[PATCH 0/3] LoongArch: KVM: Some small enhanments

[PATCH 0/3] LoongArch: KVM: Some small enhanments

[Bibo Mao]

There are some small enhancements about LoongArch KVM. There is
improvement about VM entry path in patch 1 and patch, add pending
request checking in advance and it is empty at most time.

Also change vcpu_is_preempted() from function to macro method, by actual
benchmark test such as unixbench. It is affected on physical machine
with function method. With the macro method, only one nop instruction
is added and there is no parameter parsing, there is no obvious
performance influence with physical machine.

Bibo Mao (3):
  LoongArch: KVM: Add kvm_request_pending checking in
    kvm_late_check_requests()
  LoongArch: KVM: Move host CSR_EENTRY save in context switch
  LoongArch: KVM: Set vcpu_is_preempted() macro rather than function

 arch/loongarch/include/asm/qspinlock.h | 27 +++++++++++++++++++++-----
 arch/loongarch/kernel/paravirt.c       | 15 ++------------
 arch/loongarch/kvm/vcpu.c              | 24 +++++++++++------------
 3 files changed, 36 insertions(+), 30 deletions(-)


base-commit: 1f318b96cc84d7c2ab792fcc0bfd42a7ca890681
-- 
2.39.3

[PATCH 1/3] LoongArch: KVM: Add kvm_request_pending checking in kvm_late_check_requests()

[Bibo Mao]

Add kvm_request_pending() checking firstly in kvm_late_check_requests(),
at most time there is no pending request, the following pending bit
checking can be skipped then.

Also move function kvm_check_pmu() in kvm_late_check_requests(), and put
it after kvm_request_pending() checking.

Signed-off-by: Bibo Mao <maobibo@loongson.cn>
---
 arch/loongarch/kvm/vcpu.c | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
index 09e137f2f841..6104982b53c8 100644
--- a/arch/loongarch/kvm/vcpu.c
+++ b/arch/loongarch/kvm/vcpu.c
@@ -149,14 +149,6 @@ static void kvm_lose_pmu(struct kvm_vcpu *vcpu)
 	kvm_restore_host_pmu(vcpu);
 }
 
-static void kvm_check_pmu(struct kvm_vcpu *vcpu)
-{
-	if (kvm_check_request(KVM_REQ_PMU, vcpu)) {
-		kvm_own_pmu(vcpu);
-		vcpu->arch.aux_inuse |= KVM_LARCH_PMU;
-	}
-}
-
 static void kvm_update_stolen_time(struct kvm_vcpu *vcpu)
 {
 	u32 version;
@@ -232,6 +224,14 @@ static int kvm_check_requests(struct kvm_vcpu *vcpu)
 static void kvm_late_check_requests(struct kvm_vcpu *vcpu)
 {
 	lockdep_assert_irqs_disabled();
+	if (!kvm_request_pending(vcpu))
+		return;
+
+	if (kvm_check_request(KVM_REQ_PMU, vcpu)) {
+		kvm_own_pmu(vcpu);
+		vcpu->arch.aux_inuse |= KVM_LARCH_PMU;
+	}
+
 	if (kvm_check_request(KVM_REQ_TLB_FLUSH_GPA, vcpu))
 		if (vcpu->arch.flush_gpa != INVALID_GPA) {
 			kvm_flush_tlb_gpa(vcpu, vcpu->arch.flush_gpa);
@@ -312,7 +312,6 @@ static int kvm_pre_enter_guest(struct kvm_vcpu *vcpu)
 		/* Make sure the vcpu mode has been written */
 		smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 		kvm_check_vpid(vcpu);
-		kvm_check_pmu(vcpu);
 
 		/*
 		 * Called after function kvm_check_vpid()
-- 
2.39.3

[PATCH 2/3] LoongArch: KVM: Move host CSR_EENTRY save in context switch

[Bibo Mao]

CSR register LOONGARCH_CSR_EENTRY is shared between host cpu and guest
vCPU, KVM need save and restore LOONGARCH_CSR_EENTRY register. Here move
CSR register LOONGARCH_CSR_EENTRY saving in context switch function,
rather than VM enter entry.

