| Message ID | 1457729240-3846-2-git-send-email-dmatlack@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Mar 11, 2016 at 12:47 PM, David Matlack <dmatlack@google.com> wrote: > From: Eric Northup <digitaleric@google.com> > > Add a percpu boolean, tracking whether a KVM vCPU is running on the > host CPU. KVM will set and clear it as it loads/unloads guest XCR0. > (Note that the rest of the guest FPU load/restore is safe, because > kvm_load_guest_fpu and kvm_put_guest_fpu call __kernel_fpu_begin() > and __kernel_fpu_end(), respectively.) irq_fpu_usable() will then > also check for this percpu boolean. Is this better than just always keeping the host's XCR0 loaded outside if the KVM interrupts-disabled region? --Andy -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Fri, Mar 11, 2016 at 1:14 PM, Andy Lutomirski <luto@amacapital.net> wrote: > > On Fri, Mar 11, 2016 at 12:47 PM, David Matlack <dmatlack@google.com> wrote: > > From: Eric Northup <digitaleric@google.com> > > > > Add a percpu boolean, tracking whether a KVM vCPU is running on the > > host CPU. KVM will set and clear it as it loads/unloads guest XCR0. > > (Note that the rest of the guest FPU load/restore is safe, because > > kvm_load_guest_fpu and kvm_put_guest_fpu call __kernel_fpu_begin() > > and __kernel_fpu_end(), respectively.) irq_fpu_usable() will then > > also check for this percpu boolean. > > Is this better than just always keeping the host's XCR0 loaded outside > if the KVM interrupts-disabled region? Probably not. AFAICT KVM does not rely on it being loaded outside that region. xsetbv isn't insanely expensive, is it? Maybe to minimize the time spent with interrupts disabled it was put outside. I do like that your solution would be contained to KVM. > > --Andy -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 11/03/2016 22:33, David Matlack wrote: > > Is this better than just always keeping the host's XCR0 loaded outside > > if the KVM interrupts-disabled region? > > Probably not. AFAICT KVM does not rely on it being loaded outside that > region. xsetbv isn't insanely expensive, is it? Maybe to minimize the > time spent with interrupts disabled it was put outside. > > I do like that your solution would be contained to KVM. I agree with Andy. We do want a fix for recent kernels because of the !eager_fpu case that Guangrong mentioned. Paolo ps: while Andy is planning to kill lazy FPU, I want to benchmark it with KVM... Remember that with a single pre-xsave host in your cluster, your virt management might happily default your VMs to a Westmere or Nehalem CPU model. GCC might be a pretty good testbench for this (e.g. a kernel compile with very high make -j), because outside of the lexer (which plays SIMD games) it never uses the FPU. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Mon, Mar 14, 2016 at 6:17 AM, Paolo Bonzini <pbonzini@redhat.com> wrote: > > > On 11/03/2016 22:33, David Matlack wrote: >> > Is this better than just always keeping the host's XCR0 loaded outside >> > if the KVM interrupts-disabled region? >> >> Probably not. AFAICT KVM does not rely on it being loaded outside that >> region. xsetbv isn't insanely expensive, is it? Maybe to minimize the >> time spent with interrupts disabled it was put outside. >> >> I do like that your solution would be contained to KVM. > > I agree with Andy. We do want a fix for recent kernels because of the > !eager_fpu case that Guangrong mentioned. > > Paolo > > ps: while Andy is planning to kill lazy FPU, I want to benchmark it with > KVM... Remember that with a single pre-xsave host in your cluster, your > virt management might happily default your VMs to a Westmere or Nehalem > CPU model. GCC might be a pretty good testbench for this (e.g. a kernel > compile with very high make -j), because outside of the lexer (which > plays SIMD games) it never uses the FPU. Aren't pre-xsave CPUs really, really old? A brief search suggests that Intel Core added it somewhere in the middle of the cycle. For pre-xsave, it could indeed hurt performance a tiny bit under workloads that use the FPU and then stop completely because the xsaveopt and init optimizations aren't available. But even that is probably a very small effect, especially because pre-xsave CPUs have smaller FPU state sizes.
