Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Michael Tokarev]

Hi!

It looks like this commit:

commit 87fa7f3e98a1310ef1ac1900e7ee7f9610a038bc
Author: Thomas Gleixner <tglx@linutronix.de>
Date:   Wed Jul 8 21:51:54 2020 +0200

     x86/kvm: Move context tracking where it belongs

     Context tracking for KVM happens way too early in the vcpu_run()
     code. Anything after guest_enter_irqoff() and before guest_exit_irqoff()
     cannot use RCU and should also be not instrumented.

     The current way of doing this covers way too much code. Move it closer to
     the actual vmenter/exit code.

broke kvm guest cpu time accounting - after this commit, when running
qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
elsewhere) is always 0.

I dunno why it happened, but it happened, and all kernels after 5.9
are affected by this.

This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.

Thanks,

/mjt

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Sean Christopherson]

On Tue, Apr 06, 2021, Michael Tokarev wrote:
> Hi!
> 
> It looks like this commit:
> 
> commit 87fa7f3e98a1310ef1ac1900e7ee7f9610a038bc
> Author: Thomas Gleixner <tglx@linutronix.de>
> Date:   Wed Jul 8 21:51:54 2020 +0200
> 
>     x86/kvm: Move context tracking where it belongs
> 
>     Context tracking for KVM happens way too early in the vcpu_run()
>     code. Anything after guest_enter_irqoff() and before guest_exit_irqoff()
>     cannot use RCU and should also be not instrumented.
> 
>     The current way of doing this covers way too much code. Move it closer to
>     the actual vmenter/exit code.
> 
> broke kvm guest cpu time accounting - after this commit, when running
> qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
> elsewhere) is always 0.
> 
> I dunno why it happened, but it happened, and all kernels after 5.9
> are affected by this.
> 
> This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.

Yes :-(

There's a bugzilla[1] and two proposed fixes[2][3].  I don't particularly like
either of the fixes, but an elegant solution hasn't presented itself.

Thomas/Paolo, can you please weigh in?

[1] https://bugzilla.kernel.org/show_bug.cgi?id=209831
[2] https://lkml.kernel.org/r/1617011036-11734-1-git-send-email-wanpengli@tencent.com
[3] https://lkml.kernel.org/r/20210206004218.312023-1-seanjc@google.com

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Thomas Gleixner]

On Tue, Apr 06 2021 at 21:47, Sean Christopherson wrote:
> On Tue, Apr 06, 2021, Michael Tokarev wrote:
>> broke kvm guest cpu time accounting - after this commit, when running
>> qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
>> elsewhere) is always 0.
>> 
>> I dunno why it happened, but it happened, and all kernels after 5.9
>> are affected by this.
>> 
>> This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.
>
> Yes :-(
>
> There's a bugzilla[1] and two proposed fixes[2][3].  I don't particularly like
> either of the fixes, but an elegant solution hasn't presented itself.
>
> Thomas/Paolo, can you please weigh in?
>
> [1] https://bugzilla.kernel.org/show_bug.cgi?id=209831
> [2] https://lkml.kernel.org/r/1617011036-11734-1-git-send-email-wanpengli@tencent.com
> [3] https://lkml.kernel.org/r/20210206004218.312023-1-seanjc@google.com

All of the solutions I looked at so far are ugly as hell. The problem is
that the accounting is plumbed into the context tracking and moving
context tracking around to a different place is just wrong.

I think the right solution is to seperate the time accounting logic out
from guest_enter/exit_irqoff() and have virt time specific helpers which
can be placed at the proper spots in kvm.

