Linux Virtualization - May 2019 - [RFC PATCH 5/6] x86/mm/tlb: Flush remote and local TLBs concurrently

On Sat, May 25, 2019 at 10:54:50AM +0200, Juergen Gross
wrote:
> On 25/05/2019 10:22, Nadav Amit wrote:
> > diff --git a/arch/x86/include/asm/paravirt_types.h
b/arch/x86/include/asm/paravirt_types.h
> > index 946f8f1f1efc..3a156e63c57d 100644
> > --- a/arch/x86/include/asm/paravirt_types.h
> > +++ b/arch/x86/include/asm/paravirt_types.h
> > @@ -211,6 +211,12 @@ struct pv_mmu_ops {
> >  	void (*flush_tlb_user)(void);
> >  	void (*flush_tlb_kernel)(void);
> >  	void (*flush_tlb_one_user)(unsigned long addr);
> > +	/*
> > +	 * flush_tlb_multi() is the preferred interface. When it is used,
> > +	 * flush_tlb_others() should return false.
> 
> This comment does not make sense. flush_tlb_others() return type is
> void.
I suspect that is an artifact from before the static_key; an attempt to
make the pv interface less awkward.

Something like the below would work for KVM I suspect, the others
(Hyper-V and Xen are more 'interesting').

---
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -580,7 +580,7 @@ static void __init kvm_apf_trap_init(voi
 
 static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
 
-static void kvm_flush_tlb_others(const struct cpumask *cpumask,
+static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
 			const struct flush_tlb_info *info)
 {
 	u8 state;
@@ -594,6 +594,9 @@ static void kvm_flush_tlb_others(const s
 	 * queue flush_on_enter for pre-empted vCPUs
 	 */
 	for_each_cpu(cpu, flushmask) {
+		if (cpu == smp_processor_id())
+			continue;
+
 		src = &per_cpu(steal_time, cpu);
 		state = READ_ONCE(src->preempted);
 		if ((state & KVM_VCPU_PREEMPTED)) {
@@ -603,7 +606,7 @@ static void kvm_flush_tlb_others(const s
 		}
 	}
 
-	native_flush_tlb_others(flushmask, info);
+	native_flush_tlb_multi(flushmask, info);
 }
 
 static void __init kvm_guest_init(void)
@@ -628,9 +631,8 @@ static void __init kvm_guest_init(void)
 	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
 	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
 	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
-		pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
+		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
 		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
-		static_key_disable(&flush_tlb_multi_enabled.key);
 	}
 
 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))

On 27/05/19 11:47, Peter Zijlstra wrote:
> On Sat, May 25, 2019 at 10:54:50AM +0200, Juergen Gross wrote:
>> On 25/05/2019 10:22, Nadav Amit wrote:
> 
>>> diff --git a/arch/x86/include/asm/paravirt_types.h
b/arch/x86/include/asm/paravirt_types.h
>>> index 946f8f1f1efc..3a156e63c57d 100644
>>> --- a/arch/x86/include/asm/paravirt_types.h
>>> +++ b/arch/x86/include/asm/paravirt_types.h
>>> @@ -211,6 +211,12 @@ struct pv_mmu_ops {
>>>  	void (*flush_tlb_user)(void);
>>>  	void (*flush_tlb_kernel)(void);
>>>  	void (*flush_tlb_one_user)(unsigned long addr);
>>> +	/*
>>> +	 * flush_tlb_multi() is the preferred interface. When it is used,
>>> +	 * flush_tlb_others() should return false.
>>
>> This comment does not make sense. flush_tlb_others() return type is
>> void.
> 
> I suspect that is an artifact from before the static_key; an attempt to
> make the pv interface less awkward.
> 
> Something like the below would work for KVM I suspect, the others
> (Hyper-V and Xen are more 'interesting').
> 
> ---
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -580,7 +580,7 @@ static void __init kvm_apf_trap_init(voi
>  
>  static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
>  
> -static void kvm_flush_tlb_others(const struct cpumask *cpumask,
> +static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
>  			const struct flush_tlb_info *info)
>  {
>  	u8 state;
> @@ -594,6 +594,9 @@ static void kvm_flush_tlb_others(const s
>  	 * queue flush_on_enter for pre-empted vCPUs
>  	 */
>  	for_each_cpu(cpu, flushmask) {
> +		if (cpu == smp_processor_id())
> +			continue;
> +
Even this would be just an optimization; the vCPU you're running on
cannot be preempted.  You can just change others to multi.

