This patch adds KVM PMU support. We save/restore the host PMU CSR
values to kvm_context. guest entry saves the host PMU CSR and
guest exit restores the host PMU CSR when the host supports PMU
Signed-off-by: Song Gao <[email protected]>
---
arch/loongarch/include/asm/kvm_csr.h | 5 +
arch/loongarch/include/asm/kvm_host.h | 14 +++
arch/loongarch/include/asm/kvm_vcpu.h | 2 +
arch/loongarch/include/asm/loongarch.h | 1 +
arch/loongarch/kvm/exit.c | 7 ++
arch/loongarch/kvm/vcpu.c | 140 ++++++++++++++++++++++++-
6 files changed, 167 insertions(+), 2 deletions(-)
diff --git a/arch/loongarch/include/asm/kvm_csr.h b/arch/loongarch/include/asm/kvm_csr.h
index 724ca8b7b401..dce556516c79 100644
--- a/arch/loongarch/include/asm/kvm_csr.h
+++ b/arch/loongarch/include/asm/kvm_csr.h
@@ -208,4 +208,9 @@ static __always_inline void kvm_change_sw_gcsr(struct loongarch_csrs *csr,
csr->csrs[gid] |= val & _mask;
}
+#define KVM_PMU_PLV_ENABLE (CSR_PERFCTRL_PLV0 | \
+ CSR_PERFCTRL_PLV1 | \
+ CSR_PERFCTRL_PLV2 | \
+ CSR_PERFCTRL_PLV3)
+
#endif /* __ASM_LOONGARCH_KVM_CSR_H__ */
diff --git a/arch/loongarch/include/asm/kvm_host.h b/arch/loongarch/include/asm/kvm_host.h
index 2d62f7b0d377..0daa0bde7742 100644
--- a/arch/loongarch/include/asm/kvm_host.h
+++ b/arch/loongarch/include/asm/kvm_host.h
@@ -54,6 +54,9 @@ struct kvm_arch_memory_slot {
struct kvm_context {
unsigned long vpid_cache;
struct kvm_vcpu *last_vcpu;
+ /* Save host pmu csr */
+ u64 perf_ctrl[4];
+ u64 perf_cntr[4];
};
struct kvm_world_switch {
@@ -99,6 +102,7 @@ enum emulation_result {
#define KVM_LARCH_LASX (0x1 << 2)
#define KVM_LARCH_SWCSR_LATEST (0x1 << 3)
#define KVM_LARCH_HWCSR_USABLE (0x1 << 4)
+#define KVM_LARCH_PERF (0x1 << 5)
struct kvm_vcpu_arch {
/*
@@ -195,6 +199,16 @@ static inline bool kvm_guest_has_lasx(struct kvm_vcpu_arch *arch)
return arch->cpucfg[2] & CPUCFG2_LASX;
}
+static inline bool kvm_guest_has_pmu(struct kvm_vcpu_arch *arch)
+{
+ return arch->cpucfg[6] & CPUCFG6_PMP;
+}
+
+static inline int kvm_get_pmu_num(struct kvm_vcpu_arch *arch)
+{
+ return (arch->cpucfg[6] & CPUCFG6_PMNUM) >> CPUCFG6_PMNUM_SHIFT;
+}
+
/* Debug: dump vcpu state */
int kvm_arch_vcpu_dump_regs(struct kvm_vcpu *vcpu);
diff --git a/arch/loongarch/include/asm/kvm_vcpu.h b/arch/loongarch/include/asm/kvm_vcpu.h
index 0cb4fdb8a9b5..7212428022d1 100644
--- a/arch/loongarch/include/asm/kvm_vcpu.h
+++ b/arch/loongarch/include/asm/kvm_vcpu.h
@@ -75,6 +75,8 @@ static inline void kvm_save_lasx(struct loongarch_fpu *fpu) { }
static inline void kvm_restore_lasx(struct loongarch_fpu *fpu) { }
#endif
+int kvm_own_pmu(struct kvm_vcpu *vcpu);
+
void kvm_init_timer(struct kvm_vcpu *vcpu, unsigned long hz);
void kvm_reset_timer(struct kvm_vcpu *vcpu);
void kvm_save_timer(struct kvm_vcpu *vcpu);
diff --git a/arch/loongarch/include/asm/loongarch.h b/arch/loongarch/include/asm/loongarch.h
index 46366e783c84..644380b6ebec 100644
--- a/arch/loongarch/include/asm/loongarch.h
+++ b/arch/loongarch/include/asm/loongarch.h
@@ -119,6 +119,7 @@
#define CPUCFG6_PMP BIT(0)
#define CPUCFG6_PAMVER GENMASK(3, 1)
#define CPUCFG6_PMNUM GENMASK(7, 4)
+#define CPUCFG6_PMNUM_SHIFT 4
#define CPUCFG6_PMBITS GENMASK(13, 8)
#define CPUCFG6_UPM BIT(14)
diff --git a/arch/loongarch/kvm/exit.c b/arch/loongarch/kvm/exit.c
index ed1d89d53e2e..ab17d9eb711f 100644
--- a/arch/loongarch/kvm/exit.c
+++ b/arch/loongarch/kvm/exit.c
@@ -83,6 +83,13 @@ static int kvm_handle_csr(struct kvm_vcpu *vcpu, larch_inst inst)
rj = inst.reg2csr_format.rj;
csrid = inst.reg2csr_format.csr;
+ if (LOONGARCH_CSR_PERFCTRL0 <= csrid && csrid <= LOONGARCH_CSR_PERFCNTR3) {
+ if(!kvm_own_pmu(vcpu)) {
+ vcpu->arch.pc -= 4;
+ return EMULATE_DONE;
+ }
+ }
+