At most time VM enter/exit is much more frequent than vCPU thread context
switch.

Signed-off-by: Bibo Mao <maobibo@loongson.cn>
---
 arch/loongarch/kvm/vcpu.c | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
index 6104982b53c8..044627c6ddc7 100644
--- a/arch/loongarch/kvm/vcpu.c
+++ b/arch/loongarch/kvm/vcpu.c
@@ -319,7 +319,6 @@ static int kvm_pre_enter_guest(struct kvm_vcpu *vcpu)
 		 * and it may also clear KVM_REQ_TLB_FLUSH_GPA pending bit
 		 */
 		kvm_late_check_requests(vcpu);
-		vcpu->arch.host_eentry = csr_read64(LOONGARCH_CSR_EENTRY);
 		/* Clear KVM_LARCH_SWCSR_LATEST as CSR will change when enter guest */
 		vcpu->arch.aux_inuse &= ~KVM_LARCH_SWCSR_LATEST;
 
@@ -1624,9 +1623,11 @@ static int _kvm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	 * If not, any old guest state from this vCPU will have been clobbered.
 	 */
 	context = per_cpu_ptr(vcpu->kvm->arch.vmcs, cpu);
-	if (migrated || (context->last_vcpu != vcpu))
+	if (migrated || (context->last_vcpu != vcpu)) {
 		vcpu->arch.aux_inuse &= ~KVM_LARCH_HWCSR_USABLE;
-	context->last_vcpu = vcpu;
+		context->last_vcpu = vcpu;
+		vcpu->arch.host_eentry = csr_read64(LOONGARCH_CSR_EENTRY);
+	}
 
 	/* Restore timer state regardless */
 	kvm_restore_timer(vcpu);
-- 
2.39.3

[PATCH 3/3] LoongArch: KVM: Set vcpu_is_preempted() macro rather than function

[Bibo Mao]

vcpu_is_preempted() is performance sensitive called in function
osq_lock(), here set it as macro. So that parameter is not parsed
at most time, it can avoid cache line thrashing across numa node.

Here is part of unixbench result on 3C5000 DualWay machine with 32
Cores and 2 Numa node.
          origin    with patch    CONFIG_PARAVIRT disabled
execl     6871.9    7134.2        7190.8
fstime    425.5     959.9         956.1

From the test result, with macro method it is almost the same with
CONFIG_PARAVIRT disabled, and there is some improvment compared with
function method.

Signed-off-by: Bibo Mao <maobibo@loongson.cn>
---
 arch/loongarch/include/asm/qspinlock.h | 27 +++++++++++++++++++++-----
 arch/loongarch/kernel/paravirt.c       | 15 ++------------
 2 files changed, 24 insertions(+), 18 deletions(-)

diff --git a/arch/loongarch/include/asm/qspinlock.h b/arch/loongarch/include/asm/qspinlock.h
index 66244801db67..305afd37cdcb 100644
--- a/arch/loongarch/include/asm/qspinlock.h
+++ b/arch/loongarch/include/asm/qspinlock.h
@@ -5,8 +5,10 @@
 #include <linux/jump_label.h>
 
 #ifdef CONFIG_PARAVIRT
-
+#include <asm/kvm_para.h>
 DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
+DECLARE_STATIC_KEY_FALSE(virt_preempt_key);
+DECLARE_PER_CPU(struct kvm_steal_time, steal_time);
 
 #define virt_spin_lock virt_spin_lock
 
@@ -34,10 +36,25 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
 	return true;
 }
 
