diff mbox series

[v2,28/43] KVM: VMX: Remove vCPU from PI wakeup list before updating PID.NV

Message ID 20211009021236.4122790-29-seanjc@google.com
State New
Headers show
Series KVM: Halt-polling and x86 APICv overhaul | expand

Commit Message

Sean Christopherson Oct. 9, 2021, 2:12 a.m. UTC
Remove the vCPU from the wakeup list before updating the notification
vector in the posted interrupt post-block helper.  There is no need to
wake the current vCPU as it is by definition not blocking.  Practically
speaking this is a nop as it only shaves a few meager cycles in the
unlikely case that the vCPU was migrated and the previous pCPU gets a
wakeup IRQ right before PID.NV is updated.  The real motivation is to
allow for more readable code in the future, when post-block is merged
with vmx_vcpu_pi_load(), at which point removal from the list will be
conditional on the old notification vector.

Opportunistically add comments to document why KVM has a per-CPU spinlock
that, at first glance, appears to be taken only on the owning CPU.
Explicitly call out that the spinlock must be taken with IRQs disabled, a
detail that was "lost" when KVM switched from spin_lock_irqsave() to
spin_lock(), with IRQs disabled for the entirety of the relevant path.

Signed-off-by: Sean Christopherson <seanjc@google.com>
---
 arch/x86/kvm/vmx/posted_intr.c | 49 +++++++++++++++++++++++-----------
 1 file changed, 33 insertions(+), 16 deletions(-)

Comments

Maxim Levitsky Oct. 28, 2021, 12:53 p.m. UTC | #1
On Fri, 2021-10-08 at 19:12 -0700, Sean Christopherson wrote:
> Remove the vCPU from the wakeup list before updating the notification
> vector in the posted interrupt post-block helper.  There is no need to
> wake the current vCPU as it is by definition not blocking.  Practically
> speaking this is a nop as it only shaves a few meager cycles in the
> unlikely case that the vCPU was migrated and the previous pCPU gets a
> wakeup IRQ right before PID.NV is updated.  The real motivation is to
> allow for more readable code in the future, when post-block is merged
> with vmx_vcpu_pi_load(), at which point removal from the list will be
> conditional on the old notification vector.
> 
> Opportunistically add comments to document why KVM has a per-CPU spinlock
> that, at first glance, appears to be taken only on the owning CPU.
> Explicitly call out that the spinlock must be taken with IRQs disabled, a
> detail that was "lost" when KVM switched from spin_lock_irqsave() to
> spin_lock(), with IRQs disabled for the entirety of the relevant path.
> 
> Signed-off-by: Sean Christopherson <seanjc@google.com>
> ---
>  arch/x86/kvm/vmx/posted_intr.c | 49 +++++++++++++++++++++++-----------
>  1 file changed, 33 insertions(+), 16 deletions(-)
> 
> diff --git a/arch/x86/kvm/vmx/posted_intr.c b/arch/x86/kvm/vmx/posted_intr.c
> index 2b2206339174..901b7a5f7777 100644
> --- a/arch/x86/kvm/vmx/posted_intr.c
> +++ b/arch/x86/kvm/vmx/posted_intr.c
> @@ -10,10 +10,22 @@
>  #include "vmx.h"
>  
>  /*
> - * We maintain a per-CPU linked-list of vCPU, so in wakeup_handler() we
> - * can find which vCPU should be waken up.
> + * Maintain a per-CPU list of vCPUs that need to be awakened by wakeup_handler()
Nit: While at it, it would be nice to rename this to pi_wakeup_hanlder() so that it can be more easilly
found.


> + * when a WAKEUP_VECTOR interrupted is posted.  vCPUs are added to the list when
> + * the vCPU is scheduled out and is blocking (e.g. in HLT) with IRQs enabled.
s/interrupted/interrupt ?

Isn't that comment incorrect? As I see, the PI hardware is setup to use the WAKEUP_VECTOR
when vcpu blocks (in pi_pre_block) and then that vcpu is added to the list.
The pi_wakeup_hanlder just goes over the list and wakes up all vcpus on the lsit.


> + * The vCPUs posted interrupt descriptor is updated at the same time to set its
> + * notification vector to WAKEUP_VECTOR, so that posted interrupt from devices
> + * wake the target vCPUs.  vCPUs are removed from the list and the notification
> + * vector is reset when the vCPU is scheduled in.
>   */
>  static DEFINE_PER_CPU(struct list_head, blocked_vcpu_on_cpu);
Also while at it, why not to rename this to 'blocked_vcpu_list'?
to explain that this is list of blocked vcpus. Its a per-cpu variable
so 'on_cpu' suffix isn't needed IMHO.

