Re: [PATCH v3 2/2] vhost-user-blk: delay vhost_user_blk_disconnect

[Dima Stepanov]

On Mon, May 25, 2020 at 12:00:10AM -0400, Raphael Norwitz wrote:
> I'm mostly happy with this. A couple comments.
> 
> On Wed, May 20, 2020 at 11:54 AM Dima Stepanov <address@hidden> wrote:
> >
> > A socket write during vhost-user communication may trigger a disconnect
> > event, calling vhost_user_blk_disconnect() and clearing all the
> > vhost_dev structures holding data that vhost-user functions expect to
> > remain valid to roll back initialization correctly. Delay the cleanup to
> > keep vhost_dev structure valid.
> > There are two possible states to handle:
> > 1. RUN_STATE_PRELAUNCH: skip bh oneshot call and perform disconnect in
> > the caller routine.
> > 2. RUN_STATE_RUNNING: delay by using bh
> >
> > BH changes are based on the similar changes for the vhost-user-net
> > device:
> >   commit e7c83a885f865128ae3cf1946f8cb538b63cbfba
> >   "vhost-user: delay vhost_user_stop"
> >
> I'd also give credit to Li Feng here - he sent a similar patch:
> 
> https://lists.gnu.org/archive/html/qemu-devel/2020-04/msg02255.html
Yes, thanks for pointing me to it.

> 
> > Signed-off-by: Dima Stepanov <address@hidden>
> > ---
> >  hw/block/vhost-user-blk.c | 49 
> > +++++++++++++++++++++++++++++++++++++++++------
> >  1 file changed, 43 insertions(+), 6 deletions(-)
> >
> > diff --git a/hw/block/vhost-user-blk.c b/hw/block/vhost-user-blk.c
> > index 9d8c0b3..447fc9c 100644
> > --- a/hw/block/vhost-user-blk.c
> > +++ b/hw/block/vhost-user-blk.c
> > @@ -337,11 +337,6 @@ static void vhost_user_blk_disconnect(DeviceState *dev)
> >      VirtIODevice *vdev = VIRTIO_DEVICE(dev);
> >      VHostUserBlk *s = VHOST_USER_BLK(vdev);
> >
> > -    if (!s->connected) {
> > -        return;
> > -    }
> > -    s->connected = false;
> > -
> >      if (s->dev.started) {
> >          vhost_user_blk_stop(vdev);
> >      }
> > @@ -349,6 +344,19 @@ static void vhost_user_blk_disconnect(DeviceState *dev)
> >      vhost_dev_cleanup(&s->dev);
> >  }
> >
> > +static void vhost_user_blk_event(void *opaque, QEMUChrEvent event);
> > +
> > +static void vhost_user_blk_chr_closed_bh(void *opaque)
> > +{
> > +    DeviceState *dev = opaque;
> > +    VirtIODevice *vdev = VIRTIO_DEVICE(dev);
> > +    VHostUserBlk *s = VHOST_USER_BLK(vdev);
> > +
> > +    vhost_user_blk_disconnect(dev);
> > +    qemu_chr_fe_set_handlers(&s->chardev, NULL, NULL, vhost_user_blk_event,
> > +            NULL, opaque, NULL, true);
> > +}
> > +
> >  static void vhost_user_blk_event(void *opaque, QEMUChrEvent event)
> >  {
> >      DeviceState *dev = opaque;
> > @@ -363,7 +371,28 @@ static void vhost_user_blk_event(void *opaque, 
> > QEMUChrEvent event)
> >          }
> >          break;
> >      case CHR_EVENT_CLOSED:
> > -        vhost_user_blk_disconnect(dev);
> > +        /*
> > +         * A close event may happen during a read/write, but vhost
> > +         * code assumes the vhost_dev remains setup, so delay the
> > +         * stop & clear. There are two possible paths to hit this
> > +         * disconnect event:
> > +         * 1. When VM is in the RUN_STATE_PRELAUNCH state. The
> > +         * vhost_user_blk_device_realize() is a caller.
> > +         * 2. In tha main loop phase after VM start.
> > +         *
> > +         * For p2 the disconnect event will be delayed. We can't
> > +         * do the same for p1, because we are not running the loop
> > +         * at this moment. So just skip this step and perform
> > +         * disconnect in the caller function.
> > +         */
> > +        if (s->connected && runstate_is_running()) {
> > +            AioContext *ctx = qemu_get_current_aio_context();
> > +
> > +            qemu_chr_fe_set_handlers(&s->chardev, NULL, NULL, NULL, NULL,
> > +                    NULL, NULL, false);
> > +            aio_bh_schedule_oneshot(ctx, vhost_user_blk_chr_closed_bh, 
> > opaque);
> > +        }
> > +        s->connected = false;
> >          break;
> >      case CHR_EVENT_BREAK:
> >      case CHR_EVENT_MUX_IN:
> > @@ -428,6 +457,14 @@ reconnect:
> >
> >      ret = vhost_dev_get_config(&s->dev, (uint8_t *)&s->blkcfg,
> >                                 sizeof(struct virtio_blk_config));
> 
> I find checking s->connected before ret a little confusing. I think we
> should also enforce a reconnect if s->connected is false. AFIK if
> s->connected is false, ret must also be less than 0, but to be safe
> I’d prefer something like:
> 
> if (ret < 0 || !s->connected) {
>     if (!s->connected) {
>         /*
>          * Perform disconnect before making reconnect. More detailed
>          * comment why it was delayed is in the vhost_user_blk_event()
>          * routine.
>          */
>           vhost_user_blk_disconnect(dev);
>     }
>     if (ret < 0) {
>            error_report(“vhost-user-blk: get block config failed”);
>     }
>     goto reconnect;
> }
> 
True. Thanks to Li Feng's patch i understood that i've overcomplicated the
logic. We don't need this disconnect call here at all.
I'll send a smaller reworked patch in this e-mail thread, just to
continue review and discussion.

> > +    if (!s->connected) {
> > +        /*
> > +         * Perform disconnect before making reconnect. More detailed
> > +         * comment why it was delayed is in the vhost_user_blk_event()
> > +         * routine.
> > +         */
> > +        vhost_user_blk_disconnect(dev);
> > +    }
> >      if (ret < 0) {
> >          error_report("vhost-user-blk: get block config failed");
> >          goto reconnect;
> > --
> > 2.7.4
> >
> >