-#define vcpu_is_preempted vcpu_is_preempted
-
-bool vcpu_is_preempted(int cpu);
-
+/*
+ * Macro is better than inline function here
+ * With inline function, parameter cpu is parsed even though it is not used.
+ * This may cause cache line thrashing across NUMA node.
+ * With macro method, parameter cpu is parsed only when it is used.
+ */
+#define vcpu_is_preempted(cpu)						\
+({									\
+	bool __val;							\
+									\
+	if (!static_branch_unlikely(&virt_preempt_key))			\
+		__val = false;						\
+	else {								\
+		struct kvm_steal_time *src;				\
+		src = &per_cpu(steal_time, cpu);			\
+		__val = !!(src->preempted & KVM_VCPU_PREEMPTED);	\
+	}								\
+	__val;								\
+})
 #endif /* CONFIG_PARAVIRT */
 
 #include <asm-generic/qspinlock.h>
diff --git a/arch/loongarch/kernel/paravirt.c b/arch/loongarch/kernel/paravirt.c
index b74fe6db49ab..2d1206e486e2 100644
--- a/arch/loongarch/kernel/paravirt.c
+++ b/arch/loongarch/kernel/paravirt.c
@@ -10,8 +10,8 @@
 #include <asm/paravirt.h>
 
 static int has_steal_clock;
-static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
-static DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
+DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
+DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
 DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
 
 static bool steal_acc = true;
@@ -261,17 +261,6 @@ static int pv_time_cpu_down_prepare(unsigned int cpu)
 	return 0;
 }
 
-bool vcpu_is_preempted(int cpu)
-{
-	struct kvm_steal_time *src;
-
-	if (!static_branch_unlikely(&virt_preempt_key))
-		return false;
-
-	src = &per_cpu(steal_time, cpu);
-	return !!(src->preempted & KVM_VCPU_PREEMPTED);
-}
-EXPORT_SYMBOL(vcpu_is_preempted);
 #endif
 
 static void pv_cpu_reboot(void *unused)
-- 
2.39.3

Re: [PATCH 3/3] LoongArch: KVM: Set vcpu_is_preempted() macro rather than function

[Huacai Chen]

Hi, Bibo,

On Wed, Mar 11, 2026 at 11:21 AM Bibo Mao <maobibo@loongson.cn> wrote:
>
> vcpu_is_preempted() is performance sensitive called in function
> osq_lock(), here set it as macro. So that parameter is not parsed
> at most time, it can avoid cache line thrashing across numa node.
I remember that I had suggested you define vcpu_is_preempted() in the
header file but you refused.

