From: Zachary Amsden Subject: Re: [RFC, PATCH 17/24] i386 Vmi msr patch Date: Tue, 14 Mar 2006 08:32:56 -0800 Message-ID: <4416F038.90707@vmware.com> References: <200603131812.k2DICGJE005747@zach-dev.vmware.com> <200603141723.54365.ak@suse.de> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <200603141723.54365.ak@suse.de> Sender: linux-kernel-owner@vger.kernel.org To: Andi Kleen Cc: virtualization@lists.osdl.org, Linus Torvalds , Linux Kernel Mailing List , Xen-devel , Andrew Morton , Dan Hecht , Dan Arai , Anne Holler , Pratap Subrahmanyam , Christopher Li , Joshua LeVasseur , Chris Wright , Rik Van Riel , Jyothy Reddy , Jack Lo , Kip Macy , Jan Beulich , Ky Srinivasan , Wim Coekaerts , Leendert van Doorn List-Id: xen-devel@lists.xenproject.org Andi Kleen wrote: > On Monday 13 March 2006 19:12, Zachary Amsden wrote: > >> Fairly straightforward code motion of MSR / TSC / PMC accessors >> to the sub-arch level. Note that rdmsr/wrmsr_safe functions are >> not moved; Linux relies on the fault behavior here in the event >> that certain MSRs are not supported on hardware, and combining >> this with a VMI wrapper is overly complicated. The instructions >> are virtualizable with trap and emulate, not on critical code >> paths, and only used as part of the MSR /proc device, which is >> highly sketchy to use inside a virtual machine, but must be >> allowed as part of the compile, since it is useful on native. >> > > I'm not aware of any MSR access being on a critical code > path on a 32bit kernel. > There aren't really any. There are some unexpected ones - such as setting SYSENTER_CS during a context switch, but only if leaving v8086 mode, which isn't a common case. But most importantly, the MSR functions were challenging to get correct, because they combine two novel elements - 64 bit values, as well as non-C calling conventions. They were actually some of the first functions I inlined, because I knew there would be problems, and the solutions would yield more powerful inlining macros. > And I don't think it's a good idea to virtualize the TSC > without CPU support. > We currently don't support configurations without a TSC. But we're not trying to virtualize the TSC without CPU support. It is possible. But I have no idea _why_ you would want to do such a thing. Zach