[v4,07/14] RISC-V: KVM: Add skeleton support for perf

Commit Message

Atish Kumar Patra Feb. 1, 2023, 11:12 p.m. UTC

This patch only adds barebone structure of perf implementation. Most of
the function returns zero at this point and will be implemented
fully in the future.

Signed-off-by: Atish Patra <[e-mail]>
---
 arch/riscv/include/asm/kvm_host.h     |   4 +
 arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
 arch/riscv/kvm/Makefile               |   1 +
 arch/riscv/kvm/vcpu.c                 |   7 ++
 arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
 5 files changed, 226 insertions(+)
 create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
 create mode 100644 arch/riscv/kvm/vcpu_pmu.c

Comments

Anup Patel Feb. 2, 2023, 4:02 a.m. UTC | #1

On Thu, Feb 2, 2023 at 4:42 AM Atish Patra <[e-mail]> wrote:
>
> This patch only adds barebone structure of perf implementation. Most of
> the function returns zero at this point and will be implemented
> fully in the future.
>
> Signed-off-by: Atish Patra <[e-mail]>

Looks good to me.

Reviewed-by: Anup Patel <[e-mail]>

Regards,
Anup

> ---
>  arch/riscv/include/asm/kvm_host.h     |   4 +
>  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
>  arch/riscv/kvm/Makefile               |   1 +
>  arch/riscv/kvm/vcpu.c                 |   7 ++
>  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
>  5 files changed, 226 insertions(+)
>  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
>  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
>
> diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> index 93f43a3..b90be9a 100644
> --- a/arch/riscv/include/asm/kvm_host.h
> +++ b/arch/riscv/include/asm/kvm_host.h
> @@ -18,6 +18,7 @@
>  #include <asm/kvm_vcpu_insn.h>
>  #include <asm/kvm_vcpu_sbi.h>
>  #include <asm/kvm_vcpu_timer.h>
> +#include <asm/kvm_vcpu_pmu.h>
>
>  #define KVM_MAX_VCPUS                  1024
>
> @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
>
>         /* Don't run the VCPU (blocked) */
>         bool pause;
> +
> +       /* Performance monitoring context */
> +       struct kvm_pmu pmu_context;
>  };
>
>  static inline void kvm_arch_hardware_unsetup(void) {}
> diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> new file mode 100644
> index 0000000..e2b4038
> --- /dev/null
> +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> @@ -0,0 +1,78 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (c) 2023 Rivos Inc
> + *
> + * Authors:
> + *     Atish Patra <[e-mail]>
> + */
> +
> +#ifndef __KVM_VCPU_RISCV_PMU_H
> +#define __KVM_VCPU_RISCV_PMU_H
> +
> +#include <linux/perf/riscv_pmu.h>
> +#include <asm/kvm_vcpu_sbi.h>
> +#include <asm/sbi.h>
> +
> +#ifdef CONFIG_RISCV_PMU_SBI
> +#define RISCV_KVM_MAX_FW_CTRS  32
> +
> +#if RISCV_KVM_MAX_FW_CTRS > 32
> +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
> +#endif
> +
> +#define RISCV_MAX_COUNTERS      64
> +
> +/* Per virtual pmu counter data */
> +struct kvm_pmc {
> +       u8 idx;
> +       struct perf_event *perf_event;
> +       uint64_t counter_val;
> +       union sbi_pmu_ctr_info cinfo;
> +       /* Event monitoring status */
> +       bool started;
> +};
> +
> +/* PMU data structure per vcpu */
> +struct kvm_pmu {
> +       struct kvm_pmc pmc[RISCV_MAX_COUNTERS];
> +       /* Number of the virtual firmware counters available */
> +       int num_fw_ctrs;
> +       /* Number of the virtual hardware counters available */
> +       int num_hw_ctrs;
> +       /* A flag to indicate that pmu initialization is done */
> +       bool init_done;
> +       /* Bit map of all the virtual counter used */
> +       DECLARE_BITMAP(pmc_in_use, RISCV_MAX_COUNTERS);
> +};
> +
> +#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu_context)
> +#define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu_context))
> +