/* Process CSR ops */
switch (rj) {
case 0: /* process csrrd */
diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
index 3a8779065f73..6fb59b805807 100644
--- a/arch/loongarch/kvm/vcpu.c
+++ b/arch/loongarch/kvm/vcpu.c
@@ -333,6 +333,12 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
case LOONGARCH_CPUCFG5:
*v = GENMASK(31, 0);
return 0;
+ case LOONGARCH_CPUCFG6:
+ if (cpu_has_pmp)
+ *v = GENMASK(14, 0);
+ else
+ *v = 0;
+ return 0;
case LOONGARCH_CPUCFG16:
*v = GENMASK(16, 0);
return 0;
@@ -351,7 +357,7 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
static int kvm_check_cpucfg(int id, u64 val)
{
- int ret;
+ int ret, host;
u64 mask = 0;
ret = _kvm_get_cpucfg_mask(id, &mask);
@@ -377,6 +383,18 @@ static int kvm_check_cpucfg(int id, u64 val)
/* LASX architecturally implies LSX and FP but val does not satisfy that */
return -EINVAL;
return 0;
+ case LOONGARCH_CPUCFG6:
+ if (val & CPUCFG6_PMP) {
+ host = read_cpucfg(LOONGARCH_CPUCFG6);
+ if ((val & CPUCFG6_PMBITS) != (host & CPUCFG6_PMBITS))
+ /* Guest pmbits must be the same with host */
+ return -EINVAL;
+ if ((val & CPUCFG6_PMNUM) > (host & CPUCFG6_PMNUM))
+ return -EINVAL;
+ if ((val & CPUCFG6_UPM) && !(host & CPUCFG6_UPM))
+ return -EINVAL;
+ }
+ return 0;
default:
/*
* Values for the other CPUCFG IDs are not being further validated
@@ -552,7 +570,8 @@ static int kvm_loongarch_cpucfg_has_attr(struct kvm_vcpu *vcpu,
struct kvm_device_attr *attr)
{
switch (attr->attr) {
- case 2:
+ case LOONGARCH_CPUCFG2:
+ case LOONGARCH_CPUCFG6:
return 0;
default:
return -ENXIO;
@@ -844,6 +863,119 @@ void kvm_lose_fpu(struct kvm_vcpu *vcpu)
preempt_enable();
}
+static inline void kvm_save_host_pmu(struct kvm_vcpu *vcpu)
+{
+ struct kvm_context *context;
+
+ context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
+ context->perf_ctrl[0] = read_csr_perfctrl0();
+ context->perf_cntr[0] = read_csr_perfcntr0();
+ context->perf_ctrl[1] = read_csr_perfctrl1();
+ context->perf_cntr[1] = read_csr_perfcntr1();
+ context->perf_ctrl[2] = read_csr_perfctrl2();
+ context->perf_cntr[2] = read_csr_perfcntr2();
+ context->perf_ctrl[3] = read_csr_perfctrl3();
+ context->perf_cntr[3] = read_csr_perfcntr3();
+}
+
+static inline void kvm_restore_host_pmu(struct kvm_vcpu *vcpu)
+{
+ struct kvm_context *context;
+
+ context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
+ write_csr_perfctrl0(context->perf_ctrl[0]);
+ write_csr_perfcntr0(context->perf_cntr[0]);
+ write_csr_perfctrl1(context->perf_ctrl[1]);
+ write_csr_perfcntr1(context->perf_cntr[1]);
+ write_csr_perfctrl2(context->perf_ctrl[2]);
+ write_csr_perfcntr2(context->perf_cntr[2]);
+ write_csr_perfctrl3(context->perf_ctrl[3]);
+ write_csr_perfcntr3(context->perf_cntr[3]);
+}
+
+static inline void kvm_save_guest_pmu(struct kvm_vcpu *vcpu)
+{
+ struct loongarch_csrs *csr = vcpu->arch.csr;
+
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
+ kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
+ kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL0, 0);
+ kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL1, 0);
+ kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL2, 0);
+ kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL3, 0);
+}
+
+static inline void kvm_restore_guest_pmu(struct kvm_vcpu *vcpu)
+{
+ struct loongarch_csrs *csr = vcpu->arch.csr;
+
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
+ kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
+}
+
+static void _kvm_own_pmu(struct kvm_vcpu *vcpu)
+{
+ unsigned long val;
+
+ kvm_save_host_pmu(vcpu);
+ /* Set PM0-PM(num) to Guest */
+ val = read_csr_gcfg() & ~CSR_GCFG_GPERF;
+ val |= (kvm_get_pmu_num(&vcpu->arch) + 1) << CSR_GCFG_GPERF_SHIFT;