Huacai

>
> Here is part of unixbench result on 3C5000 DualWay machine with 32
> Cores and 2 Numa node.
>           origin    with patch    CONFIG_PARAVIRT disabled
> execl     6871.9    7134.2        7190.8
> fstime    425.5     959.9         956.1
>
> From the test result, with macro method it is almost the same with
> CONFIG_PARAVIRT disabled, and there is some improvment compared with
> function method.
>
> Signed-off-by: Bibo Mao <maobibo@loongson.cn>
> ---
>  arch/loongarch/include/asm/qspinlock.h | 27 +++++++++++++++++++++-----
>  arch/loongarch/kernel/paravirt.c       | 15 ++------------
>  2 files changed, 24 insertions(+), 18 deletions(-)
>
> diff --git a/arch/loongarch/include/asm/qspinlock.h b/arch/loongarch/include/asm/qspinlock.h
> index 66244801db67..305afd37cdcb 100644
> --- a/arch/loongarch/include/asm/qspinlock.h
> +++ b/arch/loongarch/include/asm/qspinlock.h
> @@ -5,8 +5,10 @@
>  #include <linux/jump_label.h>
>
>  #ifdef CONFIG_PARAVIRT
> -
> +#include <asm/kvm_para.h>
>  DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
> +DECLARE_STATIC_KEY_FALSE(virt_preempt_key);
> +DECLARE_PER_CPU(struct kvm_steal_time, steal_time);
>
>  #define virt_spin_lock virt_spin_lock
>
> @@ -34,10 +36,25 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
>         return true;
>  }
>
> -#define vcpu_is_preempted vcpu_is_preempted
> -
> -bool vcpu_is_preempted(int cpu);
> -
> +/*
> + * Macro is better than inline function here
> + * With inline function, parameter cpu is parsed even though it is not used.
> + * This may cause cache line thrashing across NUMA node.
> + * With macro method, parameter cpu is parsed only when it is used.
> + */
> +#define vcpu_is_preempted(cpu)                                         \
> +({                                                                     \
> +       bool __val;                                                     \
> +                                                                       \
> +       if (!static_branch_unlikely(&virt_preempt_key))                 \
> +               __val = false;                                          \
> +       else {                                                          \
> +               struct kvm_steal_time *src;                             \
> +               src = &per_cpu(steal_time, cpu);                        \
> +               __val = !!(src->preempted & KVM_VCPU_PREEMPTED);        \
> +       }                                                               \
> +       __val;                                                          \
> +})
>  #endif /* CONFIG_PARAVIRT */
>
>  #include <asm-generic/qspinlock.h>
> diff --git a/arch/loongarch/kernel/paravirt.c b/arch/loongarch/kernel/paravirt.c
> index b74fe6db49ab..2d1206e486e2 100644
> --- a/arch/loongarch/kernel/paravirt.c
> +++ b/arch/loongarch/kernel/paravirt.c
> @@ -10,8 +10,8 @@
>  #include <asm/paravirt.h>
>
>  static int has_steal_clock;
> -static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
> -static DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
> +DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
> +DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
>  DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
>
>  static bool steal_acc = true;
> @@ -261,17 +261,6 @@ static int pv_time_cpu_down_prepare(unsigned int cpu)
>         return 0;
>  }
>
> -bool vcpu_is_preempted(int cpu)
> -{
> -       struct kvm_steal_time *src;
> -
> -       if (!static_branch_unlikely(&virt_preempt_key))
> -               return false;
> -
> -       src = &per_cpu(steal_time, cpu);
> -       return !!(src->preempted & KVM_VCPU_PREEMPTED);
> -}
> -EXPORT_SYMBOL(vcpu_is_preempted);
>  #endif
>
>  static void pv_cpu_reboot(void *unused)
> --
> 2.39.3
>
>

Re: [PATCH 3/3] LoongArch: KVM: Set vcpu_is_preempted() macro rather than function

[Bibo Mao]

On 2026/3/11 上午11:35, Huacai Chen wrote:
> Hi, Bibo,
> 
> On Wed, Mar 11, 2026 at 11:21 AM Bibo Mao <maobibo@loongson.cn> wrote:
>>
>> vcpu_is_preempted() is performance sensitive called in function
>> osq_lock(), here set it as macro. So that parameter is not parsed
>> at most time, it can avoid cache line thrashing across numa node.
> I remember that I had suggested you define vcpu_is_preempted() in the
> header file but you refused.
yes, I refuse then, that is a tragedy, I am sorry for that:(

By some benchmarks, vcpu_is_preempted() is called in busy-loop method, 
more obvious with more CPUs, it may be a bottleneck.