> +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> +                               struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +                                unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> +                                struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +                               unsigned long ctr_mask, unsigned long flag,
> +                               struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +                                    unsigned long ctr_mask, unsigned long flag,
> +                                    unsigned long eidx, uint64_t evtdata,
> +                                    struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> +                               struct kvm_vcpu_sbi_return *retdata);
> +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu);
> +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu);
> +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu);
> +
> +#else
> +struct kvm_pmu {
> +};
> +
> +static inline void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu) {}
> +static inline void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu) {}
> +static inline void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu) {}
> +#endif /* CONFIG_RISCV_PMU_SBI */
> +#endif /* !__KVM_VCPU_RISCV_PMU_H */
> diff --git a/arch/riscv/kvm/Makefile b/arch/riscv/kvm/Makefile
> index 019df920..5de1053 100644
> --- a/arch/riscv/kvm/Makefile
> +++ b/arch/riscv/kvm/Makefile
> @@ -25,3 +25,4 @@ kvm-y += vcpu_sbi_base.o
>  kvm-y += vcpu_sbi_replace.o
>  kvm-y += vcpu_sbi_hsm.o
>  kvm-y += vcpu_timer.o
> +kvm-$(CONFIG_RISCV_PMU_SBI) += vcpu_pmu.o
> diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> index 7c08567..7d010b0 100644
> --- a/arch/riscv/kvm/vcpu.c
> +++ b/arch/riscv/kvm/vcpu.c
> @@ -138,6 +138,8 @@ static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu)
>         WRITE_ONCE(vcpu->arch.irqs_pending, 0);
>         WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0);
>
> +       kvm_riscv_vcpu_pmu_reset(vcpu);
> +
>         vcpu->arch.hfence_head = 0;
>         vcpu->arch.hfence_tail = 0;
>         memset(vcpu->arch.hfence_queue, 0, sizeof(vcpu->arch.hfence_queue));
> @@ -194,6 +196,9 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
>         /* Setup VCPU timer */
>         kvm_riscv_vcpu_timer_init(vcpu);
>
> +       /* setup performance monitoring */
> +       kvm_riscv_vcpu_pmu_init(vcpu);
> +
>         /* Reset VCPU */
>         kvm_riscv_reset_vcpu(vcpu);
>
> @@ -216,6 +221,8 @@ void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
>         /* Cleanup VCPU timer */
>         kvm_riscv_vcpu_timer_deinit(vcpu);
>
> +       kvm_riscv_vcpu_pmu_deinit(vcpu);
> +
>         /* Free unused pages pre-allocated for G-stage page table mappings */
>         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
>  }
> diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
> new file mode 100644
> index 0000000..2dad37f
> --- /dev/null
> +++ b/arch/riscv/kvm/vcpu_pmu.c
> @@ -0,0 +1,136 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Copyright (c) 2023 Rivos Inc
> + *
> + * Authors:
> + *     Atish Patra <[e-mail]>
> + */
> +
> +#include <linux/errno.h>
> +#include <linux/err.h>
> +#include <linux/kvm_host.h>
> +#include <linux/perf/riscv_pmu.h>
> +#include <asm/csr.h>
> +#include <asm/kvm_vcpu_sbi.h>
> +#include <asm/kvm_vcpu_pmu.h>
> +#include <linux/kvm_host.h>
> +
> +#define kvm_pmu_num_counters(pmu) ((pmu)->num_hw_ctrs + (pmu)->num_fw_ctrs)
> +
> +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata)
> +{
> +       struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> +
> +       retdata->out_val = kvm_pmu_num_counters(kvpmu);
> +
> +       return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> +                               struct kvm_vcpu_sbi_return *retdata)
> +{
