Re: [PATCH 1/2] virtio: reset region cache when on queue deletion

[Michael S. Tsirkin]

I guess it somehow has to do with the following:

    if (proxy->disable_legacy == ON_OFF_AUTO_AUTO) {
        proxy->disable_legacy = pcie_port ? ON_OFF_AUTO_ON : ON_OFF_AUTO_OFF;
    }

so by default device on an express port does not have a legacy interface.

Somehow having a legacy interface fixes things?
Does enabling legacy on q35 without this patch fix things?
And does disabling legacy on PIIX without this patch break thing?

How can having a legacy interface fix things if it's not
even active? I don't know. Is there a chance windows drivers use the
legacy interface on a transitional device by mistake?


On Mon, Jan 06, 2020 at 11:10:18AM +0200, Yuri Benditovich wrote:
> Michael,
> Can you please comment on Jason's question: why we have a problem only with 
> q35
> and not with legacy pc?
> If you have a simple answer, it will help us in further work with other hot
> plug/unplug problems.
> 
> Thanks,
> Yuri Benditovich
> 
> On Sun, Jan 5, 2020 at 6:21 PM Yuri Benditovich <address@hidden>
> wrote:
> 
> 
> 
>     On Sun, Jan 5, 2020 at 1:39 PM Michael S. Tsirkin <address@hidden> wrote:
> 
>         On Thu, Jan 02, 2020 at 09:09:04AM +0200, Yuri Benditovich wrote:
>         >
>         >
>         > On Thu, Jan 2, 2020 at 1:50 AM Michael S. Tsirkin <address@hidden>
>         wrote:
>         >
>         >     On Thu, Dec 26, 2019 at 11:29:50AM +0200, Yuri Benditovich 
> wrote:
>         >     > On Thu, Dec 26, 2019 at 10:58 AM Jason Wang <
>         address@hidden> wrote:
>         >     > >
>         >     > >
>         >     > > On 2019/12/26 下午12:36, Yuri Benditovich wrote:
>         >     > > > https://bugzilla.redhat.com/show_bug.cgi?id=1708480
>         >     > > > Fix leak of region reference that prevents complete
>         >     > > > device deletion on hot unplug.
>         >     > >
>         >     > >
>         >     > > More information is needed here, the bug said only q35 can
>         meet this
>         >     > > issue. What makes q35 different here?
>         >     > >
>         >     >
>         >     > I do not have any ready answer, I did not dig into it too 
> much.
>         >     > Probably Michael Tsirkin or Paolo Bonzini can answer without
>         digging.
>         >
>         >
>         >
>         >     > >
>         >     > > >
>         >     > > > Signed-off-by: Yuri Benditovich <
>         address@hidden>
>         >     > > > ---
>         >     > > >   hw/virtio/virtio.c | 5 +++++
>         >     > > >   1 file changed, 5 insertions(+)
>         >     > > >
>         >     > > > diff --git a/hw/virtio/virtio.c b/hw/virtio/virtio.c
>         >     > > > index 04716b5f6c..baadec8abc 100644
>         >     > > > --- a/hw/virtio/virtio.c
>         >     > > > +++ b/hw/virtio/virtio.c
>         >     > > > @@ -2340,6 +2340,11 @@ void virtio_del_queue(VirtIODevice
>         *vdev, int
>         >     n)
>         >     > > >       vdev->vq[n].vring.num_default = 0;
>         >     > > >       vdev->vq[n].handle_output = NULL;
>         >     > > >       vdev->vq[n].handle_aio_output = NULL;
>         >     > > > +    /*
>         >     > > > +     * with vring.num = 0 the queue will be ignored
>         >     > > > +     * in later loops of region cache reset
>         >     > > > +     */
>         >     > >
>         >     > >
>         >     > > I can't get the meaning of this comment.
>         >     > >
>         >     > > Thanks
>         >     > >
>         >     > >
>         >     > > > +    virtio_virtqueue_reset_region_cache(&vdev->vq[n]);
>         >
>         >
>         >     Do we need to drop this from virtio_device_free_virtqueues then?
>         >
>         >
>         >
>         > Not mandatory. Repetitive  virtio_virtqueue_reset_region_cache does
>         not do
>         > anything bad.
>         > Some of virtio devices do not do 'virtio_del_queue' at all.
>         Currently 
>         > virtio_device_free_virtqueues resets region cache for them.
>         > IMO, not calling 'virtio_del_queue' is a bug, but not in the scope 
> of
>         current
>         > series, I'll take care of that later.
> 
>         Maybe we should just del all queues in virtio_device_unrealize?
>         Will allow us to drop some logic tracking which vqs were created.
> 
> 
> 
>     Yes, this is also possible with some rework of
>     virtio_device_free_virtqueues.
>     virtio-net has some additional operations around queue deletion, it 
> deletes
>     queues when switches from single queue to multiple.
>      
> 
> 
>         >
>         >     > > >       g_free(vdev->vq[n].used_elems);
>         >     > > >   }
>         >     > > >
>         >     > >
>         >
>         >
> 
>