Thanks,

        tglx

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Wanpeng Li]

On Thu, 8 Apr 2021 at 21:19, Thomas Gleixner <tglx@linutronix.de> wrote:
>
> On Tue, Apr 06 2021 at 21:47, Sean Christopherson wrote:
> > On Tue, Apr 06, 2021, Michael Tokarev wrote:
> >> broke kvm guest cpu time accounting - after this commit, when running
> >> qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
> >> elsewhere) is always 0.
> >>
> >> I dunno why it happened, but it happened, and all kernels after 5.9
> >> are affected by this.
> >>
> >> This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.
> >
> > Yes :-(
> >
> > There's a bugzilla[1] and two proposed fixes[2][3].  I don't particularly like
> > either of the fixes, but an elegant solution hasn't presented itself.
> >
> > Thomas/Paolo, can you please weigh in?
> >
> > [1] https://bugzilla.kernel.org/show_bug.cgi?id=209831
> > [2] https://lkml.kernel.org/r/1617011036-11734-1-git-send-email-wanpengli@tencent.com
> > [3] https://lkml.kernel.org/r/20210206004218.312023-1-seanjc@google.com
>
> All of the solutions I looked at so far are ugly as hell. The problem is
> that the accounting is plumbed into the context tracking and moving
> context tracking around to a different place is just wrong.
>
> I think the right solution is to seperate the time accounting logic out
> from guest_enter/exit_irqoff() and have virt time specific helpers which
> can be placed at the proper spots in kvm.

Good suggestion, I will have a try. :)

    Wanpeng

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Wanpeng Li]

On Thu, 8 Apr 2021 at 21:19, Thomas Gleixner <tglx@linutronix.de> wrote:
>
> On Tue, Apr 06 2021 at 21:47, Sean Christopherson wrote:
> > On Tue, Apr 06, 2021, Michael Tokarev wrote:
> >> broke kvm guest cpu time accounting - after this commit, when running
> >> qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
> >> elsewhere) is always 0.
> >>
> >> I dunno why it happened, but it happened, and all kernels after 5.9
> >> are affected by this.
> >>
> >> This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.
> >
> > Yes :-(
> >
> > There's a bugzilla[1] and two proposed fixes[2][3].  I don't particularly like
> > either of the fixes, but an elegant solution hasn't presented itself.
> >
> > Thomas/Paolo, can you please weigh in?
> >
> > [1] https://bugzilla.kernel.org/show_bug.cgi?id=209831
> > [2] https://lkml.kernel.org/r/1617011036-11734-1-git-send-email-wanpengli@tencent.com
> > [3] https://lkml.kernel.org/r/20210206004218.312023-1-seanjc@google.com
>
> All of the solutions I looked at so far are ugly as hell. The problem is
> that the accounting is plumbed into the context tracking and moving
> context tracking around to a different place is just wrong.
>
> I think the right solution is to seperate the time accounting logic out
> from guest_enter/exit_irqoff() and have virt time specific helpers which
> can be placed at the proper spots in kvm.

For x86 part, how about something like below:

diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 48b396f3..7aeb724 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -3730,6 +3730,7 @@ static noinstr void svm_vcpu_enter_exit(struct
kvm_vcpu *vcpu)
     lockdep_hardirqs_on_prepare(CALLER_ADDR0);
     instrumentation_end();

+    account_guest_enter();
     guest_enter_irqoff();
     lockdep_hardirqs_on(CALLER_ADDR0);

@@ -3759,6 +3760,8 @@ static noinstr void svm_vcpu_enter_exit(struct
kvm_vcpu *vcpu)
      */
     lockdep_hardirqs_off(CALLER_ADDR0);
     guest_exit_irqoff();
+    if (vtime_accounting_enabled_this_cpu())
+        account_guest_exit();

     instrumentation_begin();
     trace_hardirqs_off_finish();
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index c05e6e2..5f6c30c 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -6617,6 +6617,7 @@ static noinstr void vmx_vcpu_enter_exit(struct
kvm_vcpu *vcpu,
     lockdep_hardirqs_on_prepare(CALLER_ADDR0);
     instrumentation_end();

+    account_guest_enter();
     guest_enter_irqoff();
     lockdep_hardirqs_on(CALLER_ADDR0);

