From: "Jan Beulich" Subject: Re: [PATCH] enable port accesses with (almost) full register context Date: Tue, 12 Sep 2006 11:32:23 +0100 Message-ID: <4506A8D7.76E4.0078.0@novell.com> References: <450693EC.76E4.0078.0@novell.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: Content-Disposition: inline List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: Keir Fraser Cc: xen-devel@lists.xensource.com List-Id: xen-devel@lists.xenproject.org >>> Keir Fraser 12.09.06 11:50 >>> >On 12/9/06 10:03, "Jan Beulich" wrote: > >> Again, I'm using self-modifying code there (to store the port number, as I >> can't reliably use %dx for it if the original instruction happened to be one >> with immediate operand, and %edx/%rdx happens to carry relevant data >> for the SMI handler), which is what needs synchronization. > >Ok, I see. I think it would be neater to build the code on the stack, or >some other per-cpu area, and avoid the synchronisation. We have no plans to >use the PAGE_NX flag in Xen itself, and x86/64 already has stack >trampolines. Perhaps the register save/restore code could be tidied too, >since it's not performance critical. It's not at all uniform like I'd >expect, with those interleaved push/pop/mov instructions. How about >something more like: > pushad; call restore_guest_regs; ; popad >Where restore_guest_regs takes a regparm, and (obviously) restores the >regparm register last. I'd only do it as a call because it'd be ugly to >dynamically build that amount of code. Hm, I don't like this on-the-fly building of code very much, and I also don't like writing assembly code that can obviously written to perform better. Also, on 64-bits the code wouldn't look so much nicer since there's no {push,pop}ad. But certainly, if you refuse to take the patch without changing that... >I'm not sure about the full extent of the interface changes either. How >about we add a new sysctl for specifying ports which need 'direct >execution'. It makes sense to make it a sysctl because this is a property of >the I/O port (or assumptions about it encoded in the platform firmware) >rather than a per-domain issue, or something that I think should be visible >at the physdev_op interface. > >We'd test the per-port direct-execution flag for any port access by any >domain. After all, the only reason we don't use the new code for *all* port >accesses is concern about performance. I think calling this 'direct >execution' versus 'emulation' at the interface is fair -- even though we >emulate in all cases, in the former case it will be Xen's responsibility to >do all that is necessary to make it appear to the BIOS that the instruction >was executed directly, as when running natively. That sounds right (and better than the current way). I'll do that change, though I guess I'd still not call it direct execution. Jan