Regards
Bibo Mao

> 
> Huacai
> 
>>
>> Here is part of unixbench result on 3C5000 DualWay machine with 32
>> Cores and 2 Numa node.
>>            origin    with patch    CONFIG_PARAVIRT disabled
>> execl     6871.9    7134.2        7190.8
>> fstime    425.5     959.9         956.1
>>
>>  From the test result, with macro method it is almost the same with
>> CONFIG_PARAVIRT disabled, and there is some improvment compared with
>> function method.
>>
>> Signed-off-by: Bibo Mao <maobibo@loongson.cn>
>> ---
>>   arch/loongarch/include/asm/qspinlock.h | 27 +++++++++++++++++++++-----
>>   arch/loongarch/kernel/paravirt.c       | 15 ++------------
>>   2 files changed, 24 insertions(+), 18 deletions(-)
>>
>> diff --git a/arch/loongarch/include/asm/qspinlock.h b/arch/loongarch/include/asm/qspinlock.h
>> index 66244801db67..305afd37cdcb 100644
>> --- a/arch/loongarch/include/asm/qspinlock.h
>> +++ b/arch/loongarch/include/asm/qspinlock.h
>> @@ -5,8 +5,10 @@
>>   #include <linux/jump_label.h>
>>
>>   #ifdef CONFIG_PARAVIRT
>> -
>> +#include <asm/kvm_para.h>
>>   DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
>> +DECLARE_STATIC_KEY_FALSE(virt_preempt_key);
>> +DECLARE_PER_CPU(struct kvm_steal_time, steal_time);
>>
>>   #define virt_spin_lock virt_spin_lock
>>
>> @@ -34,10 +36,25 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
>>          return true;
>>   }
>>
>> -#define vcpu_is_preempted vcpu_is_preempted
>> -
>> -bool vcpu_is_preempted(int cpu);
>> -
>> +/*
>> + * Macro is better than inline function here
>> + * With inline function, parameter cpu is parsed even though it is not used.
>> + * This may cause cache line thrashing across NUMA node.
>> + * With macro method, parameter cpu is parsed only when it is used.
>> + */
>> +#define vcpu_is_preempted(cpu)                                         \
>> +({                                                                     \
>> +       bool __val;                                                     \
>> +                                                                       \
>> +       if (!static_branch_unlikely(&virt_preempt_key))                 \
>> +               __val = false;                                          \
>> +       else {                                                          \
>> +               struct kvm_steal_time *src;                             \
>> +               src = &per_cpu(steal_time, cpu);                        \
>> +               __val = !!(src->preempted & KVM_VCPU_PREEMPTED);        \
>> +       }                                                               \
>> +       __val;                                                          \
>> +})
>>   #endif /* CONFIG_PARAVIRT */
>>
>>   #include <asm-generic/qspinlock.h>
>> diff --git a/arch/loongarch/kernel/paravirt.c b/arch/loongarch/kernel/paravirt.c
>> index b74fe6db49ab..2d1206e486e2 100644
>> --- a/arch/loongarch/kernel/paravirt.c
>> +++ b/arch/loongarch/kernel/paravirt.c
>> @@ -10,8 +10,8 @@
>>   #include <asm/paravirt.h>
>>
>>   static int has_steal_clock;
>> -static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
>> -static DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
>> +DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
>> +DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
>>   DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
>>
>>   static bool steal_acc = true;
>> @@ -261,17 +261,6 @@ static int pv_time_cpu_down_prepare(unsigned int cpu)
>>          return 0;
>>   }
>>
>> -bool vcpu_is_preempted(int cpu)
>> -{
>> -       struct kvm_steal_time *src;
>> -
>> -       if (!static_branch_unlikely(&virt_preempt_key))
>> -               return false;
>> -
>> -       src = &per_cpu(steal_time, cpu);
>> -       return !!(src->preempted & KVM_VCPU_PREEMPTED);
>> -}
>> -EXPORT_SYMBOL(vcpu_is_preempted);
>>   #endif
>>
>>   static void pv_cpu_reboot(void *unused)
>> --
>> 2.39.3
>>
>>

Re: [PATCH 3/3] LoongArch: KVM: Set vcpu_is_preempted() macro rather than function

[Huacai Chen]