@@ -6648,6 +6649,8 @@ static noinstr void vmx_vcpu_enter_exit(struct
kvm_vcpu *vcpu,
      */
     lockdep_hardirqs_off(CALLER_ADDR0);
     guest_exit_irqoff();
+    if (vtime_accounting_enabled_this_cpu())
+        account_guest_exit();

     instrumentation_begin();
     trace_hardirqs_off_finish();
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 16fb395..33422c0 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -9229,6 +9229,8 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
     ++vcpu->stat.exits;
     local_irq_disable();
     kvm_after_interrupt(vcpu);
+    if (!vtime_accounting_enabled_this_cpu())
+        account_guest_exit();

     if (lapic_in_kernel(vcpu)) {
         s64 delta = vcpu->arch.apic->lapic_timer.advance_expire_delta;
diff --git a/include/linux/context_tracking.h b/include/linux/context_tracking.h
index bceb064..ff70229 100644
--- a/include/linux/context_tracking.h
+++ b/include/linux/context_tracking.h
@@ -104,8 +104,7 @@ static inline void context_tracking_init(void) { }


 #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
-/* must be called with irqs disabled */
-static __always_inline void guest_enter_irqoff(void)
+static __always_inline void account_guest_enter(void)
 {
     instrumentation_begin();
     if (vtime_accounting_enabled_this_cpu())
@@ -113,7 +112,11 @@ static __always_inline void guest_enter_irqoff(void)
     else
         current->flags |= PF_VCPU;
     instrumentation_end();
+}

+/* must be called with irqs disabled */
+static __always_inline void guest_enter_irqoff(void)
+{
     if (context_tracking_enabled())
         __context_tracking_enter(CONTEXT_GUEST);

@@ -131,11 +134,8 @@ static __always_inline void guest_enter_irqoff(void)
     }
 }

-static __always_inline void guest_exit_irqoff(void)
+static __always_inline void account_guest_exit(void)
 {
-    if (context_tracking_enabled())
-        __context_tracking_exit(CONTEXT_GUEST);
-
     instrumentation_begin();
     if (vtime_accounting_enabled_this_cpu())
         vtime_guest_exit(current);
@@ -144,8 +144,14 @@ static __always_inline void guest_exit_irqoff(void)
     instrumentation_end();
 }

+static __always_inline void guest_exit_irqoff(void)
+{
+    if (context_tracking_enabled())
+        __context_tracking_exit(CONTEXT_GUEST);
+}
+
 #else
-static __always_inline void guest_enter_irqoff(void)
+static __always_inline void account_guest_enter(void)
 {
     /*
      * This is running in ioctl context so its safe
@@ -155,11 +161,17 @@ static __always_inline void guest_enter_irqoff(void)
     instrumentation_begin();
     vtime_account_kernel(current);
     current->flags |= PF_VCPU;
+    instrumentation_end();
+}
+
+static __always_inline void guest_enter_irqoff(void)
+{
+    instrumentation_begin();
     rcu_virt_note_context_switch(smp_processor_id());
     instrumentation_end();
 }

-static __always_inline void guest_exit_irqoff(void)
+static __always_inline void account_guest_exit(void)
 {
     instrumentation_begin();
     /* Flush the guest cputime we spent on the guest */
@@ -167,6 +179,11 @@ static __always_inline void guest_exit_irqoff(void)
     current->flags &= ~PF_VCPU;
     instrumentation_end();
 }
+
+static __always_inline void guest_exit_irqoff(void)
+{
+
+}
 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */

 static inline void guest_exit(void)

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Thomas Gleixner]

On Fri, Apr 09 2021 at 16:13, Wanpeng Li wrote:
> On Thu, 8 Apr 2021 at 21:19, Thomas Gleixner <tglx@linutronix.de> wrote:
>
> +    account_guest_enter();

This wants to move into the instrumentation_begin/end() section above.

