Re: [PATCH V2] vhost: fix check for # of outstanding buffers

[Shirley Ma <mashirle@us.ibm.com>]

On Thu, 2011-07-21 at 11:06 +0300, Michael S. Tsirkin wrote:
> On Wed, Jul 20, 2011 at 10:23:12AM -0700, Shirley Ma wrote:
> > Fix the check for number of outstanding buffers returns incorrect
> > results due to vq->pend_idx wrap around;
> > 
> > Signed-off-by: Shirley Ma <xma@us.ibm.com>
> 
> OK, the logic's right now, and it's not worse
> than what we had, so I applied this after
> fixing up the comment (it's upend_idx and English
> sentences don't need to end with a semicolumn ;)
> 
> However, I would like to see the effect of the bug
> noted in the log in the future.
> 
> And the reason I mention this here, is that
> I think that the whole VHOST_MAX_PEND thing
> does not work as advertised: this logic only
> triggers when the ring is empty, so we will happily push
> more than VHOST_MAX_PEND packets if the guest manages
> to give them to us.
> 
> I'm not sure why we have the limit, either: the wmem
> limit in the socket still applies and seems more
> effective to prevent denial of service by a malicious guest.

Vhost can push more than VHOST_MAX_PEND if the guest manages to give
more. That's managed by wmem limit.

MAX_PEND is max of outstanding used buffers which lower level device
can't DMAed on time. socket destructor remains unchanged, so it can't
managed by wmem.

Since vhost handle_tx always calls vhost_zerocopy_singal_used() so this
condition is unlikely hit unless the lower device can't DMAed TX
MAX_PEND packets.

Thanks
Shirley