+ write_csr_gcfg(val);
+ kvm_restore_guest_pmu(vcpu);
+}
+
+int kvm_own_pmu(struct kvm_vcpu *vcpu)
+{
+ if (!kvm_guest_has_pmu(&vcpu->arch))
+ return -EINVAL;
+
+ preempt_disable();
+ _kvm_own_pmu(vcpu);
+ vcpu->arch.aux_inuse |= KVM_LARCH_PERF;
+ preempt_enable();
+ return 0;
+}
+
+static void kvm_lose_pmu(struct kvm_vcpu *vcpu)
+{
+ struct loongarch_csrs *csr = vcpu->arch.csr;
+
+ if (!(vcpu->arch.aux_inuse & KVM_LARCH_PERF))
+ return;
+
+ kvm_save_guest_pmu(vcpu);
+ /* Disable pmu access from guest */
+ write_csr_gcfg(read_csr_gcfg() & ~CSR_GCFG_GPERF);
+ if (((kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0) |
+ kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1) |
+ kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2) |
+ kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3))
+ & KVM_PMU_PLV_ENABLE) == 0)
+ vcpu->arch.aux_inuse &= ~KVM_LARCH_PERF;
+ kvm_restore_host_pmu(vcpu);
+}
+
+static void kvm_restore_pmu(struct kvm_vcpu *vcpu)
+{
+ if (!(vcpu->arch.aux_inuse & KVM_LARCH_PERF))
+ return;
+
+ _kvm_own_pmu(vcpu);
+}
+
int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu, struct kvm_interrupt *irq)
{
int intr = (int)irq->irq;
@@ -982,6 +1114,9 @@ static int _kvm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
/* Control guest page CCA attribute */
change_csr_gcfg(CSR_GCFG_MATC_MASK, CSR_GCFG_MATC_ROOT);
+ /* Restore hardware perf csr */
+ kvm_restore_pmu(vcpu);
+
/* Don't bother restoring registers multiple times unless necessary */
if (vcpu->arch.aux_inuse & KVM_LARCH_HWCSR_USABLE)
return 0;
@@ -1065,6 +1200,7 @@ static int _kvm_vcpu_put(struct kvm_vcpu *vcpu, int cpu)
struct loongarch_csrs *csr = vcpu->arch.csr;
kvm_lose_fpu(vcpu);
+ kvm_lose_pmu(vcpu);
/*
* Update CSR state from hardware if software CSR state is stale,
--
2.39.3
On 2024/4/10 下午5:58, Song Gao wrote:
> This patch adds KVM PMU support. We save/restore the host PMU CSR
> values to kvm_context. guest entry saves the host PMU CSR and
> guest exit restores the host PMU CSR when the host supports PMU
It will be better if there is description about hw/sw pmu virtualization
method, such as is it shared or separated with host pmu hardware, or is
it software emulated?
>
> Signed-off-by: Song Gao <[email protected]>
> ---
> arch/loongarch/include/asm/kvm_csr.h | 5 +
> arch/loongarch/include/asm/kvm_host.h | 14 +++
> arch/loongarch/include/asm/kvm_vcpu.h | 2 +
> arch/loongarch/include/asm/loongarch.h | 1 +
> arch/loongarch/kvm/exit.c | 7 ++
> arch/loongarch/kvm/vcpu.c | 140 ++++++++++++++++++++++++-
> 6 files changed, 167 insertions(+), 2 deletions(-)
>
> diff --git a/arch/loongarch/include/asm/kvm_csr.h b/arch/loongarch/include/asm/kvm_csr.h
> index 724ca8b7b401..dce556516c79 100644
> --- a/arch/loongarch/include/asm/kvm_csr.h
> +++ b/arch/loongarch/include/asm/kvm_csr.h
> @@ -208,4 +208,9 @@ static __always_inline void kvm_change_sw_gcsr(struct loongarch_csrs *csr,
> csr->csrs[gid] |= val & _mask;
> }
>
> +#define KVM_PMU_PLV_ENABLE (CSR_PERFCTRL_PLV0 | \
> + CSR_PERFCTRL_PLV1 | \
> + CSR_PERFCTRL_PLV2 | \
> + CSR_PERFCTRL_PLV3)
> +
How about KVM_PMU_ENABLED or KVM_PMU_EVENT_ENABLED?
> #endif /* __ASM_LOONGARCH_KVM_CSR_H__ */
> diff --git a/arch/loongarch/include/asm/kvm_host.h b/arch/loongarch/include/asm/kvm_host.h
> index 2d62f7b0d377..0daa0bde7742 100644
> --- a/arch/loongarch/include/asm/kvm_host.h
> +++ b/arch/loongarch/include/asm/kvm_host.h
> @@ -54,6 +54,9 @@ struct kvm_arch_memory_slot {
> struct kvm_context {
> unsigned long vpid_cache;
> struct kvm_vcpu *last_vcpu;
> + /* Save host pmu csr */
> + u64 perf_ctrl[4];
> + u64 perf_cntr[4];
It will be better if macro is used rather than hardcoded number 4.