On Wed, Mar 11, 2026 at 11:51 AM Bibo Mao <maobibo@loongson.cn> wrote:
>
>
>
> On 2026/3/11 上午11:35, Huacai Chen wrote:
> > Hi, Bibo,
> >
> > On Wed, Mar 11, 2026 at 11:21 AM Bibo Mao <maobibo@loongson.cn> wrote:
> >>
> >> vcpu_is_preempted() is performance sensitive called in function
> >> osq_lock(), here set it as macro. So that parameter is not parsed
> >> at most time, it can avoid cache line thrashing across numa node.
> > I remember that I had suggested you define vcpu_is_preempted() in the
> > header file but you refused.
> yes, I refuse then, that is a tragedy, I am sorry for that:(
>
> By some benchmarks, vcpu_is_preempted() is called in busy-loop method,
> more obvious with more CPUs, it may be a bottleneck.
I also remember that you said the static key and percpu steal_time
structure should be EXPORT_SYMBOL, but you haven't done this patch.

Huacai

>
> Regards
> Bibo Mao
>
> >
> > Huacai
> >
> >>
> >> Here is part of unixbench result on 3C5000 DualWay machine with 32
> >> Cores and 2 Numa node.
> >>            origin    with patch    CONFIG_PARAVIRT disabled
> >> execl     6871.9    7134.2        7190.8
> >> fstime    425.5     959.9         956.1
> >>
> >>  From the test result, with macro method it is almost the same with
> >> CONFIG_PARAVIRT disabled, and there is some improvment compared with
> >> function method.
> >>
> >> Signed-off-by: Bibo Mao <maobibo@loongson.cn>
> >> ---
> >>   arch/loongarch/include/asm/qspinlock.h | 27 +++++++++++++++++++++-----
> >>   arch/loongarch/kernel/paravirt.c       | 15 ++------------
> >>   2 files changed, 24 insertions(+), 18 deletions(-)
> >>
> >> diff --git a/arch/loongarch/include/asm/qspinlock.h b/arch/loongarch/include/asm/qspinlock.h
> >> index 66244801db67..305afd37cdcb 100644
> >> --- a/arch/loongarch/include/asm/qspinlock.h
> >> +++ b/arch/loongarch/include/asm/qspinlock.h
> >> @@ -5,8 +5,10 @@
> >>   #include <linux/jump_label.h>
> >>
> >>   #ifdef CONFIG_PARAVIRT
> >> -
> >> +#include <asm/kvm_para.h>
> >>   DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
> >> +DECLARE_STATIC_KEY_FALSE(virt_preempt_key);
> >> +DECLARE_PER_CPU(struct kvm_steal_time, steal_time);
> >>
> >>   #define virt_spin_lock virt_spin_lock
> >>
> >> @@ -34,10 +36,25 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
> >>          return true;
> >>   }
> >>
> >> -#define vcpu_is_preempted vcpu_is_preempted
> >> -
> >> -bool vcpu_is_preempted(int cpu);
> >> -
> >> +/*
> >> + * Macro is better than inline function here
> >> + * With inline function, parameter cpu is parsed even though it is not used.
> >> + * This may cause cache line thrashing across NUMA node.
> >> + * With macro method, parameter cpu is parsed only when it is used.
> >> + */
> >> +#define vcpu_is_preempted(cpu)                                         \
> >> +({                                                                     \
> >> +       bool __val;                                                     \
> >> +                                                                       \
> >> +       if (!static_branch_unlikely(&virt_preempt_key))                 \
> >> +               __val = false;                                          \
> >> +       else {                                                          \
> >> +               struct kvm_steal_time *src;                             \
> >> +               src = &per_cpu(steal_time, cpu);                        \
> >> +               __val = !!(src->preempted & KVM_VCPU_PREEMPTED);        \
> >> +       }                                                               \
> >> +       __val;                                                          \
> >> +})
> >>   #endif /* CONFIG_PARAVIRT */
> >>
> >>   #include <asm-generic/qspinlock.h>
> >> diff --git a/arch/loongarch/kernel/paravirt.c b/arch/loongarch/kernel/paravirt.c
> >> index b74fe6db49ab..2d1206e486e2 100644
> >> --- a/arch/loongarch/kernel/paravirt.c
> >> +++ b/arch/loongarch/kernel/paravirt.c
> >> @@ -10,8 +10,8 @@
> >>   #include <asm/paravirt.h>
> >>
> >>   static int has_steal_clock;
> >> -static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
> >> -static DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
> >> +DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
> >> +DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
> >>   DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
> >>
> >>   static bool steal_acc = true;
> >> @@ -261,17 +261,6 @@ static int pv_time_cpu_down_prepare(unsigned int cpu)
> >>          return 0;
> >>   }
> >>
> >> -bool vcpu_is_preempted(int cpu)
> >> -{
> >> -       struct kvm_steal_time *src;
> >> -
> >> -       if (!static_branch_unlikely(&virt_preempt_key))
> >> -               return false;
> >> -
> >> -       src = &per_cpu(steal_time, cpu);
> >> -       return !!(src->preempted & KVM_VCPU_PREEMPTED);
> >> -}
> >> -EXPORT_SYMBOL(vcpu_is_preempted);
> >>   #endif
> >>
> >>   static void pv_cpu_reboot(void *unused)
> >> --
> >> 2.39.3
> >>
> >>
>
>