> +/*
> + * Protect the per-CPU list with a per-CPU spinlock to handle task migration.
> + * When a blocking vCPU is awakened _and_ migrated to a different pCPU, the
> + * ->sched_in() path will need to take the vCPU off the list of the _previous_
> + * CPU.  IRQs must be disabled when taking this lock, otherwise deadlock will
> + * occur if a wakeup IRQ arrives and attempts to acquire the lock.
> + */
>  static DEFINE_PER_CPU(spinlock_t, blocked_vcpu_on_cpu_lock);
>  
>  static inline struct pi_desc *vcpu_to_pi_desc(struct kvm_vcpu *vcpu)
> @@ -101,23 +113,28 @@ static void __pi_post_block(struct kvm_vcpu *vcpu)
>  	WARN(pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR,
>  	     "Wakeup handler not enabled while the vCPU was blocking");
>  
> -	dest = cpu_physical_id(vcpu->cpu);
> -	if (!x2apic_mode)
> -		dest = (dest << 8) & 0xFF00;
> -
> -	do {
> -		old.control = new.control = READ_ONCE(pi_desc->control);
> -
> -		new.ndst = dest;
> -
> -		/* set 'NV' to 'notification vector' */
> -		new.nv = POSTED_INTR_VECTOR;
> -	} while (cmpxchg64(&pi_desc->control, old.control,
> -			   new.control) != old.control);
> -
> +	/*
> +	 * Remove the vCPU from the wakeup list of the _previous_ pCPU, which
> +	 * will not be the same as the current pCPU if the task was migrated.
> +	 */
>  	spin_lock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
>  	list_del(&vcpu->blocked_vcpu_list);
>  	spin_unlock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> +
> +	dest = cpu_physical_id(vcpu->cpu);
> +	if (!x2apic_mode)
> +		dest = (dest << 8) & 0xFF00;
It would be nice to have a function for this, this appears in this file twice.
Maybe there is a function already somewhere?


> +
> +	do {
> +		old.control = new.control = READ_ONCE(pi_desc->control);
> +
> +		new.ndst = dest;
> +
> +		/* set 'NV' to 'notification vector' */
> +		new.nv = POSTED_INTR_VECTOR;
> +	} while (cmpxchg64(&pi_desc->control, old.control,
> +			   new.control) != old.control);
> +
>  	vcpu->pre_pcpu = -1;
>  }
>  

Best regards,
	Maxim Levitsky
Sean Christopherson Oct. 28, 2021, 5:19 p.m. UTC | #2
On Thu, Oct 28, 2021, Maxim Levitsky wrote:
> On Fri, 2021-10-08 at 19:12 -0700, Sean Christopherson wrote:
> > Remove the vCPU from the wakeup list before updating the notification
> > vector in the posted interrupt post-block helper.  There is no need to
> > wake the current vCPU as it is by definition not blocking.  Practically
> > speaking this is a nop as it only shaves a few meager cycles in the
> > unlikely case that the vCPU was migrated and the previous pCPU gets a
> > wakeup IRQ right before PID.NV is updated.  The real motivation is to
> > allow for more readable code in the future, when post-block is merged
> > with vmx_vcpu_pi_load(), at which point removal from the list will be
> > conditional on the old notification vector.
> > 
> > Opportunistically add comments to document why KVM has a per-CPU spinlock
> > that, at first glance, appears to be taken only on the owning CPU.
> > Explicitly call out that the spinlock must be taken with IRQs disabled, a
> > detail that was "lost" when KVM switched from spin_lock_irqsave() to
> > spin_lock(), with IRQs disabled for the entirety of the relevant path.
> > 
> > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > ---
> >  arch/x86/kvm/vmx/posted_intr.c | 49 +++++++++++++++++++++++-----------
> >  1 file changed, 33 insertions(+), 16 deletions(-)
> > 
> > diff --git a/arch/x86/kvm/vmx/posted_intr.c b/arch/x86/kvm/vmx/posted_intr.c
> > index 2b2206339174..901b7a5f7777 100644
> > --- a/arch/x86/kvm/vmx/posted_intr.c
> > +++ b/arch/x86/kvm/vmx/posted_intr.c
> > @@ -10,10 +10,22 @@
> >  #include "vmx.h"
> >  
> >  /*
> > - * We maintain a per-CPU linked-list of vCPU, so in wakeup_handler() we
> > - * can find which vCPU should be waken up.
> > + * Maintain a per-CPU list of vCPUs that need to be awakened by wakeup_handler()
> Nit: While at it, it would be nice to rename this to pi_wakeup_hanlder() so
> that it can be more easilly found.