Paolo
>  		src = &per_cpu(steal_time, cpu);
>  		state = READ_ONCE(src->preempted);
>  		if ((state & KVM_VCPU_PREEMPTED)) {
> @@ -603,7 +606,7 @@ static void kvm_flush_tlb_others(const s
>  		}
>  	}
>  
> -	native_flush_tlb_others(flushmask, info);
> +	native_flush_tlb_multi(flushmask, info);
>  }
>  
>  static void __init kvm_guest_init(void)
> @@ -628,9 +631,8 @@ static void __init kvm_guest_init(void)
>  	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
>  	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
>  	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
> -		pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
> +		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
>  		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
> -		static_key_disable(&flush_tlb_multi_enabled.key);
>  	}
>  
>  	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
>

On Mon, May 27, 2019 at 12:21:59PM +0200, Paolo Bonzini
wrote:
> On 27/05/19 11:47, Peter Zijlstra wrote:
> > --- a/arch/x86/kernel/kvm.c
> > +++ b/arch/x86/kernel/kvm.c
> > @@ -580,7 +580,7 @@ static void __init kvm_apf_trap_init(voi
> >  
> >  static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
> >  
> > -static void kvm_flush_tlb_others(const struct cpumask *cpumask,
> > +static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
> >  			const struct flush_tlb_info *info)
> >  {
> >  	u8 state;
> > @@ -594,6 +594,9 @@ static void kvm_flush_tlb_others(const s
> >  	 * queue flush_on_enter for pre-empted vCPUs
> >  	 */
> >  	for_each_cpu(cpu, flushmask) {
> > +		if (cpu == smp_processor_id())
> > +			continue;
> > +
> 
> Even this would be just an optimization; the vCPU you're running on
> cannot be preempted.  You can just change others to multi.
Yeah, I know, but it felt weird so I added the explicit skip. No strong
feelings though.

> On May 27, 2019, at 2:47 AM, Peter Zijlstra <peterz at infradead.org>
wrote:
> 
> On Sat, May 25, 2019 at 10:54:50AM +0200, Juergen Gross wrote:
>> On 25/05/2019 10:22, Nadav Amit wrote:
> 
>>> diff --git a/arch/x86/include/asm/paravirt_types.h
b/arch/x86/include/asm/paravirt_types.h
>>> index 946f8f1f1efc..3a156e63c57d 100644
>>> --- a/arch/x86/include/asm/paravirt_types.h
>>> +++ b/arch/x86/include/asm/paravirt_types.h
>>> @@ -211,6 +211,12 @@ struct pv_mmu_ops {
>>> 	void (*flush_tlb_user)(void);
>>> 	void (*flush_tlb_kernel)(void);
>>> 	void (*flush_tlb_one_user)(unsigned long addr);
>>> +	/*
>>> +	 * flush_tlb_multi() is the preferred interface. When it is used,
>>> +	 * flush_tlb_others() should return false.
>> 
>> This comment does not make sense. flush_tlb_others() return type is
>> void.
> 
> I suspect that is an artifact from before the static_key; an attempt to
> make the pv interface less awkward.
Yes, remainders that should have been removed - I will remove them for the
next version.
> Something like the below would work for KVM I suspect, the others
> (Hyper-V and Xen are more 'interesting').
> 
> ---
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -580,7 +580,7 @@ static void __init kvm_apf_trap_init(voi
> 
> static DEFINE_PER_CPU(cpumask_var_t, __pv_tlb_mask);
> 
> -static void kvm_flush_tlb_others(const struct cpumask *cpumask,
> +static void kvm_flush_tlb_multi(const struct cpumask *cpumask,
> 			const struct flush_tlb_info *info)
> {
> 	u8 state;
> @@ -594,6 +594,9 @@ static void kvm_flush_tlb_others(const s
> 	 * queue flush_on_enter for pre-empted vCPUs
> 	 */
> 	for_each_cpu(cpu, flushmask) {
> +		if (cpu == smp_processor_id())
> +			continue;
> +
> 		src = &per_cpu(steal_time, cpu);
> 		state = READ_ONCE(src->preempted);
> 		if ((state & KVM_VCPU_PREEMPTED)) {
> @@ -603,7 +606,7 @@ static void kvm_flush_tlb_others(const s
> 		}
> 	}
> 
> -	native_flush_tlb_others(flushmask, info);
> +	native_flush_tlb_multi(flushmask, info);
> }
> 
> static void __init kvm_guest_init(void)
> @@ -628,9 +631,8 @@ static void __init kvm_guest_init(void)
> 	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
> 	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
> 	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME)) {
> -		pv_ops.mmu.flush_tlb_others = kvm_flush_tlb_others;
> +		pv_ops.mmu.flush_tlb_multi = kvm_flush_tlb_multi;
> 		pv_ops.mmu.tlb_remove_table = tlb_remove_table;
> -		static_key_disable(&flush_tlb_multi_enabled.key);
> 	}
> 
> 	if (kvm_para_has_feature(KVM_FEATURE_PV_EOI))
That?s what I have as well ;-).

As you mentioned (in another email), specifically hyper-v code seems
convoluted to me. In general, I prefer not to touch KVM/Xen/hyper-v, but you
twist my arm, I will send a compile-tested version for Xen and hyper-v.