Virtualization - Apr 2022 - [PATCH V2 4/5] virtio-pci: implement synchronize_vqs()

On Tue, Apr 12 2022, Halil Pasic <pasic at linux.ibm.com> wrote:
> On Mon, 11 Apr 2022 16:27:41 +0200
> Cornelia Huck <cohuck at redhat.com> wrote:
>> My main concern is that we would need to synchronize against a single
>> interrupt that covers all kinds of I/O interrupts, not just a single
>> device...
>> 
>
> Could we synchronize on struct airq_info's lock member? If we were
> to grab all of these that might be involved...
Hm, that could possibly narrow the sync down to a subset, which seems
better. For devices still using classic interrupts, per-device sync
would be easy.
>
> AFAIU for the synchronize implementation we need a lock or a set of locks
> that contain all the possible vring_interrupt() calls with the queuues
> that belong to the given device as a critical section. That way, one
> has the acquire's and release's in place so that the
vrign_interrupt()
> either guaranteed to finish before the change of driver_ready is
> guaranteed to be complete, or it is guaranteed to see the change.
>
> In any case, I guess we should first get clear on the first part. I.e.
> when do we want to allow host->guest notifications.
Also, whether we just care about vring interrupts, or general device
interrupts (not sure if a config change interrupt may also trigger
things we do not want to trigger?)

On Wed, Apr 13, 2022 at 12:49 AM Cornelia Huck <cohuck at redhat.com>
wrote:
>
> On Tue, Apr 12 2022, Halil Pasic <pasic at linux.ibm.com> wrote:
>
> > On Mon, 11 Apr 2022 16:27:41 +0200
> > Cornelia Huck <cohuck at redhat.com> wrote:
>
> >> My main concern is that we would need to synchronize against a
single
> >> interrupt that covers all kinds of I/O interrupts, not just a
single
> >> device...
> >>
> >
> > Could we synchronize on struct airq_info's lock member? If we were
> > to grab all of these that might be involved...
>
> Hm, that could possibly narrow the sync down to a subset, which seems
> better. For devices still using classic interrupts, per-device sync
> would be easy.
>
> >
> > AFAIU for the synchronize implementation we need a lock or a set of
locks
> > that contain all the possible vring_interrupt() calls with the queuues
> > that belong to the given device as a critical section. That way, one
> > has the acquire's and release's in place so that the
vrign_interrupt()
> > either guaranteed to finish before the change of driver_ready is
> > guaranteed to be complete, or it is guaranteed to see the change.
> >
> > In any case, I guess we should first get clear on the first part. I.e.
> > when do we want to allow host->guest notifications.
>
> Also, whether we just care about vring interrupts, or general device
> interrupts (not sure if a config change interrupt may also trigger
> things we do not want to trigger?)
I think only vring interrupts, since the config interrupt hardening is
done via 22b7050a024d7 ("virtio: defer config changed notifications")

Thanks
>