From: Radim =?utf-8?B?S3LEjW3DocWZ?= Subject: Re: [PATCH] kvm: x86: svm: remove SVM_EXIT_READ_CR* intercepts Date: Mon, 16 Mar 2015 18:34:59 +0100 Message-ID: <20150316173459.GB14216@potion.brq.redhat.com> References: <20150312201746.2953.11716.stgit@joelvm2.amd.com> <20150312212002.GA1711@potion.brq.redhat.com> <550701C5.8050603@amd.com> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: Gleb Natapov , Paolo Bonzini , kvm@vger.kernel.org, Joerg Roedel , Borislav Petkov , linux-kernel@vger.kernel.org, David Kaplan To: Joel Schopp Return-path: Content-Disposition: inline In-Reply-To: <550701C5.8050603@amd.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org 2015-03-16 11:16-0500, Joel Schopp: > On 03/12/2015 04:20 PM, Radim Kr=C4=8Dm=C3=A1=C5=99 wrote: > > 2015-03-12 15:17-0500, Joel Schopp: > >> There isn't really a valid reason for kvm to intercept cr* reads > >> on svm hardware. The current kvm code just ends up returning > >> the register > > There is no need to intercept CR* if the value that the guest shoul= d see > > is equal to what we set there, but that is not always the case: > > - CR0 might differ from what the guest should see because of lazy f= pu > Based on our previous conversations I understand why we have to trap = the > write to the CR0 ts bit for lazy fpu, but don't understand why that > should affect a read. KVM keeps one CR0 with guest's state (svm.vcpu.arch.cr0) and a second one that is loaded to hardware CR0 on VMRUN (svm.vmcb->save.cr0); these two might not match. If we didn't intercept read, it would return hardware CR0, so the guest could do CLTS (change svm.vcpu.arch.cr0) and read CR0.TS =3D 1, because= of lazy FPU. Correct emulation is what we want. > > CR2 and CR8 already aren't intercepted, so it looks like only CR0 a= nd > > CR4 could use some optimizations. > I'll send out a v2 with these less aggressive optimizations. Thanks.