> };
>
> struct kvm_world_switch {
> @@ -99,6 +102,7 @@ enum emulation_result {
> #define KVM_LARCH_LASX (0x1 << 2)
> #define KVM_LARCH_SWCSR_LATEST (0x1 << 3)
> #define KVM_LARCH_HWCSR_USABLE (0x1 << 4)
> +#define KVM_LARCH_PERF (0x1 << 5)
>
> struct kvm_vcpu_arch {
> /*
> @@ -195,6 +199,16 @@ static inline bool kvm_guest_has_lasx(struct kvm_vcpu_arch *arch)
> return arch->cpucfg[2] & CPUCFG2_LASX;
> }
>
> +static inline bool kvm_guest_has_pmu(struct kvm_vcpu_arch *arch)
> +{
> + return arch->cpucfg[6] & CPUCFG6_PMP;
macro LOONGARCH_CPUCFG6 will be better.
> +}
> +
> +static inline int kvm_get_pmu_num(struct kvm_vcpu_arch *arch)
> +{
> + return (arch->cpucfg[6] & CPUCFG6_PMNUM) >> CPUCFG6_PMNUM_SHIFT;
Ditto
Regards
Bibo Mao
> +}
> +
> /* Debug: dump vcpu state */
> int kvm_arch_vcpu_dump_regs(struct kvm_vcpu *vcpu);
>
> diff --git a/arch/loongarch/include/asm/kvm_vcpu.h b/arch/loongarch/include/asm/kvm_vcpu.h
> index 0cb4fdb8a9b5..7212428022d1 100644
> --- a/arch/loongarch/include/asm/kvm_vcpu.h
> +++ b/arch/loongarch/include/asm/kvm_vcpu.h
> @@ -75,6 +75,8 @@ static inline void kvm_save_lasx(struct loongarch_fpu *fpu) { }
> static inline void kvm_restore_lasx(struct loongarch_fpu *fpu) { }
> #endif
>
> +int kvm_own_pmu(struct kvm_vcpu *vcpu);
> +
> void kvm_init_timer(struct kvm_vcpu *vcpu, unsigned long hz);
> void kvm_reset_timer(struct kvm_vcpu *vcpu);
> void kvm_save_timer(struct kvm_vcpu *vcpu);
> diff --git a/arch/loongarch/include/asm/loongarch.h b/arch/loongarch/include/asm/loongarch.h
> index 46366e783c84..644380b6ebec 100644
> --- a/arch/loongarch/include/asm/loongarch.h
> +++ b/arch/loongarch/include/asm/loongarch.h
> @@ -119,6 +119,7 @@
> #define CPUCFG6_PMP BIT(0)
> #define CPUCFG6_PAMVER GENMASK(3, 1)
> #define CPUCFG6_PMNUM GENMASK(7, 4)
> +#define CPUCFG6_PMNUM_SHIFT 4
> #define CPUCFG6_PMBITS GENMASK(13, 8)
> #define CPUCFG6_UPM BIT(14)
>
> diff --git a/arch/loongarch/kvm/exit.c b/arch/loongarch/kvm/exit.c
> index ed1d89d53e2e..ab17d9eb711f 100644
> --- a/arch/loongarch/kvm/exit.c
> +++ b/arch/loongarch/kvm/exit.c
> @@ -83,6 +83,13 @@ static int kvm_handle_csr(struct kvm_vcpu *vcpu, larch_inst inst)
> rj = inst.reg2csr_format.rj;
> csrid = inst.reg2csr_format.csr;
>
> + if (LOONGARCH_CSR_PERFCTRL0 <= csrid && csrid <= LOONGARCH_CSR_PERFCNTR3) {
> + if(!kvm_own_pmu(vcpu)) {
> + vcpu->arch.pc -= 4;
> + return EMULATE_DONE;
> + }
> + }
> +
> /* Process CSR ops */
> switch (rj) {
> case 0: /* process csrrd */
> diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
> index 3a8779065f73..6fb59b805807 100644
> --- a/arch/loongarch/kvm/vcpu.c
> +++ b/arch/loongarch/kvm/vcpu.c
> @@ -333,6 +333,12 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
> case LOONGARCH_CPUCFG5:
> *v = GENMASK(31, 0);
> return 0;
> + case LOONGARCH_CPUCFG6:
> + if (cpu_has_pmp)
> + *v = GENMASK(14, 0);
> + else
> + *v = 0;
> + return 0;
> case LOONGARCH_CPUCFG16:
> *v = GENMASK(16, 0);
> return 0;
> @@ -351,7 +357,7 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
>
> static int kvm_check_cpucfg(int id, u64 val)
> {
> - int ret;
> + int ret, host;
> u64 mask = 0;
>
> ret = _kvm_get_cpucfg_mask(id, &mask);
> @@ -377,6 +383,18 @@ static int kvm_check_cpucfg(int id, u64 val)
> /* LASX architecturally implies LSX and FP but val does not satisfy that */
> return -EINVAL;
> return 0;
> + case LOONGARCH_CPUCFG6:
> + if (val & CPUCFG6_PMP) {
> + host = read_cpucfg(LOONGARCH_CPUCFG6);
> + if ((val & CPUCFG6_PMBITS) != (host & CPUCFG6_PMBITS))
> + /* Guest pmbits must be the same with host */
> + return -EINVAL;
> + if ((val & CPUCFG6_PMNUM) > (host & CPUCFG6_PMNUM))
> + return -EINVAL;
> + if ((val & CPUCFG6_UPM) && !(host & CPUCFG6_UPM))