>      guest_enter_irqoff();
>      lockdep_hardirqs_on(CALLER_ADDR0);
>
> @@ -3759,6 +3760,8 @@ static noinstr void svm_vcpu_enter_exit(struct
> kvm_vcpu *vcpu)
>       */
>      lockdep_hardirqs_off(CALLER_ADDR0);
>      guest_exit_irqoff();
> +    if (vtime_accounting_enabled_this_cpu())
> +        account_guest_exit();

This time below. Aside of that I'd suggest to have two inlines instead
of having the conditional here.

>
>  #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
> -/* must be called with irqs disabled */
> -static __always_inline void guest_enter_irqoff(void)
> +static __always_inline void account_guest_enter(void)
>  {
>      instrumentation_begin();
>      if (vtime_accounting_enabled_this_cpu())
> @@ -113,7 +112,11 @@ static __always_inline void guest_enter_irqoff(void)
>      else
>          current->flags |= PF_VCPU;
>      instrumentation_end();

If you move the invocation into the instrumentable section then this
instrumentation_begin/end() can be removed.

Something like the below +/- the obligatory bikeshed painting
vs. function names.

Thanks,

        tglx
---
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -3782,6 +3782,7 @@ static noinstr void svm_vcpu_enter_exit(
 	 * accordingly.
 	 */
 	instrumentation_begin();
+	vtime_account_guest_enter();
 	trace_hardirqs_on_prepare();
 	lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 	instrumentation_end();
@@ -3816,6 +3817,7 @@ static noinstr void svm_vcpu_enter_exit(
 
 	instrumentation_begin();
 	trace_hardirqs_off_finish();
+	vtime_account_guest_exit();
 	instrumentation_end();
 }
 
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -6655,6 +6655,7 @@ static noinstr void vmx_vcpu_enter_exit(
 	 * accordingly.
 	 */
 	instrumentation_begin();
+	vtime_account_guest_enter();
 	trace_hardirqs_on_prepare();
 	lockdep_hardirqs_on_prepare(CALLER_ADDR0);
 	instrumentation_end();
@@ -6693,6 +6694,7 @@ static noinstr void vmx_vcpu_enter_exit(
 
 	instrumentation_begin();
 	trace_hardirqs_off_finish();
+	vtime_account_guest_exit();
 	instrumentation_end();
 }
 
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -9200,6 +9200,7 @@ static int vcpu_enter_guest(struct kvm_v
 	++vcpu->stat.exits;
 	local_irq_disable();
 	kvm_after_interrupt(vcpu);
+	vcpu_account_guest_exit();
 
 	if (lapic_in_kernel(vcpu)) {
 		s64 delta = vcpu->arch.apic->lapic_timer.advance_expire_delta;
--- a/include/linux/context_tracking.h
+++ b/include/linux/context_tracking.h
@@ -107,13 +107,6 @@ static inline void context_tracking_init
 /* must be called with irqs disabled */
 static __always_inline void guest_enter_irqoff(void)
 {
-	instrumentation_begin();
-	if (vtime_accounting_enabled_this_cpu())
-		vtime_guest_enter(current);
-	else
-		current->flags |= PF_VCPU;
-	instrumentation_end();
-
 	if (context_tracking_enabled())
 		__context_tracking_enter(CONTEXT_GUEST);
 
@@ -135,37 +128,18 @@ static __always_inline void guest_exit_i
 {
 	if (context_tracking_enabled())
 		__context_tracking_exit(CONTEXT_GUEST);
-
-	instrumentation_begin();
-	if (vtime_accounting_enabled_this_cpu())
-		vtime_guest_exit(current);
-	else
-		current->flags &= ~PF_VCPU;
-	instrumentation_end();
 }
 