> +       struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> +
> +       if (cidx > RISCV_MAX_COUNTERS || cidx == 1) {
> +               retdata->err_val = SBI_ERR_INVALID_PARAM;
> +               return 0;
> +       }
> +
> +       retdata->out_val = kvpmu->pmc[cidx].cinfo.value;
> +
> +       return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +                                unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> +                                struct kvm_vcpu_sbi_return *retdata)
> +{
> +       /* TODO */
> +       return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +                               unsigned long ctr_mask, unsigned long flag,
> +                               struct kvm_vcpu_sbi_return *retdata)
> +{
> +       /* TODO */
> +       return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +                                    unsigned long ctr_mask, unsigned long flag,
> +                                    unsigned long eidx, uint64_t evtdata,
> +                                    struct kvm_vcpu_sbi_return *retdata)
> +{
> +       /* TODO */
> +       return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> +                               struct kvm_vcpu_sbi_return *retdata)
> +{
> +       /* TODO */
> +       return 0;
> +}
> +
> +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
> +{
> +       int i = 0, ret, num_hw_ctrs = 0, hpm_width = 0;
> +       struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> +       struct kvm_pmc *pmc;
> +
> +       ret = riscv_pmu_get_hpm_info(&hpm_width, &num_hw_ctrs);
> +       if (ret < 0 || !hpm_width || !num_hw_ctrs)
> +               return;
> +
> +       /*
> +        * It is guranteed that RISCV_KVM_MAX_FW_CTRS can't exceed 32 as
> +        * that may exceed total number of counters more than RISCV_MAX_COUNTERS
> +        */
> +       kvpmu->num_hw_ctrs = num_hw_ctrs;
> +       kvpmu->num_fw_ctrs = RISCV_KVM_MAX_FW_CTRS;
> +
> +       /*
> +        * There is no correlation between the logical hardware counter and virtual counters.
> +        * However, we need to encode a hpmcounter CSR in the counter info field so that
> +        * KVM can trap n emulate the read. This works well in the migration use case as
> +        * KVM doesn't care if the actual hpmcounter is available in the hardware or not.
> +        */
> +       for (i = 0; i < kvm_pmu_num_counters(kvpmu); i++) {
> +               /* TIME CSR shouldn't be read from perf interface */
> +               if (i == 1)
> +                       continue;
> +               pmc = &kvpmu->pmc[i];
> +               pmc->idx = i;
> +               if (i < kvpmu->num_hw_ctrs) {
> +                       pmc->cinfo.type = SBI_PMU_CTR_TYPE_HW;
> +                       if (i < 3)
> +                               /* CY, IR counters */
> +                               pmc->cinfo.width = 63;
> +                       else
> +                               pmc->cinfo.width = hpm_width;
> +                       /*
> +                        * The CSR number doesn't have any relation with the logical
> +                        * hardware counters. The CSR numbers are encoded sequentially
> +                        * to avoid maintaining a map between the virtual counter
> +                        * and CSR number.
> +                        */
> +                       pmc->cinfo.csr = CSR_CYCLE + i;
> +               } else {
> +                       pmc->cinfo.type = SBI_PMU_CTR_TYPE_FW;
> +                       pmc->cinfo.width = BITS_PER_LONG - 1;
> +               }
> +       }
> +
> +       kvpmu->init_done = true;
> +}
> +
> +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu)
> +{
> +       /* TODO */
> +}
> +
> +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu)
> +{
> +       kvm_riscv_vcpu_pmu_deinit(vcpu);
> +}
> --
> 2.25.1
>