> + return -EINVAL;
> + }
> + return 0;
> default:
> /*
> * Values for the other CPUCFG IDs are not being further validated
> @@ -552,7 +570,8 @@ static int kvm_loongarch_cpucfg_has_attr(struct kvm_vcpu *vcpu,
> struct kvm_device_attr *attr)
> {
> switch (attr->attr) {
> - case 2:
> + case LOONGARCH_CPUCFG2:
> + case LOONGARCH_CPUCFG6:
> return 0;
> default:
> return -ENXIO;
> @@ -844,6 +863,119 @@ void kvm_lose_fpu(struct kvm_vcpu *vcpu)
> preempt_enable();
> }
>
> +static inline void kvm_save_host_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct kvm_context *context;
> +
> + context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
> + context->perf_ctrl[0] = read_csr_perfctrl0();
> + context->perf_cntr[0] = read_csr_perfcntr0();
> + context->perf_ctrl[1] = read_csr_perfctrl1();
> + context->perf_cntr[1] = read_csr_perfcntr1();
> + context->perf_ctrl[2] = read_csr_perfctrl2();
> + context->perf_cntr[2] = read_csr_perfcntr2();
> + context->perf_ctrl[3] = read_csr_perfctrl3();
> + context->perf_cntr[3] = read_csr_perfcntr3();
> +}
> +
> +static inline void kvm_restore_host_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct kvm_context *context;
> +
> + context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
> + write_csr_perfctrl0(context->perf_ctrl[0]);
> + write_csr_perfcntr0(context->perf_cntr[0]);
> + write_csr_perfctrl1(context->perf_ctrl[1]);
> + write_csr_perfcntr1(context->perf_cntr[1]);
> + write_csr_perfctrl2(context->perf_ctrl[2]);
> + write_csr_perfcntr2(context->perf_cntr[2]);
> + write_csr_perfctrl3(context->perf_ctrl[3]);
> + write_csr_perfcntr3(context->perf_cntr[3]);
> +}
> +
> +static inline void kvm_save_guest_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct loongarch_csrs *csr = vcpu->arch.csr;
> +
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL0, 0);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL1, 0);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL2, 0);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL3, 0);
> +}
> +
> +static inline void kvm_restore_guest_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct loongarch_csrs *csr = vcpu->arch.csr;
> +
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
> +}
> +
> +static void _kvm_own_pmu(struct kvm_vcpu *vcpu)
> +{
> + unsigned long val;
> +
> + kvm_save_host_pmu(vcpu);
> + /* Set PM0-PM(num) to Guest */
> + val = read_csr_gcfg() & ~CSR_GCFG_GPERF;
> + val |= (kvm_get_pmu_num(&vcpu->arch) + 1) << CSR_GCFG_GPERF_SHIFT;
> + write_csr_gcfg(val);
> + kvm_restore_guest_pmu(vcpu);
> +}
> +
> +int kvm_own_pmu(struct kvm_vcpu *vcpu)
> +{
> + if (!kvm_guest_has_pmu(&vcpu->arch))
> + return -EINVAL;
> +
> + preempt_disable();
> + _kvm_own_pmu(vcpu);
> + vcpu->arch.aux_inuse |= KVM_LARCH_PERF;
> + preempt_enable();
> + return 0;
> +}
> +
> +static void kvm_lose_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct loongarch_csrs *csr = vcpu->arch.csr;
> +
> + if (!(vcpu->arch.aux_inuse & KVM_LARCH_PERF))
> + return;
> +
> + kvm_save_guest_pmu(vcpu);
> + /* Disable pmu access from guest */
> + write_csr_gcfg(read_csr_gcfg() & ~CSR_GCFG_GPERF);
> + if (((kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0) |
> + kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1) |
> + kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2) |
> + kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3))
> + & KVM_PMU_PLV_ENABLE) == 0)
> + vcpu->arch.aux_inuse &= ~KVM_LARCH_PERF;
> + kvm_restore_host_pmu(vcpu);
> +}
> +
> +static void kvm_restore_pmu(struct kvm_vcpu *vcpu)
> +{
> + if (!(vcpu->arch.aux_inuse & KVM_LARCH_PERF))
> + return;
> +
> + _kvm_own_pmu(vcpu);
> +}
> +