 #else
 static __always_inline void guest_enter_irqoff(void)
 {
-	/*
-	 * This is running in ioctl context so its safe
-	 * to assume that it's the stime pending cputime
-	 * to flush.
-	 */
 	instrumentation_begin();
-	vtime_account_kernel(current);
-	current->flags |= PF_VCPU;
 	rcu_virt_note_context_switch(smp_processor_id());
 	instrumentation_end();
 }
 
 static __always_inline void guest_exit_irqoff(void)
 {
-	instrumentation_begin();
-	/* Flush the guest cputime we spent on the guest */
-	vtime_account_kernel(current);
-	current->flags &= ~PF_VCPU;
-	instrumentation_end();
 }
 #endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */
 
@@ -178,4 +152,24 @@ static inline void guest_exit(void)
 	local_irq_restore(flags);
 }
 
+static __always_inline void vtime_account_guest_enter(void)
+{
+	if (vtime_accounting_enabled_this_cpu())
+		vtime_guest_enter(current);
+	else
+		current->flags |= PF_VCPU;
+}
+
+static __always_inline void vtime_account_guest_exit(void)
+{
+	if (vtime_accounting_enabled_this_cpu())
+		vtime_guest_exit(current);
+}
+
+static __always_inline void vcpu_account_guest_exit(void)
+{
+	if (!vtime_accounting_enabled_this_cpu())
+		current->flags &= ~PF_VCPU;
+}
+
 #endif

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Wanpeng Li]

On Wed, 7 Apr 2021 at 18:55, Michael Tokarev <mjt@tls.msk.ru> wrote:
>
> Hi!
>
> It looks like this commit:
>
> commit 87fa7f3e98a1310ef1ac1900e7ee7f9610a038bc
> Author: Thomas Gleixner <tglx@linutronix.de>
> Date:   Wed Jul 8 21:51:54 2020 +0200
>
>      x86/kvm: Move context tracking where it belongs
>
>      Context tracking for KVM happens way too early in the vcpu_run()
>      code. Anything after guest_enter_irqoff() and before guest_exit_irqoff()
>      cannot use RCU and should also be not instrumented.
>
>      The current way of doing this covers way too much code. Move it closer to
>      the actual vmenter/exit code.
>
> broke kvm guest cpu time accounting - after this commit, when running
> qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
> elsewhere) is always 0.
>
> I dunno why it happened, but it happened, and all kernels after 5.9
> are affected by this.
>
> This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.

Thanks for the report. I have a patch here which can fix it.
https://lore.kernel.org/kvm/1617011036-11734-1-git-send-email-wanpengli@tencent.com/

    Wanpeng

Re: Commit "x86/kvm: Move context tracking where it belongs" broke guest time accounting

[Wanpeng Li]

On Wed, 7 Apr 2021 at 18:55, Michael Tokarev <mjt@tls.msk.ru> wrote:
>
> Hi!
>
> It looks like this commit:
>
> commit 87fa7f3e98a1310ef1ac1900e7ee7f9610a038bc
> Author: Thomas Gleixner <tglx@linutronix.de>
> Date:   Wed Jul 8 21:51:54 2020 +0200
>
>      x86/kvm: Move context tracking where it belongs
>
>      Context tracking for KVM happens way too early in the vcpu_run()
>      code. Anything after guest_enter_irqoff() and before guest_exit_irqoff()
>      cannot use RCU and should also be not instrumented.
>
>      The current way of doing this covers way too much code. Move it closer to
>      the actual vmenter/exit code.
>
> broke kvm guest cpu time accounting - after this commit, when running
> qemu-system-x86_64 -enable-kvm, the guest time (in /proc/stat and
> elsewhere) is always 0.
>
> I dunno why it happened, but it happened, and all kernels after 5.9
> are affected by this.
>
> This commit is found in a (painful) git bisect between kernel 5.8 and 5.10.

Hi Michael,

Please have a try.
https://lore.kernel.org/kvm/1618298169-3831-1-git-send-email-wanpengli@tencent.com/

    Wanpeng