Re: [PATCH 3/3] LoongArch: KVM: Set vcpu_is_preempted() macro rather than function

[Bibo Mao]

On 2026/3/11 下午12:38, Huacai Chen wrote:
> On Wed, Mar 11, 2026 at 11:51 AM Bibo Mao <maobibo@loongson.cn> wrote:
>>
>>
>>
>> On 2026/3/11 上午11:35, Huacai Chen wrote:
>>> Hi, Bibo,
>>>
>>> On Wed, Mar 11, 2026 at 11:21 AM Bibo Mao <maobibo@loongson.cn> wrote:
>>>>
>>>> vcpu_is_preempted() is performance sensitive called in function
>>>> osq_lock(), here set it as macro. So that parameter is not parsed
>>>> at most time, it can avoid cache line thrashing across numa node.
>>> I remember that I had suggested you define vcpu_is_preempted() in the
>>> header file but you refused.
>> yes, I refuse then, that is a tragedy, I am sorry for that:(
>>
>> By some benchmarks, vcpu_is_preempted() is called in busy-loop method,
>> more obvious with more CPUs, it may be a bottleneck.
> I also remember that you said the static key and percpu steal_time
> structure should be EXPORT_SYMBOL, but you haven't done this patch.
Currently API vcpu_is_preempted() and its caller owner_on_cpu() is only 
used on files in kernel/locking/mutex.c, rwsem.c etc, it seems that it 
is only used by kernel. And there is no obvious signals for module usage 
with this API.

So I do not want to export the symbol this time, will export the symbol 
if there is actually usage for modules.

