On Tue, May 08, 2018 at 05:34:40PM +0800, Jason Wang wrote:

On 2018年05月08日 17:16, Tiwei Bie wrote:

quoted

On Tue, May 08, 2018 at 03:16:53PM +0800, Jason Wang wrote:

quoted

On 2018年05月08日 14:44, Tiwei Bie wrote:

quoted

On Tue, May 08, 2018 at 01:40:40PM +0800, Jason Wang wrote:

quoted

On 2018年05月08日 11:05, Jason Wang wrote:

quoted

quoted

Because in virtqueue_enable_cb_delayed(), we may set an
event_off which is bigger than new and both of them have
wrapped. And in this case, although new is smaller than
event_off (i.e. the third param -- old), new shouldn't
add vq->num, and actually we are expecting a very big
idx diff.

Yes, so to calculate distance correctly between event and new, we just
need to compare the warp counter and return false if it doesn't match
without the need to try to add vq.num here.

Thanks

Sorry, looks like the following should work, we need add vq.num if
used_wrap_counter does not match:

static bool vhost_vring_packed_need_event(struct vhost_virtqueue *vq,
                        __u16 off_wrap, __u16 new,
                        __u16 old)
{
      bool wrap = off_wrap >> 15;
      int off = off_wrap & ~(1 << 15);
      __u16 d1, d2;

      if (wrap != vq->used_wrap_counter)
          d1 = new + vq->num - off - 1;

Just to draw your attention (maybe you have already
noticed this).

I miss this, thanks!

quoted

In this case (i.e. wrap != vq->used_wrap_counter),
it's also possible that (off < new) is true. Because,

when virtqueue_enable_cb_delayed_packed() is used,
`off` is calculated in driver in a way like this:

	off = vq->last_used_idx + bufs;
	if (off >= vq->vring_packed.num) {
		off -= vq->vring_packed.num;
		wrap_counter ^= 1;
	}

And when `new` (in vhost) is close to vq->num. The
vq->last_used_idx + bufs (in driver) can be bigger
than vq->vring_packed.num, and:

1. `off` will wrap;
2. wrap counters won't match;
3. off < new;

And d1 (i.e. new + vq->num - off - 1) will be a value
bigger than vq->num. I'm okay with this, although it's
a bit weird.

So I'm considering something more compact by reusing vring_need_event() by
pretending a larger queue size and adding vq->num back when necessary:

static bool vhost_vring_packed_need_event(struct vhost_virtqueue *vq,
                       __u16 off_wrap, __u16 new,
                       __u16 old)
{
     bool wrap = vq->used_wrap_counter;

If the wrap counter is obtained from the vq,
I think `new` should also be obtained from
the vq. Or the wrap counter should be carried
in `new`.

quoted

     int off = off_wrap & ~(1 << 15);
     __u16 d1, d2;

     if (new < old) {
         new += vq->num;
         wrap ^= 1;
     }

     if (wrap != off_wrap >> 15)
         off += vq->num;

When `new` and `old` wraps, and `off` doesn't wrap,
wrap != (off_wrap >> 15) will be true. In this case,
`off` is bigger than `new`, and what we should do
is `off -= vq->num` instead of `off += vq->num`.

If I understand this correctly, if we track old correctly, it won't happen
if guest driver behave correctly. That means it should only happen for a
buggy driver (e.g trying to move off_wrap back).

If vhost is faster than virtio driver, I guess above
case may happen. The `old` and `new` will be updated
each time we want to notify the driver. If the driver
is slower, `old` and `new` in vhost may wrap before
the `off` which is set by driver wraps.

Best regards,
Tiwei Bie

Thanks

quoted

Best regards,
Tiwei Bie

quoted

     return vring_need_event(off, new, old);
}

quoted

Best regards,
Tiwei Bie

quoted

      else
          d1 = new - off - 1;

      if (new > old)
          d2 = new - old;
      else
          d2 = new + vq->num - old;

      return d1 < d2;
}

Thanks