> int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu, struct kvm_interrupt *irq)
> {
> int intr = (int)irq->irq;
> @@ -982,6 +1114,9 @@ static int _kvm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> /* Control guest page CCA attribute */
> change_csr_gcfg(CSR_GCFG_MATC_MASK, CSR_GCFG_MATC_ROOT);
>
> + /* Restore hardware perf csr */
> + kvm_restore_pmu(vcpu);
> +
> /* Don't bother restoring registers multiple times unless necessary */
> if (vcpu->arch.aux_inuse & KVM_LARCH_HWCSR_USABLE)
> return 0;
> @@ -1065,6 +1200,7 @@ static int _kvm_vcpu_put(struct kvm_vcpu *vcpu, int cpu)
> struct loongarch_csrs *csr = vcpu->arch.csr;
>
> kvm_lose_fpu(vcpu);
> + kvm_lose_pmu(vcpu);
>
> /*
> * Update CSR state from hardware if software CSR state is stale,
>
On 2024/4/10 下午5:58, Song Gao wrote:
> This patch adds KVM PMU support. We save/restore the host PMU CSR
> values to kvm_context. guest entry saves the host PMU CSR and
> guest exit restores the host PMU CSR when the host supports PMU
>
> Signed-off-by: Song Gao <[email protected]>
> ---
> arch/loongarch/include/asm/kvm_csr.h | 5 +
> arch/loongarch/include/asm/kvm_host.h | 14 +++
> arch/loongarch/include/asm/kvm_vcpu.h | 2 +
> arch/loongarch/include/asm/loongarch.h | 1 +
> arch/loongarch/kvm/exit.c | 7 ++
> arch/loongarch/kvm/vcpu.c | 140 ++++++++++++++++++++++++-
> 6 files changed, 167 insertions(+), 2 deletions(-)
>
> diff --git a/arch/loongarch/include/asm/kvm_csr.h b/arch/loongarch/include/asm/kvm_csr.h
> index 724ca8b7b401..dce556516c79 100644
> --- a/arch/loongarch/include/asm/kvm_csr.h
> +++ b/arch/loongarch/include/asm/kvm_csr.h
> @@ -208,4 +208,9 @@ static __always_inline void kvm_change_sw_gcsr(struct loongarch_csrs *csr,
> csr->csrs[gid] |= val & _mask;
> }
>
> +#define KVM_PMU_PLV_ENABLE (CSR_PERFCTRL_PLV0 | \
> + CSR_PERFCTRL_PLV1 | \
> + CSR_PERFCTRL_PLV2 | \
> + CSR_PERFCTRL_PLV3)
> +
> #endif /* __ASM_LOONGARCH_KVM_CSR_H__ */
> diff --git a/arch/loongarch/include/asm/kvm_host.h b/arch/loongarch/include/asm/kvm_host.h
> index 2d62f7b0d377..0daa0bde7742 100644
> --- a/arch/loongarch/include/asm/kvm_host.h
> +++ b/arch/loongarch/include/asm/kvm_host.h
> @@ -54,6 +54,9 @@ struct kvm_arch_memory_slot {
> struct kvm_context {
> unsigned long vpid_cache;
> struct kvm_vcpu *last_vcpu;
> + /* Save host pmu csr */
> + u64 perf_ctrl[4];
> + u64 perf_cntr[4];
> };
>
> struct kvm_world_switch {
> @@ -99,6 +102,7 @@ enum emulation_result {
> #define KVM_LARCH_LASX (0x1 << 2)
> #define KVM_LARCH_SWCSR_LATEST (0x1 << 3)
> #define KVM_LARCH_HWCSR_USABLE (0x1 << 4)
> +#define KVM_LARCH_PERF (0x1 << 5)
>
> struct kvm_vcpu_arch {
> /*
> @@ -195,6 +199,16 @@ static inline bool kvm_guest_has_lasx(struct kvm_vcpu_arch *arch)
> return arch->cpucfg[2] & CPUCFG2_LASX;
> }
>
> +static inline bool kvm_guest_has_pmu(struct kvm_vcpu_arch *arch)
> +{
> + return arch->cpucfg[6] & CPUCFG6_PMP;
> +}
> +
> +static inline int kvm_get_pmu_num(struct kvm_vcpu_arch *arch)
> +{
> + return (arch->cpucfg[6] & CPUCFG6_PMNUM) >> CPUCFG6_PMNUM_SHIFT;
> +}
> +
> /* Debug: dump vcpu state */
> int kvm_arch_vcpu_dump_regs(struct kvm_vcpu *vcpu);
>
> diff --git a/arch/loongarch/include/asm/kvm_vcpu.h b/arch/loongarch/include/asm/kvm_vcpu.h
> index 0cb4fdb8a9b5..7212428022d1 100644
> --- a/arch/loongarch/include/asm/kvm_vcpu.h
> +++ b/arch/loongarch/include/asm/kvm_vcpu.h
> @@ -75,6 +75,8 @@ static inline void kvm_save_lasx(struct loongarch_fpu *fpu) { }
> static inline void kvm_restore_lasx(struct loongarch_fpu *fpu) { }
> #endif
>
> +int kvm_own_pmu(struct kvm_vcpu *vcpu);
> +
> void kvm_init_timer(struct kvm_vcpu *vcpu, unsigned long hz);
> void kvm_reset_timer(struct kvm_vcpu *vcpu);
