On Wed, Aug 4, 2021 at 8:41 AM Stefano Garzarella <sgarz...@redhat.com>
wrote:
>
> On Tue, Aug 03, 2021 at 11:58:27AM -0700, Jiang Wang . wrote:
> >On Wed, Jul 7, 2021 at 10:27 AM Stefano Garzarella <sgarz...@redhat.com>
> >wrote:
> >> On Wed, Jul 07, 2021 at 09:52:46AM -0700, Jiang Wang . wrote:
> >> >On Wed, Jul 7, 2021 at 1:33 AM Stefano Garzarella <sgarz...@redhat.com>
> >> >wrote:
> >> >> On Tue, Jul 06, 2021 at 10:26:07PM +0000, Jiang Wang wrote:
>
> [...]
>
> >> >> >+
> >> >> >+ if (nvqs < 0)
> >> >> >+ nvqs = MAX_VQS_WITHOUT_DGRAM;
> >> >> >+
> >> >> >+ if (nvqs == MAX_VQS_WITH_DGRAM) {
> >> >> >+ vvc->dgram_recv_vq = virtio_add_queue(vdev,
> >> >> >VHOST_VSOCK_QUEUE_SIZE,
> >> >> >+
> >> >> >vhost_vsock_common_handle_output);
> >> >> >+ vvc->dgram_trans_vq = virtio_add_queue(vdev,
> >> >> >VHOST_VSOCK_QUEUE_SIZE,
> >> >> >+
> >> >> >vhost_vsock_common_handle_output);
> >> >> >+ }
> >> >> >+
> >> >> > /* The event queue belongs to QEMU */
> >> >> > vvc->event_vq = virtio_add_queue(vdev, VHOST_VSOCK_QUEUE_SIZE,
> >> >> >
> >> >> > vhost_vsock_common_handle_output);
> >> >>
> >> >> Did you do a test with a guest that doesn't support datagram with QEMU
> >> >> and hosts that do?
> >> >>
> >> >Yes, and it works.
> >> >
> >> >> I repost my thoughts that I had on v2:
> >> >>
> >> >> What happen if the guest doesn't support dgram?
> >> >>
> >> >> I think we should dynamically use the 3rd queue or the 5th queue
> >> >> for
> >> >> the events at runtime after the guest acked the features.
> >> >>
> >> >> Maybe better to switch to an array of VirtQueue.
> >> >>
> >> >I think in current V3, it already dynamically use 3rd or 5th queue
> >> >depending
> >> >on the feature bit.
> >>
> >> I'm not sure. IIUC when vhost_vsock_common_realize() is called, we don't
> >> know the features acked by the guest, so how can it be dynamic?
> >>
> >> Here we should know only if the host kernel supports it.
> >>
> >> Maybe it works, because in QEMU we use the event queue only after a
> >> migration to send a reset event, so you can try to migrate a guest to
> >> check this path.
> >>
> >I tried VM migration and didn't see any problems. The migration looks fine
> >and vsock dgram still works after migration. Is there any more specific test
> >you want to run to check for this code path?
> >
>
> I meant a migration of a guest from QEMU without this patch to a QEMU
> with this patch. Of course in that case testing a socket stream.
>
Sorry, I meant the opposite.
You should try to migrate a guest that does not support dgrams starting
from a QEMU with this patch (and kernel that supports dgram, so qemu
uses the 5th queue for event), to a QEMU without this patch.
I still think the event queue should remain the third and those dgrams
the next 2.
Thanks,
Stefano
