From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jeremy Fitzhardinge Subject: Re: Xen with dom0 pvops on ultra-recent "git tip" kernel on x86_64 Date: Fri, 16 Jan 2009 09:52:33 -0800 Message-ID: <4970C961.9090808@goop.org> References: <49690103.5020706@wol.de> <1231619339.5616.15.camel@leto.intern.saout.de> <20090110203706.GW15052@edu.joroinen.fi> <1231622029.4778.1.camel@leto.intern.saout.de> <1231683478.4809.43.camel@leto.intern.saout.de> <1231685561.4809.50.camel@leto.intern.saout.de> <1231691732.5342.3.camel@leto.intern.saout.de> <1231769868.4838.4.camel@leto.intern.saout.de> <1231778480.4742.3.camel@leto.intern.saout.de> <1231838888.4823.5.camel@leto.intern.saout.de> <20090113153302.GZ15052@edu.joroinen.fi> <1231864918.5257.6.camel@leto.intern.saout.de> <1231865847.5257.8.camel@leto.intern.saout.de> <4970C2A7.8010702@goop.org> <1232127636.4946.28.camel@leto.intern.saout.de> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-15; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <1232127636.4946.28.camel@leto.intern.saout.de> List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: Christophe Saout Cc: xen-devel List-Id: xen-devel@lists.xenproject.org Christophe Saout wrote: > It's not really exciting stuff, mostly some cleanups which were > moving stuff around and such (and some iotlb dma things which were not > entirely trivial but pretty straight-forward). > Yeah the iotlb stuff was in a bit of a flux, but I think I understand how it ended up. >> I haven't been brave enough to even attempt starting X; there's a lot of >> work to be done if you want to do anything beyond dumb framebuffer. >> >> Anyway, in this case, I definitely haven't added this call, so it will >> need to be done somewhere. Does it need to be called once at startup, >> or on every context switch? There may well be a suitable place to hook >> this into already, but we can add a new pvop if its really needed. >> > > Looking at the original Xen kernel the call has to be made on every > context switch if the io permissions change between the two threads. > In the past days I have sort of started trying to put some new paravirt > call together in the various places. This unfortunately also meant > changing a bit of the native stuff to pass through a unified > "set_io_bitmap" call. I haven't checked if that works, but at least the > Xen versions of the call doesn't get X up and running. Now I get a > black screen and dead machine. I guess I will have to do some > debugging. ;-) > There are several pvops calls made already during context switch, so I wouldn't be surprised if you can sneak it into one of those (load_sp0 looks like a candidate, since it's already passed the tss). Just make sure the hypercall gets batched with the rest of the context-switch multicall so that the latency doesn't get worse. > If you're interested, I can send you what I have so far, but I guess you > have more pressing things to do first. :-) > Yeah, I'm working on rebasing the patch queue at the moment, then I'll try to catch up on the various bugs/observations that people have reported. J