Re: [PATCH v2] packed-ring: fix example code

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Dec 04, 2018 at 06:49:29PM +0800, Tiwei Bie wrote:
> Driver can't just check whether USED bit equals to the used
> wrap counter when checking whether a descriptor is a used
> descriptor. Below is an example:
> 
> Assuming ring size is 4, ring's initial state will be:
> 
> +----+----+----+----+
> | 00 | 00 | 00 | 00 |
> +----+----+----+----+
> 
> 00 means AVAIL=0 USED=0, 01 means AVAIL=0 USED=1
> 10 means AVAIL=1 USED=0, 11 means AVAIL=1 USED=1
> 
> After driver made two descriptor chains available and each
> chain consists of two descriptors, the ring could be:
> 
> +----+-----------+----+-----------+
> | 10 | 10 (id=0) | 10 | 10 (id=1) |
> +----+-----------+----+-----------+
> 
> After device processed all the available descriptors and made
> them used (e.g. in order), the ring could be:
> 
> +-----------+----+-----------+----+
> | 11 (id=0) | 10 | 11 (id=1) | 10 |
> +-----------+----+-----------+----+
> 
> After driver processed all the used descriptors and made
> one descriptor (not chained, just one descriptor) available,
> the ring could be:
> 
> +-----------+----+----+----+
> | 01 (id=0) | 10 | 11 | 10 |
> +-----------+----+----+----+
> 
> After device made that descriptor used, the ring will be:
> 
> +-----------+----+----+----+
> | 00 (id=0) | 10 | 11 | 10 |
> +-----------+----+----+----+
> 
> If driver just checks whether USED bit equals to the used
> wrap counter when checking whether a descriptor is a used
> descriptor, after processing the first descriptor (whose
> AVAIL and USED bits are both 0), and advancing vq->next_used
> pointer, it will then also treat the next descriptor, i.e.
> the second descriptor (whose AVAIL and USED bits are 1 and
> 0 respectively) as a used descriptor which is wrong. The
> most straightforward way to fix it is to also check whether
> AVAIL bit equals to USED bit.
> 
> Fixes: https://github.com/oasis-tcs/virtio-spec/issues/29
> Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
> ---
> v2:
> - Add "Fixes" tag;
> - Refine commit log;
> 
>  packed-ring.tex | 6 ++++--
>  1 file changed, 4 insertions(+), 2 deletions(-)
> 
> diff --git a/packed-ring.tex b/packed-ring.tex
> index f24f49b..c913768 100644
> --- a/packed-ring.tex
> +++ b/packed-ring.tex
> @@ -688,15 +688,17 @@ for (;;) {
>          struct pvirtq_desc *d = vq->desc[vq->next_used];
>  
>          flags = d->flags;
> +        bool avail = flags & VIRTQ_DESC_F_AVAIL;
>          bool used = flags & VIRTQ_DESC_F_USED;
>  
> -        if (used != vq->used_wrap_count) {
> +        if (avail != used || used != vq->used_wrap_count) {

I think I prefer avail != vq->used_wrap_count.
This is imho clearer because both bits are controlled by the driver.

And maybe add a comment along the lines of

/*
 * Check that
 * 1. Descriptor has been made available.
 *    Note: there are many other way to check this, e.g.
 *    track number of outstanding available descriptors or buffers
 *    and check that it's not 0.
 * 2. Descriptor has been used by device.
 */

>                  vq->driver_event.flags = RING_EVENT_FLAGS_ENABLE;
>                  memory_barrier();
>  
>                  flags = d->flags;
> +                bool avail = flags & VIRTQ_DESC_F_AVAIL;
>                  bool used = flags & VIRTQ_DESC_F_USED;
> -                if (used != vq->used_wrap_count) {
> +                if (avail != used || used != vq->used_wrap_count) {

So why do we need to re-test avail here? I guess it's in case
driver is making descriptors available in parallel?
If so maybe explain that:

/* 
 * Re-test in case another thread submitted more descriptors
 * and/or device used more descriptors before driver enabled events.
 */


>                          break;
>                  }


>  
> -- 
> 2.17.1