On 15/03/2016 19:27, Andy Lutomirski wrote: > On Mon, Mar 14, 2016 at 6:17 AM, Paolo Bonzini <pbonzini@redhat.com> wrote: >> >> >> On 11/03/2016 22:33, David Matlack wrote: >>>> Is this better than just always keeping the host's XCR0 loaded outside >>>> if the KVM interrupts-disabled region? >>> >>> Probably not. AFAICT KVM does not rely on it being loaded outside that >>> region. xsetbv isn't insanely expensive, is it? Maybe to minimize the >>> time spent with interrupts disabled it was put outside. >>> >>> I do like that your solution would be contained to KVM. >> >> I agree with Andy. We do want a fix for recent kernels because of the >> !eager_fpu case that Guangrong mentioned. >> >> Paolo >> >> ps: while Andy is planning to kill lazy FPU, I want to benchmark it with >> KVM... Remember that with a single pre-xsave host in your cluster, your >> virt management might happily default your VMs to a Westmere or Nehalem >> CPU model. GCC might be a pretty good testbench for this (e.g. a kernel >> compile with very high make -j), because outside of the lexer (which >> plays SIMD games) it never uses the FPU. > > Aren't pre-xsave CPUs really, really old? A brief search suggests > that Intel Core added it somewhere in the middle of the cycle. I am fairly sure it was added in Sandy Bridge, together with AVX. But what really matters for eager FPU is not xsave, it's xsaveopt, and I think AMD has never even produced a microprocessor that supports it. > For pre-xsave, it could indeed hurt performance a tiny bit under > workloads that use the FPU and then stop completely because the > xsaveopt and init optimizations aren't available. But even that is > probably a very small effect, especially because pre-xsave CPUs have > smaller FPU state sizes. It's still a few cache lines. Benchmarks will tell. Paolo -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 03/16/2016 03:32 AM, Paolo Bonzini wrote: > > > On 15/03/2016 19:27, Andy Lutomirski wrote: >> On Mon, Mar 14, 2016 at 6:17 AM, Paolo Bonzini <pbonzini@redhat.com> wrote: >>> >>> >>> On 11/03/2016 22:33, David Matlack wrote: >>>>> Is this better than just always keeping the host's XCR0 loaded outside >>>>> if the KVM interrupts-disabled region? >>>> >>>> Probably not. AFAICT KVM does not rely on it being loaded outside that >>>> region. xsetbv isn't insanely expensive, is it? Maybe to minimize the >>>> time spent with interrupts disabled it was put outside. >>>> >>>> I do like that your solution would be contained to KVM. >>> >>> I agree with Andy. We do want a fix for recent kernels because of the >>> !eager_fpu case that Guangrong mentioned. Relying on interrupt is not easy as XCR0 can not be automatically saved/loaded by VMCS... Once interrupt happens, it will use guest's XCR0 anyway. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 16/03/2016 04:55, Xiao Guangrong wrote: >>>>> >>>>> Probably not. AFAICT KVM does not rely on it being loaded outside that >>>>> region. xsetbv isn't insanely expensive, is it? Maybe to minimize the >>>>> time spent with interrupts disabled it was put outside. >>>>> >>>>> I do like that your solution would be contained to KVM. >>>> >>>> I agree with Andy. We do want a fix for recent kernels because of the >>>> !eager_fpu case that Guangrong mentioned. > > Relying on interrupt is not easy as XCR0 can not be automatically > saved/loaded by VMCS... Once interrupt happens, it will use guest's XCR0 anyway. Right, that's why an xsetbv while interrupts are disabled is appealing. Paolo -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/arch/x86/include/asm/i387.h b/arch/x86/include/asm/i387.h index ed8089d..ca2c173 100644 --- a/arch/x86/include/asm/i387.h +++ b/arch/x86/include/asm/i387.h @@ -14,6 +14,7 @@ #include <linux/sched.h> #include <linux/hardirq.h> +#include <linux/percpu.h> struct pt_regs; struct user_i387_struct; @@ -25,6 +26,8 @@ extern void math_state_restore(void); extern bool irq_fpu_usable(void); +DECLARE_PER_CPU(bool, kvm_xcr0_loaded); + /* * Careful: __kernel_fpu_begin/end() must be called with preempt disabled * and they don't touch the preempt state on their own. diff --git a/arch/x86/kernel/i387.c b/arch/x86/kernel/i387.c index b627746..9015828 100644 --- a/arch/x86/kernel/i387.c +++ b/arch/x86/kernel/i387.c @@ -19,6 +19,9 @@ #include <asm/fpu-internal.h> #include <asm/user.h> +DEFINE_PER_CPU(bool, kvm_xcr0_loaded); +EXPORT_PER_CPU_SYMBOL(kvm_xcr0_loaded); + /* * Were we in an interrupt that interrupted kernel mode? * @@ -33,8 +36,11 @@ */ static inline bool interrupted_kernel_fpu_idle(void) { - if (use_eager_fpu()) - return __thread_has_fpu(current); + if (use_eager_fpu()) { + /* Preempt already disabled, safe to read percpu. */ + return __thread_has_fpu(current) && + !__this_cpu_read(kvm_xcr0_loaded); + } return !__thread_has_fpu(current) && (read_cr0() & X86_CR0_TS); diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c index d21bce5..f0ba7a1 100644 --- a/arch/x86/kvm/x86.c +++ b/arch/x86/kvm/x86.c @@ -557,8 +557,10 @@ EXPORT_SYMBOL_GPL(kvm_lmsw); static void kvm_load_guest_xcr0(struct kvm_vcpu *vcpu) { + BUG_ON(this_cpu_read(kvm_xcr0_loaded) != vcpu->guest_xcr0_loaded); if (kvm_read_cr4_bits(vcpu, X86_CR4_OSXSAVE) && !vcpu->guest_xcr0_loaded) { + this_cpu_write(kvm_xcr0_loaded, 1); /* kvm_set_xcr() also depends on this */ xsetbv(XCR_XFEATURE_ENABLED_MASK, vcpu->arch.xcr0); vcpu->guest_xcr0_loaded = 1; @@ -571,7 +573,9 @@ static void kvm_put_guest_xcr0(struct kvm_vcpu *vcpu) if (vcpu->arch.xcr0 != host_xcr0) xsetbv(XCR_XFEATURE_ENABLED_MASK, host_xcr0); vcpu->guest_xcr0_loaded = 0; + this_cpu_write(kvm_xcr0_loaded, 0); } + BUG_ON(this_cpu_read(kvm_xcr0_loaded) != vcpu->guest_xcr0_loaded); } int __kvm_set_xcr(struct kvm_vcpu *vcpu, u32 index, u64 xcr)
From: Eric Northup <digitaleric@google.com> Add a percpu boolean, tracking whether a KVM vCPU is running on the host CPU. KVM will set and clear it as it loads/unloads guest XCR0. (Note that the rest of the guest FPU load/restore is safe, because kvm_load_guest_fpu and kvm_put_guest_fpu call __kernel_fpu_begin() and __kernel_fpu_end(), respectively.) irq_fpu_usable() will then also check for this percpu boolean. --- arch/x86/include/asm/i387.h | 3 +++ arch/x86/kernel/i387.c | 10 ++++++++-- arch/x86/kvm/x86.c | 4 ++++ 3 files changed, 15 insertions(+), 2 deletions(-)