Ah, good catch.

> > + * when a WAKEUP_VECTOR interrupted is posted.  vCPUs are added to the list when
> > + * the vCPU is scheduled out and is blocking (e.g. in HLT) with IRQs enabled.
> s/interrupted/interrupt ?
> 
> Isn't that comment incorrect? As I see, the PI hardware is setup to use the WAKEUP_VECTOR
> when vcpu blocks (in pi_pre_block) and then that vcpu is added to the list.
> The pi_wakeup_hanlder just goes over the list and wakes up all vcpus on the lsit.

Doh, yes.  This patch is predicting the future.  The comment becomes correct as of 

  KVM: VMX: Handle PI wakeup shenanigans during vcpu_put/load

but as of this patch the "scheduled out" piece doesn't hold true.
 
> > + * The vCPUs posted interrupt descriptor is updated at the same time to set its
> > + * notification vector to WAKEUP_VECTOR, so that posted interrupt from devices
> > + * wake the target vCPUs.  vCPUs are removed from the list and the notification
> > + * vector is reset when the vCPU is scheduled in.
> >   */
> >  static DEFINE_PER_CPU(struct list_head, blocked_vcpu_on_cpu);
> Also while at it, why not to rename this to 'blocked_vcpu_list'?
> to explain that this is list of blocked vcpus. Its a per-cpu variable
> so 'on_cpu' suffix isn't needed IMHO.

As you noted, addressed in a future patch.

> > +/*
> > + * Protect the per-CPU list with a per-CPU spinlock to handle task migration.
> > + * When a blocking vCPU is awakened _and_ migrated to a different pCPU, the
> > + * ->sched_in() path will need to take the vCPU off the list of the _previous_
> > + * CPU.  IRQs must be disabled when taking this lock, otherwise deadlock will
> > + * occur if a wakeup IRQ arrives and attempts to acquire the lock.
> > + */
> >  static DEFINE_PER_CPU(spinlock_t, blocked_vcpu_on_cpu_lock);
> >  
> >  static inline struct pi_desc *vcpu_to_pi_desc(struct kvm_vcpu *vcpu)
> > @@ -101,23 +113,28 @@ static void __pi_post_block(struct kvm_vcpu *vcpu)
> >  	WARN(pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR,
> >  	     "Wakeup handler not enabled while the vCPU was blocking");
> >  
> > -	dest = cpu_physical_id(vcpu->cpu);
> > -	if (!x2apic_mode)
> > -		dest = (dest << 8) & 0xFF00;
> > -
> > -	do {
> > -		old.control = new.control = READ_ONCE(pi_desc->control);
> > -
> > -		new.ndst = dest;
> > -
> > -		/* set 'NV' to 'notification vector' */
> > -		new.nv = POSTED_INTR_VECTOR;
> > -	} while (cmpxchg64(&pi_desc->control, old.control,
> > -			   new.control) != old.control);
> > -
> > +	/*
> > +	 * Remove the vCPU from the wakeup list of the _previous_ pCPU, which
> > +	 * will not be the same as the current pCPU if the task was migrated.
> > +	 */
> >  	spin_lock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> >  	list_del(&vcpu->blocked_vcpu_list);
> >  	spin_unlock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> > +
> > +	dest = cpu_physical_id(vcpu->cpu);
> > +	if (!x2apic_mode)
> > +		dest = (dest << 8) & 0xFF00;
> It would be nice to have a function for this, this appears in this file twice.
> Maybe there is a function already somewhere?

The second instance does go away by the aforementioned:

  KVM: VMX: Handle PI wakeup shenanigans during vcpu_put/load

I'm inclined to say we don't want a helper because there should only ever be one
path that changes PI.ndst.  But a comment would definitely help to explain the
difference between xAPIC and x2APIC IDs.
Maxim Levitsky Oct. 31, 2021, 10:52 p.m. UTC | #3
On Thu, 2021-10-28 at 17:19 +0000, Sean Christopherson wrote:
> On Thu, Oct 28, 2021, Maxim Levitsky wrote:
> > On Fri, 2021-10-08 at 19:12 -0700, Sean Christopherson wrote:
> > > Remove the vCPU from the wakeup list before updating the notification
> > > vector in the posted interrupt post-block helper.  There is no need to
> > > wake the current vCPU as it is by definition not blocking.  Practically
> > > speaking this is a nop as it only shaves a few meager cycles in the
> > > unlikely case that the vCPU was migrated and the previous pCPU gets a
> > > wakeup IRQ right before PID.NV is updated.  The real motivation is to
> > > allow for more readable code in the future, when post-block is merged
> > > with vmx_vcpu_pi_load(), at which point removal from the list will be
> > > conditional on the old notification vector.
> > > 
> > > Opportunistically add comments to document why KVM has a per-CPU spinlock
> > > that, at first glance, appears to be taken only on the owning CPU.
> > > Explicitly call out that the spinlock must be taken with IRQs disabled, a
> > > detail that was "lost" when KVM switched from spin_lock_irqsave() to
> > > spin_lock(), with IRQs disabled for the entirety of the relevant path.
> > > 
> > > Signed-off-by: Sean Christopherson <seanjc@google.com>
> > > ---
> > >  arch/x86/kvm/vmx/posted_intr.c | 49 +++++++++++++++++++++++-----------
> > >  1 file changed, 33 insertions(+), 16 deletions(-)
> > > 
> > > diff --git a/arch/x86/kvm/vmx/posted_intr.c b/arch/x86/kvm/vmx/posted_intr.c
> > > index 2b2206339174..901b7a5f7777 100644
> > > --- a/arch/x86/kvm/vmx/posted_intr.c
> > > +++ b/arch/x86/kvm/vmx/posted_intr.c
> > > @@ -10,10 +10,22 @@
> > >  #include "vmx.h"
> > >  
> > >  /*
> > > - * We maintain a per-CPU linked-list of vCPU, so in wakeup_handler() we
> > > - * can find which vCPU should be waken up.
> > > + * Maintain a per-CPU list of vCPUs that need to be awakened by wakeup_handler()
> > Nit: While at it, it would be nice to rename this to pi_wakeup_hanlder() so
> > that it can be more easilly found.
> 
> Ah, good catch.
> 
> > > + * when a WAKEUP_VECTOR interrupted is posted.  vCPUs are added to the list when
> > > + * the vCPU is scheduled out and is blocking (e.g. in HLT) with IRQs enabled.
> > s/interrupted/interrupt ?
> > 
> > Isn't that comment incorrect? As I see, the PI hardware is setup to use the WAKEUP_VECTOR
> > when vcpu blocks (in pi_pre_block) and then that vcpu is added to the list.
> > The pi_wakeup_hanlder just goes over the list and wakes up all vcpus on the lsit.
> 
> Doh, yes.  This patch is predicting the future.  The comment becomes correct as of 
> 
>   KVM: VMX: Handle PI wakeup shenanigans during vcpu_put/load
> 
> but as of this patch the "scheduled out" piece doesn't hold true.
>  
> > > + * The vCPUs posted interrupt descriptor is updated at the same time to set its
> > > + * notification vector to WAKEUP_VECTOR, so that posted interrupt from devices
> > > + * wake the target vCPUs.  vCPUs are removed from the list and the notification
> > > + * vector is reset when the vCPU is scheduled in.
> > >   */
> > >  static DEFINE_PER_CPU(struct list_head, blocked_vcpu_on_cpu);
> > Also while at it, why not to rename this to 'blocked_vcpu_list'?
> > to explain that this is list of blocked vcpus. Its a per-cpu variable
> > so 'on_cpu' suffix isn't needed IMHO.
> 
> As you noted, addressed in a future patch.
> 
> > > +/*
> > > + * Protect the per-CPU list with a per-CPU spinlock to handle task migration.
> > > + * When a blocking vCPU is awakened _and_ migrated to a different pCPU, the
> > > + * ->sched_in() path will need to take the vCPU off the list of the _previous_
> > > + * CPU.  IRQs must be disabled when taking this lock, otherwise deadlock will
> > > + * occur if a wakeup IRQ arrives and attempts to acquire the lock.
> > > + */
> > >  static DEFINE_PER_CPU(spinlock_t, blocked_vcpu_on_cpu_lock);
> > >  
> > >  static inline struct pi_desc *vcpu_to_pi_desc(struct kvm_vcpu *vcpu)
> > > @@ -101,23 +113,28 @@ static void __pi_post_block(struct kvm_vcpu *vcpu)
> > >  	WARN(pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR,
> > >  	     "Wakeup handler not enabled while the vCPU was blocking");
> > >  
> > > -	dest = cpu_physical_id(vcpu->cpu);
> > > -	if (!x2apic_mode)
> > > -		dest = (dest << 8) & 0xFF00;
> > > -
> > > -	do {
> > > -		old.control = new.control = READ_ONCE(pi_desc->control);
> > > -
> > > -		new.ndst = dest;
> > > -
> > > -		/* set 'NV' to 'notification vector' */
> > > -		new.nv = POSTED_INTR_VECTOR;
> > > -	} while (cmpxchg64(&pi_desc->control, old.control,
> > > -			   new.control) != old.control);
> > > -
> > > +	/*
> > > +	 * Remove the vCPU from the wakeup list of the _previous_ pCPU, which
> > > +	 * will not be the same as the current pCPU if the task was migrated.
> > > +	 */
> > >  	spin_lock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> > >  	list_del(&vcpu->blocked_vcpu_list);
> > >  	spin_unlock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
> > > +
> > > +	dest = cpu_physical_id(vcpu->cpu);
> > > +	if (!x2apic_mode)
> > > +		dest = (dest << 8) & 0xFF00;
> > It would be nice to have a function for this, this appears in this file twice.
> > Maybe there is a function already somewhere?
> 
> The second instance does go away by the aforementioned:

Then no need for a helper.

> 
>   KVM: VMX: Handle PI wakeup shenanigans during vcpu_put/load
> 
> I'm inclined to say we don't want a helper because there should only ever be one
> path that changes PI.ndst.  But a comment would definitely help to explain the
> difference between xAPIC and x2APIC IDs.
> 

Makes sense!

Best regards,
	Maxim Levitsky
diff mbox series

Patch

diff --git a/arch/x86/kvm/vmx/posted_intr.c b/arch/x86/kvm/vmx/posted_intr.c
index 2b2206339174..901b7a5f7777 100644
--- a/arch/x86/kvm/vmx/posted_intr.c
+++ b/arch/x86/kvm/vmx/posted_intr.c
@@ -10,10 +10,22 @@ 
 #include "vmx.h"
 
 /*
- * We maintain a per-CPU linked-list of vCPU, so in wakeup_handler() we
- * can find which vCPU should be waken up.
+ * Maintain a per-CPU list of vCPUs that need to be awakened by wakeup_handler()
+ * when a WAKEUP_VECTOR interrupted is posted.  vCPUs are added to the list when
+ * the vCPU is scheduled out and is blocking (e.g. in HLT) with IRQs enabled.
+ * The vCPUs posted interrupt descriptor is updated at the same time to set its
+ * notification vector to WAKEUP_VECTOR, so that posted interrupt from devices
+ * wake the target vCPUs.  vCPUs are removed from the list and the notification
+ * vector is reset when the vCPU is scheduled in.
  */
 static DEFINE_PER_CPU(struct list_head, blocked_vcpu_on_cpu);
+/*
+ * Protect the per-CPU list with a per-CPU spinlock to handle task migration.
+ * When a blocking vCPU is awakened _and_ migrated to a different pCPU, the
+ * ->sched_in() path will need to take the vCPU off the list of the _previous_
+ * CPU.  IRQs must be disabled when taking this lock, otherwise deadlock will
+ * occur if a wakeup IRQ arrives and attempts to acquire the lock.
+ */
 static DEFINE_PER_CPU(spinlock_t, blocked_vcpu_on_cpu_lock);
 
 static inline struct pi_desc *vcpu_to_pi_desc(struct kvm_vcpu *vcpu)
@@ -101,23 +113,28 @@  static void __pi_post_block(struct kvm_vcpu *vcpu)
 	WARN(pi_desc->nv != POSTED_INTR_WAKEUP_VECTOR,
 	     "Wakeup handler not enabled while the vCPU was blocking");
 
-	dest = cpu_physical_id(vcpu->cpu);
-	if (!x2apic_mode)
-		dest = (dest << 8) & 0xFF00;
-
-	do {
-		old.control = new.control = READ_ONCE(pi_desc->control);
-
-		new.ndst = dest;
-
-		/* set 'NV' to 'notification vector' */
-		new.nv = POSTED_INTR_VECTOR;
-	} while (cmpxchg64(&pi_desc->control, old.control,
-			   new.control) != old.control);
-
+	/*
+	 * Remove the vCPU from the wakeup list of the _previous_ pCPU, which
+	 * will not be the same as the current pCPU if the task was migrated.
+	 */
 	spin_lock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
 	list_del(&vcpu->blocked_vcpu_list);
 	spin_unlock(&per_cpu(blocked_vcpu_on_cpu_lock, vcpu->pre_pcpu));
+
+	dest = cpu_physical_id(vcpu->cpu);
+	if (!x2apic_mode)
+		dest = (dest << 8) & 0xFF00;
+
+	do {
+		old.control = new.control = READ_ONCE(pi_desc->control);
+
+		new.ndst = dest;
+
+		/* set 'NV' to 'notification vector' */
+		new.nv = POSTED_INTR_VECTOR;
+	} while (cmpxchg64(&pi_desc->control, old.control,
+			   new.control) != old.control);
+
 	vcpu->pre_pcpu = -1;
 }