[v1] riscv: skip csr restore if vcpu preempted reload

[PATCH] riscv: skip csr restore if vcpu preempted reload

Posted by Jinyu Tang 1 month, 4 weeks ago

The kvm_arch_vcpu_load() function is called in two cases for riscv:
1. When entering KVM_RUN from userspace ioctl.
2. When a preempted VCPU is scheduled back.

In the second case, if no other KVM VCPU has run on this CPU since the
current VCPU was preempted, the guest CSR values are still valid in
the hardware and do not need to be restored.

This patch is to skip the CSR write path when:
1. The VCPU was previously preempted
(vcpu->scheduled_out == 1).
2. It is being reloaded on the same physical CPU
(vcpu->arch.last_exit_cpu == cpu).
3. No other KVM VCPU has used this CPU in the meantime
(vcpu == __this_cpu_read(kvm_former_vcpu)).

This reduces many CSR writes with frequent preemption on the same CPU.

Signed-off-by: Jinyu Tang <tjytimi@163.com>
---
 arch/riscv/kvm/vcpu.c | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
index f001e5640..1c6c55ee1 100644
--- a/arch/riscv/kvm/vcpu.c
+++ b/arch/riscv/kvm/vcpu.c
@@ -25,6 +25,8 @@
 #define CREATE_TRACE_POINTS
 #include "trace.h"
 
+static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_former_vcpu);
+
 const struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
 	KVM_GENERIC_VCPU_STATS(),
 	STATS_DESC_COUNTER(VCPU, ecall_exit_stat),
@@ -581,6 +583,10 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 	struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr;
 	struct kvm_vcpu_config *cfg = &vcpu->arch.cfg;
 
+	if (vcpu->scheduled_out && vcpu == __this_cpu_read(kvm_former_vcpu) &&
+		vcpu->arch.last_exit_cpu == cpu)
+		goto csr_restore_done;
+
 	if (kvm_riscv_nacl_sync_csr_available()) {
 		nsh = nacl_shmem();
 		nacl_csr_write(nsh, CSR_VSSTATUS, csr->vsstatus);
@@ -624,6 +630,7 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 
 	kvm_riscv_mmu_update_hgatp(vcpu);
 
+csr_restore_done:
 	kvm_riscv_vcpu_timer_restore(vcpu);
 
 	kvm_riscv_vcpu_host_fp_save(&vcpu->arch.host_context);
@@ -645,6 +652,8 @@ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 	void *nsh;
 	struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr;
 
+	__this_cpu_write(kvm_former_vcpu, vcpu);
+
 	vcpu->cpu = -1;
 
 	kvm_riscv_vcpu_aia_put(vcpu);
-- 
2.43.0

Re: [PATCH] riscv: skip csr restore if vcpu preempted reload

Posted by Andrew Jones 1 month, 2 weeks ago

On Thu, Aug 07, 2025 at 07:42:20PM +0800, Jinyu Tang wrote:
> The kvm_arch_vcpu_load() function is called in two cases for riscv:
> 1. When entering KVM_RUN from userspace ioctl.
> 2. When a preempted VCPU is scheduled back.
> 
> In the second case, if no other KVM VCPU has run on this CPU since the
> current VCPU was preempted, the guest CSR values are still valid in
> the hardware and do not need to be restored.
> 
> This patch is to skip the CSR write path when:
> 1. The VCPU was previously preempted
> (vcpu->scheduled_out == 1).
> 2. It is being reloaded on the same physical CPU
> (vcpu->arch.last_exit_cpu == cpu).
> 3. No other KVM VCPU has used this CPU in the meantime
> (vcpu == __this_cpu_read(kvm_former_vcpu)).
> 
> This reduces many CSR writes with frequent preemption on the same CPU.
> 
> Signed-off-by: Jinyu Tang <tjytimi@163.com>
> ---
>  arch/riscv/kvm/vcpu.c | 9 +++++++++
>  1 file changed, 9 insertions(+)
> 
> diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> index f001e5640..1c6c55ee1 100644
> --- a/arch/riscv/kvm/vcpu.c
> +++ b/arch/riscv/kvm/vcpu.c
> @@ -25,6 +25,8 @@
>  #define CREATE_TRACE_POINTS
>  #include "trace.h"
>  
> +static DEFINE_PER_CPU(struct kvm_vcpu *, kvm_former_vcpu);
> +
>  const struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
>  	KVM_GENERIC_VCPU_STATS(),
>  	STATS_DESC_COUNTER(VCPU, ecall_exit_stat),
> @@ -581,6 +583,10 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
>  	struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr;
>  	struct kvm_vcpu_config *cfg = &vcpu->arch.cfg;
>  
> +	if (vcpu->scheduled_out && vcpu == __this_cpu_read(kvm_former_vcpu) &&
> +		vcpu->arch.last_exit_cpu == cpu)
> +		goto csr_restore_done;
> +
>  	if (kvm_riscv_nacl_sync_csr_available()) {
>  		nsh = nacl_shmem();
>  		nacl_csr_write(nsh, CSR_VSSTATUS, csr->vsstatus);
> @@ -624,6 +630,7 @@ void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
>  
>  	kvm_riscv_mmu_update_hgatp(vcpu);
>  
> +csr_restore_done:
>  	kvm_riscv_vcpu_timer_restore(vcpu);
>  
>  	kvm_riscv_vcpu_host_fp_save(&vcpu->arch.host_context);
> @@ -645,6 +652,8 @@ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
>  	void *nsh;
>  	struct kvm_vcpu_csr *csr = &vcpu->arch.guest_csr;
>  
> +	__this_cpu_write(kvm_former_vcpu, vcpu);
> +
>  	vcpu->cpu = -1;
>  
>  	kvm_riscv_vcpu_aia_put(vcpu);
> -- 
> 2.43.0
>

This looks like a good idea, but can't we also apply it to
kvm_riscv_vcpu_aia_load()? And, if we could track whether or
not the kernel uses FP and/or vector then we could also avoid
restoring those registers when they haven't been used.

Thanks,
drew