Re: [PATCH] virtio_ring: add an error code check in virtqueue_resize

From: Michael S. Tsirkin
Date: Mon Oct 23 2023 - 07:19:44 EST


On Mon, Oct 23, 2023 at 06:52:34PM +0800, Xuan Zhuo wrote:
> On Mon, 23 Oct 2023 17:52:02 +0800, Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx> wrote:
> > On Mon, 23 Oct 2023 17:50:46 +0800, Su Hui <suhui@xxxxxxxxxxxx> wrote:
> > > On 2023/10/23 13:46, Xuan Zhuo wrote:
> > > >>>>>>>> Well, what are the cases where it can happen practically?
> > > >>>>>>> Device error. Such as vp_active_vq()
> > > >>>>>>>
> > > >>>>>>> Thanks.
> > > >>>>>> Hmm interesting. OK. But do callers know to recover?
> > > >>>>> No.
> > > >>>>>
> > > >>>>> So I think WARN + broken is suitable.
> > > >>>>>
> > > >>>>> Thanks.
> > > >>>> Sorry for the late, is the following code okay?
> > > >>>>
> > > >>>> @@ -2739,7 +2739,7 @@ int virtqueue_resize(struct virtqueue *_vq, u32 num,
> > > >>>> void (*recycle)(struct virtqueue *vq, void *buf))
> > > >>>> {
> > > >>>> struct vring_virtqueue *vq = to_vvq(_vq);
> > > >>>> - int err;
> > > >>>> + int err, err_reset;
> > > >>>>
> > > >>>> if (num > vq->vq.num_max)
> > > >>>> return -E2BIG;
> > > >>>> @@ -2759,7 +2759,15 @@ int virtqueue_resize(struct virtqueue *_vq, u32 num,
> > > >>>> else
> > > >>>> err = virtqueue_resize_split(_vq, num);
> > > >>>>
> > > >>>> - return virtqueue_enable_after_reset(_vq);
> > > >>>> + err_reset = virtqueue_enable_after_reset(_vq);
> > > >>>> +
> > > >>>> + if (err) {
> > > >>> No err.
> > > >>>
> > > >>> err is not important.
> > > >>> You can remove that.
> > > >> Emm, I'm a little confused that which code should I remove ?
> > > >>
> > > >>
> > > >> like this:
> > > >> if (vq->packed_ring)
> > > >> virtqueue_resize_packed(_vq, num);
> > > >> else
> > > >> virtqueue_resize_split(_vq, num);
> > > >>
> > > >> And we should set broken and warn inside virtqueue_enable_after_reset()?
> > >
> > > In my opinion, we should return the error code of virtqueue_resize_packed() / virtqueue_resize_split().
> > > But if this err is not important, this patch makes no sense.
> > > Maybe I misunderstand somewhere...
> > > If you think it's worth sending a patch, you can send it :).(I'm not familiar with this code).
> >
> > OK.
>
> Hi Michael,
>
> The queue reset code is wrote with the CONFIG_VIRTIO_HARDEN_NOTIFICATION.
>
> When we disable the vq, the broken is true until we re-enable it.
>
> So when we re-enable it fail, the vq is broken status.
>
> Normally, this just happens on the buggy device.
> So I think that is enough.
>
> Thanks.

I don't know what to do about CONFIG_VIRTIO_HARDEN_NOTIFICATION.
It's known to be broken and it does not look like there's
active effort to revive it - should we just drop it for now?


>
> static int vp_modern_disable_vq_and_reset(struct virtqueue *vq)
> {
> [...]
>
> vp_modern_set_queue_reset(mdev, vq->index);
>
> [...]
>
> #ifdef CONFIG_VIRTIO_HARDEN_NOTIFICATION
> ->> __virtqueue_break(vq);
> #endif
>
> [...]
> }
>
> static int vp_modern_enable_vq_after_reset(struct virtqueue *vq)
> {
> [...]
>
> if (vp_modern_get_queue_reset(mdev, index))
> return -EBUSY;
>
> if (vp_modern_get_queue_enable(mdev, index))
> return -EBUSY;
>
> err = vp_active_vq(vq, info->msix_vector);
> if (err)
> return err;
>
> }
>
> [...]
>
> #ifdef CONFIG_VIRTIO_HARDEN_NOTIFICATION
> ->> __virtqueue_unbreak(vq);
> #endif
>
> [...]
> }
>
>
>
>
> >
> > Thanks.
> >
> >
> > >
> > > Thanks,
> > > Su Hui
> > >