On Mon, May 24, 2010 at 09:27:15AM -0700, David Stevens wrote: > netdev-owner@xxxxxxxxxxxxxxx wrote on 05/24/2010 09:13:51 AM: > > > On Mon, May 24, 2010 at 08:52:40AM -0700, David Stevens wrote: > > > "Michael S. Tsirkin" <mst@xxxxxxxxxx> wrote on 05/24/2010 03:17:10 AM: > > > > > > > On Thu, May 20, 2010 at 09:58:06AM -0700, David L Stevens wrote: > > > > > [for Michael Tsirkin's vhost development git tree] > > > > > > > > > > This patch fixes a race between guest and host when > > > > > adding used buffers wraps the ring. Without it, guests > > > > > can see partial packets before num_buffers is set in > > > > > the vnet header. > > > > > > > > > > Signed-off-by: David L Stevens <dlstevens@xxxxxxxxxx> > > > > > > > > Could you please explain what the race is? > > > > > > Sure. The pre-patch code in the ring-wrap case > > > does this: > > > > > > add part1 bufs > > > update used index > > > add part2 bufs > > > update used index > > > > > > After we update the used index for part1, the part1 > > > buffers are available to the guest. If the guest is > > > consuming at that point, it can process the partial > > > packet before the rest of the packet is there. In that > > > case, num_buffers will be greater than the number of > > > buffers available to the guest and it'll drop the > > > packet with a framing error. I was seeing 2 or 3 framing > > > errors every 100 million packets or so pre-patch, none > > > post-patch. > > > Actually, the second sentence is incorrect in the > > > original description-- num_buffers is up to date when > > > the guest sees it, but the used index is not. > > > > > > +-DLS > > > > so this happens always - what does wrap-around refer to? > > The 2-part update only happens when a packet spans > the end/beginning of the vring (the wrap). The framing error only > happens if the guest sees the vring-wrapping packets > before the second used-index write (the race). > So, the framing error doesn't happen always--it's > pretty rare. But with the patch, it never happens. > > +-DLS I see. The logging is still bugg though I think. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
