal. If you don't agree - please open a new thread
> > > > > > > and argue about that.
> > > > > > 
> > > > > > I sometimes wonder whether we're arguing about the same thing :(
> > > > > > 
> > > > > > Dropping scsi for virtio-1 is fine. Dropping 
> > > > > > backwards-compatibility is
> > > > > > not. If I upgrade the host, I want the guests to be able to continue
> > > > > > using scsi without needing to fence virtio-1 off manually.
> > > > > 
> > > > > Paolo's argument is that no one should be using scsi passthrough.
> > > > > 
> > > > > If the feature has users, we should bring it back into virtio 1.
> > > > > 
> > > > > If almost one uses it, then no one will suffer too much from getting
> > > > > an error message saying "please set disable-modern=on".
> > > > 
> > > > And here's where we disagree. Even if it's exotic, I don't want to
> > > > break existing users.
> > > 
> > > You should take this disagreement to the virtio TC.  QEMU merely
> > > implements what the spec voted by TC says.
> > 
> > I fail to see why this is a spec issue. It sounds like a qemu
> > implementation question to me.
> 
> Pls re-read the discussion around removing this issue then.
> 
> 
> virtio 1 should be a super-set of virtio functionality-wise.

Perhaps this assumption is the problem then? scsi seems to be the only
odd one, but still...

> 
> 
> Paolo said this specific feature has been deprecated for years, no
> one should be using it, so we dropped it from spec.

Nothing wrong with that. But "it's deprecated" does not mean "nobody's
using it", I fear. The question is: Should qemu accommodate those users?

> 
> 
> > > 
> > > > > And there's no reason to make it behave differently
> > > > > between ccw and pci.
> > > > > 
> > > > > 
> > > > > > > 
> > > > > > > 
> > > > > > > > Moreover, I will need a revision-fencing mechanism in any case, 
> > > > > > > > when we
> > > > > > > > introduce further revisions.
> > > > > > > 
> > > > > > > Why? Assuming we drop more features in the future?
> > > > > > 
> > > > > > Revisions != features. Think new or changed channel commands, for
> > > > > > example.
> > > > > 
> > > > > You likely can just add these unconditionally.
> > > > 
> > > > Backwards compatibility?
> > > 
> > > Compatibility is built-in to revision negotiation, isn't it?
> > 
> > Yes, but we still want to support migration to older releases.
> 
> I don't think you can do anything in virtio to make this work
> for PPC.  As long as you don't version machine types, cross version
> migration will be broken.

ccw is s390x :)

And we added versioned machines for 2.4, so yes, we care.

Which brings me to the next problem:

Assuming we want to make blocking VERSION_1 devices user-configurable,
we'll need a max_revision property or something like that. (pci's
legacy/modern approach just won't cut it, I fear, since we'll need to
handle higher revisions for later cross-version migration purposes as
well.) This implies we need to add this max_rev field to virtio-ccw's
migration stream _now_ (for 2.4).

I've started hacking up a patch, but I'm no way finished.

So:
- checking scsi flag in features for revision fencing won't work (we
  always need to advertise it for !VERSION_1)
- checking if it's a blk device with scsi configured for revision
  fencing is so ugly that I won't even try to code it
- which leaves setting VERSION_1 from the start and have virtio-blk
  fence VERSION_1 + scsi. For which I need the max_revision property
  with the issues I outlined above.

Not sure how to get out of this pickle.