Conor Dooley Feb. 2, 2023, 11:33 a.m. UTC | #2

On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> This patch only adds barebone structure of perf implementation. Most of
> the function returns zero at this point and will be implemented
> fully in the future.
> 
> Signed-off-by: Atish Patra <[e-mail]>
> +/* Per virtual pmu counter data */
> +struct kvm_pmc {
> +	u8 idx;
> +	struct perf_event *perf_event;
> +	uint64_t counter_val;

CI also complained that here, and elsewhere, you used uint64_t rather
than u64. Am I missing a reason for not using the regular types?

Thanks,
Conor.

> +	union sbi_pmu_ctr_info cinfo;
> +	/* Event monitoring status */
> +	bool started;

Andrew Jones Feb. 2, 2023, 5:03 p.m. UTC | #3

On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> This patch only adds barebone structure of perf implementation. Most of
> the function returns zero at this point and will be implemented
> fully in the future.
> 
> Signed-off-by: Atish Patra <[e-mail]>
> ---
>  arch/riscv/include/asm/kvm_host.h     |   4 +
>  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
>  arch/riscv/kvm/Makefile               |   1 +
>  arch/riscv/kvm/vcpu.c                 |   7 ++
>  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
>  5 files changed, 226 insertions(+)
>  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
>  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
> 
> diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> index 93f43a3..b90be9a 100644
> --- a/arch/riscv/include/asm/kvm_host.h
> +++ b/arch/riscv/include/asm/kvm_host.h
> @@ -18,6 +18,7 @@
>  #include <asm/kvm_vcpu_insn.h>
>  #include <asm/kvm_vcpu_sbi.h>
>  #include <asm/kvm_vcpu_timer.h>
> +#include <asm/kvm_vcpu_pmu.h>
>  
>  #define KVM_MAX_VCPUS			1024
>  
> @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
>  
>  	/* Don't run the VCPU (blocked) */
>  	bool pause;
> +
> +	/* Performance monitoring context */
> +	struct kvm_pmu pmu_context;
>  };
>  
>  static inline void kvm_arch_hardware_unsetup(void) {}
> diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> new file mode 100644
> index 0000000..e2b4038
> --- /dev/null
> +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> @@ -0,0 +1,78 @@
> +/* SPDX-License-Identifier: GPL-2.0-only */
> +/*
> + * Copyright (c) 2023 Rivos Inc
> + *
> + * Authors:
> + *     Atish Patra <[e-mail]>
> + */
> +
> +#ifndef __KVM_VCPU_RISCV_PMU_H
> +#define __KVM_VCPU_RISCV_PMU_H
> +
> +#include <linux/perf/riscv_pmu.h>
> +#include <asm/kvm_vcpu_sbi.h>
> +#include <asm/sbi.h>
> +
> +#ifdef CONFIG_RISCV_PMU_SBI
> +#define RISCV_KVM_MAX_FW_CTRS	32
> +
> +#if RISCV_KVM_MAX_FW_CTRS > 32
> +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"

"The number of firmware counters cannot exceed 32 without increasing RISCV_MAX_COUNTERS"

> +#endif
> +
> +#define RISCV_MAX_COUNTERS      64

But instead of that message, what I think we need is something like

 #define RISCV_KVM_MAX_HW_CTRS	32
 #define RISCV_KVM_MAX_FW_CTRS	32
 #define RISCV_MAX_COUNTERS	(RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)

 static_assert(RISCV_MAX_COUNTERS <= 64)

And then in pmu_sbi_device_probe() should ensure

  num_counters <= RISCV_MAX_COUNTERS

and pmu_sbi_get_ctrinfo() should ensure

  num_hw_ctr <= RISCV_KVM_MAX_HW_CTRS
  num_fw_ctr <= RISCV_KVM_MAX_FW_CTRS

which has to be done at runtime.

> +
> +/* Per virtual pmu counter data */
> +struct kvm_pmc {
> +	u8 idx;
> +	struct perf_event *perf_event;
> +	uint64_t counter_val;
> +	union sbi_pmu_ctr_info cinfo;
> +	/* Event monitoring status */
> +	bool started;
> +};
> +
> +/* PMU data structure per vcpu */
> +struct kvm_pmu {
> +	struct kvm_pmc pmc[RISCV_MAX_COUNTERS];
> +	/* Number of the virtual firmware counters available */
> +	int num_fw_ctrs;
> +	/* Number of the virtual hardware counters available */
> +	int num_hw_ctrs;
> +	/* A flag to indicate that pmu initialization is done */
> +	bool init_done;
> +	/* Bit map of all the virtual counter used */
> +	DECLARE_BITMAP(pmc_in_use, RISCV_MAX_COUNTERS);
> +};
> +
> +#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu_context)
> +#define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu_context))
> +
> +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> +				struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +				 unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> +				 struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +				unsigned long ctr_mask, unsigned long flag,
> +				struct kvm_vcpu_sbi_return *retdata);
> +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +				     unsigned long ctr_mask, unsigned long flag,
> +				     unsigned long eidx, uint64_t evtdata,
> +				     struct kvm_vcpu_sbi_return *retdata);

s/flag/flags/ for all the above prototypes and all the implementations
below.

