Re: [PATCH RFC v6 02/11] vhost: use batched get_vq_desc version

[Jason Wang]

On 2020/6/10 äå7:05, Michael S. Tsirkin wrote:
> > > +EXPORT_SYMBOL_GPL(vhost_get_vq_desc);
> > >    /* Reverse the effect of vhost_get_vq_desc. Useful for error handling. */
> > >    void vhost_discard_vq_desc(struct vhost_virtqueue *vq, int n)
> > >    {
> > > +	unfetch_descs(vq);
> > >    	vq->last_avail_idx -= n;
> > So unfetch_descs() has decreased last_avail_idx.
> > Can we fix this by letting unfetch_descs() return the number and then we can
> > do:
> >
> > int d = unfetch_descs(vq);
> > vq->last_avail_idx -= (n > d) ? n - d: 0;
> >
> > Thanks
> That's intentional I think - we need both.


Yes, but:


> Unfetch_descs drops the descriptors in the cache that were
> *not returned to caller*  through get_vq_desc.
>
> vhost_discard_vq_desc drops the ones that were returned through get_vq_desc.
>
> Did I miss anything?

We could count some descriptors twice, consider the case e.g we only 
cache on descriptor:

fetch_descs()
ÂÂÂ fetch_buf()
ÂÂÂ ÂÂÂ last_avail_idx++;

Then we want do discard it:
vhost_discard_avail_buf(1)
ÂÂÂ unfetch_descs()
ÂÂÂ ÂÂÂ last_avail_idx--;
ÂÂÂ last_avail_idx -= 1;

Thanks