From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([208.118.235.92]:41490) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1S4vIV-0000DJ-8e for qemu-devel@nongnu.org; Tue, 06 Mar 2012 09:23:33 -0500 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1S4vIO-0004Nb-IX for qemu-devel@nongnu.org; Tue, 06 Mar 2012 09:23:26 -0500 Received: from mx1.redhat.com ([209.132.183.28]:45594) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1S4vIO-0004NW-AI for qemu-devel@nongnu.org; Tue, 06 Mar 2012 09:23:20 -0500 Date: Tue, 6 Mar 2012 16:23:09 +0200 From: Alon Levy Message-ID: <20120306142309.GE2240@garlic.redhat.com> References: <4F54F5F8.70105@redhat.com> <4F54F76B.70403@codemonkey.ws> <20120305143142.1a1a53e2@doriath.home> <20120305180958.GC20937@garlic.tlv.redhat.com> <4F550350.7030703@redhat.com> <4F55BE82.3030205@redhat.com> <20120306092427.35103975@doriath.home> <20120306131624.GC2240@garlic.redhat.com> <4F561661.7080805@codemonkey.ws> <20120306105342.5921e785@doriath.home> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20120306105342.5921e785@doriath.home> Subject: Re: [Qemu-devel] [PATCH v2 1/2] console: add hw_screen_dump_async List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Luiz Capitulino Cc: Avi Kivity , qemu-devel@nongnu.org, Anthony Liguori , Gerd Hoffmann On Tue, Mar 06, 2012 at 10:53:42AM -0300, Luiz Capitulino wrote: > On Tue, 06 Mar 2012 07:51:29 -0600 > Anthony Liguori wrote: > > > On 03/06/2012 07:16 AM, Alon Levy wrote: > > > On Tue, Mar 06, 2012 at 09:24:27AM -0300, Luiz Capitulino wrote: > > >> On Tue, 06 Mar 2012 08:36:34 +0100 > > >> Gerd Hoffmann wrote: > > >> > > >>> Hi, > > >>> > > >>>>> How would the parallel execution facility be opaque to the implementer? > > >>>>> screendump returns, screendump_async needs to pass a closure. You can > > >>>>> automatically generate any amount of code, but you can only have a > > >>>>> single function implementation with longjmp/coroutine, or having a > > >>>>> saparate thread per command but that would mean taking locks for > > >>>>> anything not trivial, which avoids the no-change-to-implementation. Is > > >>>>> this what you have in mind? > > >>>> > > >>>> It would not be opaque to the implementer. But it would avoid > > >>>> introducing new commands and events, instead we have a unified mechanism > > >>>> to signal completion. > > >>> > > >>> Ok. We have a async mechanism today: .mhandler.cmd_async = ... > > >>> > > >>> I know it has its problems like no cancelation and is deprecated and > > >>> all. But still: how about using it as interim until QAPI-based async > > >>> monitor support is ready? We could unbreak qxl screendumps without > > >>> having to introduce a new (but temporary!) screendump_async command + > > >>> completion event. > > >> > > >> There are a few problems here, but the blocking one is that a command > > >> can't go from sync to async. This is an incompatible change. > > >> > > >> If you mind adding the temporary command and if this issue is so rare > > >> that none can reproduce it, then I'd suggest to wait for 1.2. > > >> > > > > > > There are two options really: > > > 1. revert the patches that changed qxl screendump to save the ppm > > > before (possibly) updating the framebuffer. > > > 2. introduce a new command that is really async > > > > > > The third option, what Gerd proposes, doesn't break the blocking chain > > > going from the A, the dual purpose spice client and libvirt client, > > > through libvirt, qemu, spice and back to A. > > > > > > If no one can reproduce the block then it would seem 1 makes sense. > > > > So let's start with a reproducible test case that demonstrates the problem > > before we start introducing new commands then if there's doubt about the nature > > of the problem. > > Completely agree, I was going to suggest that too. > So cutting off a part of the email is a good way to win arguments? cool trick. I agree a reproducer is a good idea, but as I mentioned in the cut part, doing the update area without keeping the monitor waiting improves performance of the guest by letting it do io exits. Why is that a bad thing?