> +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> +				struct kvm_vcpu_sbi_return *retdata);
> +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu);
> +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu);
> +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu);
> +
> +#else
> +struct kvm_pmu {
> +};
> +
> +static inline void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu) {}
> +static inline void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu) {}
> +static inline void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu) {}
> +#endif /* CONFIG_RISCV_PMU_SBI */
> +#endif /* !__KVM_VCPU_RISCV_PMU_H */
> diff --git a/arch/riscv/kvm/Makefile b/arch/riscv/kvm/Makefile
> index 019df920..5de1053 100644
> --- a/arch/riscv/kvm/Makefile
> +++ b/arch/riscv/kvm/Makefile
> @@ -25,3 +25,4 @@ kvm-y += vcpu_sbi_base.o
>  kvm-y += vcpu_sbi_replace.o
>  kvm-y += vcpu_sbi_hsm.o
>  kvm-y += vcpu_timer.o
> +kvm-$(CONFIG_RISCV_PMU_SBI) += vcpu_pmu.o
> diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> index 7c08567..7d010b0 100644
> --- a/arch/riscv/kvm/vcpu.c
> +++ b/arch/riscv/kvm/vcpu.c
> @@ -138,6 +138,8 @@ static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu)
>  	WRITE_ONCE(vcpu->arch.irqs_pending, 0);
>  	WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0);
>  
> +	kvm_riscv_vcpu_pmu_reset(vcpu);
> +
>  	vcpu->arch.hfence_head = 0;
>  	vcpu->arch.hfence_tail = 0;
>  	memset(vcpu->arch.hfence_queue, 0, sizeof(vcpu->arch.hfence_queue));
> @@ -194,6 +196,9 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
>  	/* Setup VCPU timer */
>  	kvm_riscv_vcpu_timer_init(vcpu);
>  
> +	/* setup performance monitoring */
> +	kvm_riscv_vcpu_pmu_init(vcpu);
> +
>  	/* Reset VCPU */
>  	kvm_riscv_reset_vcpu(vcpu);
>  
> @@ -216,6 +221,8 @@ void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
>  	/* Cleanup VCPU timer */
>  	kvm_riscv_vcpu_timer_deinit(vcpu);
>  
> +	kvm_riscv_vcpu_pmu_deinit(vcpu);
> +
>  	/* Free unused pages pre-allocated for G-stage page table mappings */
>  	kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
>  }
> diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
> new file mode 100644
> index 0000000..2dad37f
> --- /dev/null
> +++ b/arch/riscv/kvm/vcpu_pmu.c
> @@ -0,0 +1,136 @@
> +// SPDX-License-Identifier: GPL-2.0
> +/*
> + * Copyright (c) 2023 Rivos Inc
> + *
> + * Authors:
> + *     Atish Patra <[e-mail]>
> + */
> +
> +#include <linux/errno.h>
> +#include <linux/err.h>
> +#include <linux/kvm_host.h>
> +#include <linux/perf/riscv_pmu.h>
> +#include <asm/csr.h>
> +#include <asm/kvm_vcpu_sbi.h>
> +#include <asm/kvm_vcpu_pmu.h>
> +#include <linux/kvm_host.h>
> +
> +#define kvm_pmu_num_counters(pmu) ((pmu)->num_hw_ctrs + (pmu)->num_fw_ctrs)
> +
> +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata)
> +{
> +	struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> +
> +	retdata->out_val = kvm_pmu_num_counters(kvpmu);
> +
> +	return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> +				struct kvm_vcpu_sbi_return *retdata)
> +{
> +	struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> +
> +	if (cidx > RISCV_MAX_COUNTERS || cidx == 1) {
> +		retdata->err_val = SBI_ERR_INVALID_PARAM;
> +		return 0;
> +	}
> +
> +	retdata->out_val = kvpmu->pmc[cidx].cinfo.value;
> +
> +	return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +				 unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> +				 struct kvm_vcpu_sbi_return *retdata)
> +{
> +	/* TODO */
> +	return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +				unsigned long ctr_mask, unsigned long flag,
> +				struct kvm_vcpu_sbi_return *retdata)
> +{
> +	/* TODO */
> +	return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> +				     unsigned long ctr_mask, unsigned long flag,
> +				     unsigned long eidx, uint64_t evtdata,
> +				     struct kvm_vcpu_sbi_return *retdata)
> +{
> +	/* TODO */
> +	return 0;
> +}
> +
> +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> +				struct kvm_vcpu_sbi_return *retdata)
> +{
> +	/* TODO */
> +	return 0;
> +}
> +
> +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
> +{
> +	int i = 0, ret, num_hw_ctrs = 0, hpm_width = 0;
> +	struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> +	struct kvm_pmc *pmc;
> +
> +	ret = riscv_pmu_get_hpm_info(&hpm_width, &num_hw_ctrs);
> +	if (ret < 0 || !hpm_width || !num_hw_ctrs)
> +		return;
> +
> +	/*
> +	 * It is guranteed that RISCV_KVM_MAX_FW_CTRS can't exceed 32 as
> +	 * that may exceed total number of counters more than RISCV_MAX_COUNTERS
> +	 */
> +	kvpmu->num_hw_ctrs = num_hw_ctrs;
> +	kvpmu->num_fw_ctrs = RISCV_KVM_MAX_FW_CTRS;

If we sanity check that num_hw_ctrs <= 32 and num_fw_ctrs <= 32 at sbi_pmu
probe time, then we can also return num_fw_ctrs (or num_ctrs) along with
num_hw_ctrs from riscv_pmu_get_hpm_info(). Then, we can put the exact
number here into kvmpmu->num_fw_ctrs, rather than using its max.

