On Mon, Oct 26, 2015 at 01:33:12PM +0100, Samuel Thibault wrote: > Wei Liu, le Mon 26 Oct 2015 12:30:28 +0000, a écrit : > > On Mon, Oct 26, 2015 at 01:21:51PM +0100, Samuel Thibault wrote: > > > Wei Liu, le Mon 26 Oct 2015 12:14:43 +0000, a écrit : > > > > In my patch, mini-os notifies remote whenever it consumes a message, > > > > which I think it's slightly better because backend can start putting > > > > things in the ring as mini-os processes them. > > > > > > That makes more notifications, but that can lead to more pipelining > > > indeed. That's what the Linux driver does, so let's do the same. > > > > > > Also, I'm realizing: aren't we missing a full memory barrier between > > > the memcpy_from_ring call and xenstore_buf->rsp_cons += ? (in the two > > > places) We need to make sure to have finished copying from the ring > > > before writing the new rsp_cons. > > > > > > > You're right. > > > > I think we should just turn that wmb() into two mb()s and place them > > before xenstore_buf->rsp_cons +=. > > We *also* need some barrier between rsp_cons += and the notification, > otherwise the notified domain may miss the rsp_cons update and thus > believe it was a spurious notification. >
Actually notify_remote_via_evtchn normally implies a mb(). But I think I'd stay on the safe side. I will have a wmb() there. :-) Thanks for your review. V2 coming soon. Wei. > Samuel > > _______________________________________________ > Minios-devel mailing list > minios-de...@lists.xenproject.org > http://lists.xenproject.org/cgi-bin/mailman/listinfo/minios-devel _______________________________________________ Xen-devel mailing list Xen-devel@lists.xen.org http://lists.xen.org/xen-devel