> void kvm_save_timer(struct kvm_vcpu *vcpu);
> diff --git a/arch/loongarch/include/asm/loongarch.h b/arch/loongarch/include/asm/loongarch.h
> index 46366e783c84..644380b6ebec 100644
> --- a/arch/loongarch/include/asm/loongarch.h
> +++ b/arch/loongarch/include/asm/loongarch.h
> @@ -119,6 +119,7 @@
> #define CPUCFG6_PMP BIT(0)
> #define CPUCFG6_PAMVER GENMASK(3, 1)
> #define CPUCFG6_PMNUM GENMASK(7, 4)
> +#define CPUCFG6_PMNUM_SHIFT 4
> #define CPUCFG6_PMBITS GENMASK(13, 8)
> #define CPUCFG6_UPM BIT(14)
>
> diff --git a/arch/loongarch/kvm/exit.c b/arch/loongarch/kvm/exit.c
> index ed1d89d53e2e..ab17d9eb711f 100644
> --- a/arch/loongarch/kvm/exit.c
> +++ b/arch/loongarch/kvm/exit.c
> @@ -83,6 +83,13 @@ static int kvm_handle_csr(struct kvm_vcpu *vcpu, larch_inst inst)
> rj = inst.reg2csr_format.rj;
> csrid = inst.reg2csr_format.csr;
>
> + if (LOONGARCH_CSR_PERFCTRL0 <= csrid && csrid <= LOONGARCH_CSR_PERFCNTR3) {
> + if(!kvm_own_pmu(vcpu)) {
> + vcpu->arch.pc -= 4;
> + return EMULATE_DONE;
> + }
> + }
> +
> /* Process CSR ops */
> switch (rj) {
> case 0: /* process csrrd */
> diff --git a/arch/loongarch/kvm/vcpu.c b/arch/loongarch/kvm/vcpu.c
> index 3a8779065f73..6fb59b805807 100644
> --- a/arch/loongarch/kvm/vcpu.c
> +++ b/arch/loongarch/kvm/vcpu.c
> @@ -333,6 +333,12 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
> case LOONGARCH_CPUCFG5:
> *v = GENMASK(31, 0);
> return 0;
> + case LOONGARCH_CPUCFG6:
> + if (cpu_has_pmp)
> + *v = GENMASK(14, 0);
> + else
> + *v = 0;
> + return 0;
> case LOONGARCH_CPUCFG16:
> *v = GENMASK(16, 0);
> return 0;
> @@ -351,7 +357,7 @@ static int _kvm_get_cpucfg_mask(int id, u64 *v)
>
> static int kvm_check_cpucfg(int id, u64 val)
> {
> - int ret;
> + int ret, host;
> u64 mask = 0;
>
> ret = _kvm_get_cpucfg_mask(id, &mask);
> @@ -377,6 +383,18 @@ static int kvm_check_cpucfg(int id, u64 val)
> /* LASX architecturally implies LSX and FP but val does not satisfy that */
> return -EINVAL;
> return 0;
> + case LOONGARCH_CPUCFG6:
> + if (val & CPUCFG6_PMP) {
> + host = read_cpucfg(LOONGARCH_CPUCFG6);
> + if ((val & CPUCFG6_PMBITS) != (host & CPUCFG6_PMBITS))
> + /* Guest pmbits must be the same with host */
> + return -EINVAL;
> + if ((val & CPUCFG6_PMNUM) > (host & CPUCFG6_PMNUM))
> + return -EINVAL;
> + if ((val & CPUCFG6_UPM) && !(host & CPUCFG6_UPM))
> + return -EINVAL;
> + }
> + return 0;
> default:
> /*
> * Values for the other CPUCFG IDs are not being further validated
> @@ -552,7 +570,8 @@ static int kvm_loongarch_cpucfg_has_attr(struct kvm_vcpu *vcpu,
> struct kvm_device_attr *attr)
> {
> switch (attr->attr) {
> - case 2:
> + case LOONGARCH_CPUCFG2:
> + case LOONGARCH_CPUCFG6:
> return 0;
> default:
> return -ENXIO;
> @@ -844,6 +863,119 @@ void kvm_lose_fpu(struct kvm_vcpu *vcpu)
> preempt_enable();
> }
>
> +static inline void kvm_save_host_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct kvm_context *context;
> +
> + context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
> + context->perf_ctrl[0] = read_csr_perfctrl0();
> + context->perf_cntr[0] = read_csr_perfcntr0();
> + context->perf_ctrl[1] = read_csr_perfctrl1();
> + context->perf_cntr[1] = read_csr_perfcntr1();
> + context->perf_ctrl[2] = read_csr_perfctrl2();
> + context->perf_cntr[2] = read_csr_perfcntr2();
> + context->perf_ctrl[3] = read_csr_perfctrl3();
> + context->perf_cntr[3] = read_csr_perfcntr3();
> +}
> +
> +static inline void kvm_restore_host_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct kvm_context *context;
> +
> + context = this_cpu_ptr(vcpu->kvm->arch.vmcs);
> + write_csr_perfctrl0(context->perf_ctrl[0]);
> + write_csr_perfcntr0(context->perf_cntr[0]);
> + write_csr_perfctrl1(context->perf_ctrl[1]);