Regards
Bibo Mao
> 
> Huacai
> 
>>
>> Regards
>> Bibo Mao
>>
>>>
>>> Huacai
>>>
>>>>
>>>> Here is part of unixbench result on 3C5000 DualWay machine with 32
>>>> Cores and 2 Numa node.
>>>>             origin    with patch    CONFIG_PARAVIRT disabled
>>>> execl     6871.9    7134.2        7190.8
>>>> fstime    425.5     959.9         956.1
>>>>
>>>>   From the test result, with macro method it is almost the same with
>>>> CONFIG_PARAVIRT disabled, and there is some improvment compared with
>>>> function method.
>>>>
>>>> Signed-off-by: Bibo Mao <maobibo@loongson.cn>
>>>> ---
>>>>    arch/loongarch/include/asm/qspinlock.h | 27 +++++++++++++++++++++-----
>>>>    arch/loongarch/kernel/paravirt.c       | 15 ++------------
>>>>    2 files changed, 24 insertions(+), 18 deletions(-)
>>>>
>>>> diff --git a/arch/loongarch/include/asm/qspinlock.h b/arch/loongarch/include/asm/qspinlock.h
>>>> index 66244801db67..305afd37cdcb 100644
>>>> --- a/arch/loongarch/include/asm/qspinlock.h
>>>> +++ b/arch/loongarch/include/asm/qspinlock.h
>>>> @@ -5,8 +5,10 @@
>>>>    #include <linux/jump_label.h>
>>>>
>>>>    #ifdef CONFIG_PARAVIRT
>>>> -
>>>> +#include <asm/kvm_para.h>
>>>>    DECLARE_STATIC_KEY_FALSE(virt_spin_lock_key);
>>>> +DECLARE_STATIC_KEY_FALSE(virt_preempt_key);
>>>> +DECLARE_PER_CPU(struct kvm_steal_time, steal_time);
>>>>
>>>>    #define virt_spin_lock virt_spin_lock
>>>>
>>>> @@ -34,10 +36,25 @@ static inline bool virt_spin_lock(struct qspinlock *lock)
>>>>           return true;
>>>>    }
>>>>
>>>> -#define vcpu_is_preempted vcpu_is_preempted
>>>> -
>>>> -bool vcpu_is_preempted(int cpu);
>>>> -
>>>> +/*
>>>> + * Macro is better than inline function here
>>>> + * With inline function, parameter cpu is parsed even though it is not used.
>>>> + * This may cause cache line thrashing across NUMA node.
>>>> + * With macro method, parameter cpu is parsed only when it is used.
>>>> + */
>>>> +#define vcpu_is_preempted(cpu)                                         \
>>>> +({                                                                     \
>>>> +       bool __val;                                                     \
>>>> +                                                                       \
>>>> +       if (!static_branch_unlikely(&virt_preempt_key))                 \
>>>> +               __val = false;                                          \
>>>> +       else {                                                          \
>>>> +               struct kvm_steal_time *src;                             \
>>>> +               src = &per_cpu(steal_time, cpu);                        \
>>>> +               __val = !!(src->preempted & KVM_VCPU_PREEMPTED);        \
>>>> +       }                                                               \
>>>> +       __val;                                                          \
>>>> +})
>>>>    #endif /* CONFIG_PARAVIRT */
>>>>
>>>>    #include <asm-generic/qspinlock.h>
>>>> diff --git a/arch/loongarch/kernel/paravirt.c b/arch/loongarch/kernel/paravirt.c
>>>> index b74fe6db49ab..2d1206e486e2 100644
>>>> --- a/arch/loongarch/kernel/paravirt.c
>>>> +++ b/arch/loongarch/kernel/paravirt.c
>>>> @@ -10,8 +10,8 @@
>>>>    #include <asm/paravirt.h>
>>>>
>>>>    static int has_steal_clock;
>>>> -static DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
>>>> -static DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
>>>> +DEFINE_PER_CPU(struct kvm_steal_time, steal_time) __aligned(64);
>>>> +DEFINE_STATIC_KEY_FALSE(virt_preempt_key);
>>>>    DEFINE_STATIC_KEY_FALSE(virt_spin_lock_key);
>>>>
>>>>    static bool steal_acc = true;
>>>> @@ -261,17 +261,6 @@ static int pv_time_cpu_down_prepare(unsigned int cpu)
>>>>           return 0;
>>>>    }
>>>>
>>>> -bool vcpu_is_preempted(int cpu)
>>>> -{
>>>> -       struct kvm_steal_time *src;
>>>> -
>>>> -       if (!static_branch_unlikely(&virt_preempt_key))
>>>> -               return false;
>>>> -
>>>> -       src = &per_cpu(steal_time, cpu);
>>>> -       return !!(src->preempted & KVM_VCPU_PREEMPTED);
>>>> -}
>>>> -EXPORT_SYMBOL(vcpu_is_preempted);
>>>>    #endif
>>>>
>>>>    static void pv_cpu_reboot(void *unused)
>>>> --
>>>> 2.39.3
>>>>
>>>>
>>
>>