Re: [PATCH 1/2] x86/entry: Avoid interrupt flag save and restore

From: Paolo Bonzini
Date: Wed Jun 08 2016 - 08:34:53 EST




On 08/06/2016 14:16, Ingo Molnar wrote:
> > The guest ones are not quite as consistent. I can fix that later,
> > there's no reason also to have guest context tracking split between
> > include/linux/context_tracking.h and include/linux/kvm_host.h.
>
> Could we please first do the cleanups before complicating the code and applying
> more substantial changes?

The further cleanups wouldn't complicate the code. It's just that
guest_enter/guest_exit require IRQs off but don't have __.

I'm thinking of something like this (untested):

diff --git a/include/linux/context_tracking.h b/include/linux/context_tracking.h
index d259274238db..c2dc581ddb0e 100644
--- a/include/linux/context_tracking.h
+++ b/include/linux/context_tracking.h
@@ -84,7 +84,7 @@ static inline void context_tracking_init(void) { }


#ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
-static inline void guest_enter(void)
+static inline void __guest_enter(void)
{
if (vtime_accounting_cpu_enabled())
vtime_guest_enter(current);
@@ -93,9 +93,19 @@ static inline void guest_enter(void)

if (context_tracking_is_enabled())
__context_tracking_enter(CONTEXT_GUEST);
+
+ /* KVM does not hold any references to rcu protected data when it
+ * switches CPU into a guest mode. In fact switching to a guest mode
+ * is very similar to exiting to userspace from rcu point of view. In
+ * addition CPU may stay in a guest mode for quite a long time (up to
+ * one time slice). Lets treat guest mode as quiescent state, just like
+ * we do with user-mode execution.
+ */
+ if (!context_tracking_cpu_is_enabled())
+ rcu_virt_note_context_switch(smp_processor_id());
}

-static inline void guest_exit(void)
+static inline void __guest_exit(void)
{
if (context_tracking_is_enabled())
__context_tracking_exit(CONTEXT_GUEST);
@@ -107,7 +117,7 @@ static inline void guest_exit(void)
}

#else
-static inline void guest_enter(void)
+static inline void __guest_enter(void)
{
/*
* This is running in ioctl context so its safe
@@ -118,7 +128,7 @@ static inline void guest_enter(void)
current->flags |= PF_VCPU;
}

-static inline void guest_exit(void)
+static inline void __guest_exit(void)
{
/* Flush the guest cputime we spent on the guest */
vtime_account_system(current);
@@ -126,4 +136,23 @@ static inline void guest_exit(void)
}
#endif /* CONFIG_VIRT_CPU_ACCOUNTING_GEN */

+static inline void guest_enter(void)
+{
+ unsigned long flags;
+
+ local_irq_save(flags);
+ __guest_enter();
+ local_irq_restore(flags);
+}
+
+/* must be called with irqs disabled */
+static inline void guest_exit(void)
+{
+ unsigned long flags;
+
+ local_irq_save(flags);
+ __guest_exit();
+ local_irq_restore(flags);
+}
+
#endif
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 5276fe0916fc..d00fdaa8da15 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -861,40 +861,23 @@ static inline void kvm_iommu_unmap_pages(struct kvm *kvm,
/* must be called with irqs disabled */
static inline void __kvm_guest_enter(void)
{
- guest_enter();
- /* KVM does not hold any references to rcu protected data when it
- * switches CPU into a guest mode. In fact switching to a guest mode
- * is very similar to exiting to userspace from rcu point of view. In
- * addition CPU may stay in a guest mode for quite a long time (up to
- * one time slice). Lets treat guest mode as quiescent state, just like
- * we do with user-mode execution.
- */
- if (!context_tracking_cpu_is_enabled())
- rcu_virt_note_context_switch(smp_processor_id());
+ __guest_enter();
}

/* must be called with irqs disabled */
static inline void __kvm_guest_exit(void)
{
- guest_exit();
+ __guest_exit();
}

static inline void kvm_guest_enter(void)
{
- unsigned long flags;
-
- local_irq_save(flags);
- __kvm_guest_enter();
- local_irq_restore(flags);
+ guest_enter();
}

static inline void kvm_guest_exit(void)
{
- unsigned long flags;
-
- local_irq_save(flags);
- __kvm_guest_exit();
- local_irq_restore(flags);
+ guest_exit();
}

/*


and then removing the kvm_-prefixed functions. It's little
more than code movement.

Paolo