> + write_csr_perfcntr1(context->perf_cntr[1]);
> + write_csr_perfctrl2(context->perf_ctrl[2]);
> + write_csr_perfcntr2(context->perf_cntr[2]);
> + write_csr_perfctrl3(context->perf_ctrl[3]);
> + write_csr_perfcntr3(context->perf_cntr[3]);
> +}
> +
> +static inline void kvm_save_guest_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct loongarch_csrs *csr = vcpu->arch.csr;
> +
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
> + kvm_save_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL0, 0);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL1, 0);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL2, 0);
> + kvm_write_hw_gcsr(LOONGARCH_CSR_PERFCTRL3, 0);
> +}
> +
> +static inline void kvm_restore_guest_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct loongarch_csrs *csr = vcpu->arch.csr;
> +
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR0);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR1);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR2);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3);
> + kvm_restore_hw_gcsr(csr, LOONGARCH_CSR_PERFCNTR3);
> +}
> +
> +static void _kvm_own_pmu(struct kvm_vcpu *vcpu)
> +{
> + unsigned long val;
> +
> + kvm_save_host_pmu(vcpu);
> + /* Set PM0-PM(num) to Guest */
> + val = read_csr_gcfg() & ~CSR_GCFG_GPERF;
> + val |= (kvm_get_pmu_num(&vcpu->arch) + 1) << CSR_GCFG_GPERF_SHIFT;
> + write_csr_gcfg(val);
> + kvm_restore_guest_pmu(vcpu);
> +}
> +
> +int kvm_own_pmu(struct kvm_vcpu *vcpu)
> +{
> + if (!kvm_guest_has_pmu(&vcpu->arch))
> + return -EINVAL;
> +
> + preempt_disable();
> + _kvm_own_pmu(vcpu);
> + vcpu->arch.aux_inuse |= KVM_LARCH_PERF;
> + preempt_enable();
Disable preemption is not safe enough, maybe there will be host pmu
interrupt happened where host pmu registers are modified in pmu
interrupt handler. It will be better if KVM_REQ_PMU is added, host pmu
register is saved just before entering guest. Like this:
kvm_make_request(KVM_REQ_PMU, vcpu);
And check pmu request in function kvm_pre_enter_guest() with irq disabled.
Regards
Bibo Mao
> + return 0;
> +}
> +
> +static void kvm_lose_pmu(struct kvm_vcpu *vcpu)
> +{
> + struct loongarch_csrs *csr = vcpu->arch.csr;
> +
> + if (!(vcpu->arch.aux_inuse & KVM_LARCH_PERF))
> + return;
> +
> + kvm_save_guest_pmu(vcpu);
> + /* Disable pmu access from guest */
> + write_csr_gcfg(read_csr_gcfg() & ~CSR_GCFG_GPERF);
> + if (((kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL0) |
> + kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL1) |
> + kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL2) |
> + kvm_read_sw_gcsr(csr, LOONGARCH_CSR_PERFCTRL3))
> + & KVM_PMU_PLV_ENABLE) == 0)
> + vcpu->arch.aux_inuse &= ~KVM_LARCH_PERF;
> + kvm_restore_host_pmu(vcpu);
> +}
> +
> +static void kvm_restore_pmu(struct kvm_vcpu *vcpu)
> +{
> + if (!(vcpu->arch.aux_inuse & KVM_LARCH_PERF))
> + return;
> +
> + _kvm_own_pmu(vcpu);
> +}
> +
> int kvm_vcpu_ioctl_interrupt(struct kvm_vcpu *vcpu, struct kvm_interrupt *irq)
> {
> int intr = (int)irq->irq;
> @@ -982,6 +1114,9 @@ static int _kvm_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
> /* Control guest page CCA attribute */
> change_csr_gcfg(CSR_GCFG_MATC_MASK, CSR_GCFG_MATC_ROOT);
>
> + /* Restore hardware perf csr */
> + kvm_restore_pmu(vcpu);
> +
> /* Don't bother restoring registers multiple times unless necessary */
> if (vcpu->arch.aux_inuse & KVM_LARCH_HWCSR_USABLE)
> return 0;
> @@ -1065,6 +1200,7 @@ static int _kvm_vcpu_put(struct kvm_vcpu *vcpu, int cpu)
> struct loongarch_csrs *csr = vcpu->arch.csr;
>
> kvm_lose_fpu(vcpu);
> + kvm_lose_pmu(vcpu);
>
> /*
> * Update CSR state from hardware if software CSR state is stale,
>