From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Michael S. Tsirkin" Subject: Re: [PATCHv6 6/8] kvm: only sync when attention bits set Date: Thu, 14 Jun 2012 11:04:05 +0300 Message-ID: <20120614080404.GB26869@redhat.com> References: <56ee14064ef307b7075f6977122781456cab5165.1338474301.git.mst@redhat.com> <20120612222748.GB1973@amt.cnet> <20120613081923.GA16162@redhat.com> <20120613083506.GA17190@redhat.com> <20120613205336.GB19290@amt.cnet> <20120613210423.GA24932@redhat.com> <20120613233846.GA23955@amt.cnet> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: x86@kernel.org, kvm@vger.kernel.org, Ingo Molnar , "H. Peter Anvin" , Avi Kivity , gleb@redhat.com, Linus Torvalds , linux-kernel@vger.kernel.org, Thomas Gleixner To: Marcelo Tosatti Return-path: Content-Disposition: inline In-Reply-To: <20120613233846.GA23955@amt.cnet> Sender: linux-kernel-owner@vger.kernel.org List-Id: kvm.vger.kernel.org On Wed, Jun 13, 2012 at 08:38:46PM -0300, Marcelo Tosatti wrote: > On Thu, Jun 14, 2012 at 12:04:23AM +0300, Michael S. Tsirkin wrote: > > On Wed, Jun 13, 2012 at 05:53:36PM -0300, Marcelo Tosatti wrote: > > > On Wed, Jun 13, 2012 at 11:35:07AM +0300, Michael S. Tsirkin wrot= e: > > > > On Wed, Jun 13, 2012 at 11:19:24AM +0300, Michael S. Tsirkin wr= ote: > > > > > On Tue, Jun 12, 2012 at 07:27:48PM -0300, Marcelo Tosatti wro= te: > > > > > > On Sun, Jun 03, 2012 at 10:28:29AM +0300, Michael S. Tsirki= n wrote: > > > > > > > Commit eb0dc6d0368072236dcd086d7fdc17fd3c4574d4 introduce= d apic > > > > > > > attention bitmask but kvm still syncs lapic unconditional= ly. > > > > > > > As that commit suggested and in anticipation of adding mo= re attention > > > > > > > bits, only sync lapic if(apic_attention). > > > > > > >=20 > > > > > > > Signed-off-by: Michael S. Tsirkin > > > > > > > --- > > > > > > > arch/x86/kvm/x86.c | 3 ++- > > > > > > > 1 files changed, 2 insertions(+), 1 deletions(-) > > > > > > >=20 > > > > > > > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > > > > > > > index be6d549..2f70861 100644 > > > > > > > --- a/arch/x86/kvm/x86.c > > > > > > > +++ b/arch/x86/kvm/x86.c > > > > > > > @@ -5388,7 +5388,8 @@ static int vcpu_enter_guest(struct = kvm_vcpu *vcpu) > > > > > > > if (unlikely(vcpu->arch.tsc_always_catchup)) > > > > > > > kvm_make_request(KVM_REQ_CLOCK_UPDATE, vcpu); > > > > > > > =20 > > > > > > > - kvm_lapic_sync_from_vapic(vcpu); > > > > > > > + if (unlikely(vcpu->arch.apic_attention)) > > > > > > > + kvm_lapic_sync_from_vapic(vcpu); > > > > > >=20 > > > > > > void kvm_lapic_sync_from_vapic(struct kvm_vcpu *vcpu) > > > > > > { > > > > > > u32 data; > > > > > > void *vapic; > > > > > >=20 > > > > > > if (!test_bit(KVM_APIC_CHECK_VAPIC, &vcpu->arch.api= c_attention)) > > > > > > return; > > > > > >=20 > > > > > > Please use unlikely more carefully, when a gain is measurea= ble: > > > > > > http://lwn.net/Articles/420019/ > > > > >=20 > > > > > Do we have to measure every single thing? > > > > > Sometimes it's obvious: vapic is slow path, isn't it? > > > >=20 > > > > Just to clarify the question: I think it's obvious this conditi= on is > > > > false more often than true. By how much, depends on the workloa= d. > > > > Do you think this is enough to tag this unlikely? > > >=20 > > > Depends whether your processor supports flexpriority or not. I do= n't > > > want to argue in favour/against the particular instance > > >=20 > > > GCC docs: > > >=20 > > > " > > > =E2=80=94 Built-in Function: long __builtin_expect (long exp, lon= g c) > > >=20 > > > You may use __builtin_expect to provide the compiler with bra= nch > > > prediction information. In general, you should prefer to use actu= al > > > profile feedback for this (-fprofile-arcs), as programmers are > > > notoriously bad at predicting how their programs actually perform= =2E > > > However, there are applications in which this data is hard to col= lect. > > > " > > >=20 > > > Lately half of branches in your patches are annotated. > >=20 > > So if I instrument and show that branch is almost never taken that = is enough? > > This citation does not require measuring the perf impact. >=20 > Without flexpriority its always taken. OK I will remove this one and keep around the unlikely for the cases that my measurements show occur in less than 1% of cases.