> +
> +	/*
> +	 * There is no correlation between the logical hardware counter and virtual counters.
> +	 * However, we need to encode a hpmcounter CSR in the counter info field so that
> +	 * KVM can trap n emulate the read. This works well in the migration use case as
> +	 * KVM doesn't care if the actual hpmcounter is available in the hardware or not.
> +	 */
> +	for (i = 0; i < kvm_pmu_num_counters(kvpmu); i++) {
> +		/* TIME CSR shouldn't be read from perf interface */
> +		if (i == 1)
> +			continue;
> +		pmc = &kvpmu->pmc[i];
> +		pmc->idx = i;
> +		if (i < kvpmu->num_hw_ctrs) {
> +			pmc->cinfo.type = SBI_PMU_CTR_TYPE_HW;
> +			if (i < 3)
> +				/* CY, IR counters */
> +				pmc->cinfo.width = 63;
> +			else
> +				pmc->cinfo.width = hpm_width;
> +			/*
> +			 * The CSR number doesn't have any relation with the logical
> +			 * hardware counters. The CSR numbers are encoded sequentially
> +			 * to avoid maintaining a map between the virtual counter
> +			 * and CSR number.
> +			 */
> +			pmc->cinfo.csr = CSR_CYCLE + i;
> +		} else {
> +			pmc->cinfo.type = SBI_PMU_CTR_TYPE_FW;
> +			pmc->cinfo.width = BITS_PER_LONG - 1;
> +		}
> +	}
> +
> +	kvpmu->init_done = true;
> +}
> +
> +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu)
> +{
> +	/* TODO */
> +}
> +
> +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu)
> +{
> +	kvm_riscv_vcpu_pmu_deinit(vcpu);
> +}
> -- 
> 2.25.1
>

Thanks,
drew

Atish Patra Feb. 3, 2023, 8:04 a.m. UTC | #4

On Thu, Feb 2, 2023 at 3:34 AM Conor Dooley <[e-mail]> wrote:
>
> On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> > This patch only adds barebone structure of perf implementation. Most of
> > the function returns zero at this point and will be implemented
> > fully in the future.
> >
> > Signed-off-by: Atish Patra <[e-mail]>
> > +/* Per virtual pmu counter data */
> > +struct kvm_pmc {
> > +     u8 idx;
> > +     struct perf_event *perf_event;
> > +     uint64_t counter_val;
>
> CI also complained that here, and elsewhere, you used uint64_t rather
> than u64. Am I missing a reason for not using the regular types?
>

Nope. It was a simple oversight. I will fix it.
Do you have a link to the CI report so that I can address them all in v5 ?

> Thanks,
> Conor.
>
> > +     union sbi_pmu_ctr_info cinfo;
> > +     /* Event monitoring status */
> > +     bool started;

Conor Dooley Feb. 3, 2023, 8:08 a.m. UTC | #5

On 3 February 2023 08:04:00 GMT, Atish Patra <[e-mail]> wrote:
>On Thu, Feb 2, 2023 at 3:34 AM Conor Dooley <[e-mail]> wrote:
>>
>> On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
>> > This patch only adds barebone structure of perf implementation. Most of
>> > the function returns zero at this point and will be implemented
>> > fully in the future.
>> >
>> > Signed-off-by: Atish Patra <[e-mail]>
>> > +/* Per virtual pmu counter data */
>> > +struct kvm_pmc {
>> > +     u8 idx;
>> > +     struct perf_event *perf_event;
>> > +     uint64_t counter_val;
>>
>> CI also complained that here, and elsewhere, you used uint64_t rather
>> than u64. Am I missing a reason for not using the regular types?
>>
>
>Nope. It was a simple oversight. I will fix it.
>Do you have a link to the CI report so that I can address them all in v5 ?

Try:
:%s/uint64_t/u64
It was just this patch, and checkpatch --strict should show it.

>
>> Thanks,
>> Conor.
>>
>> > +     union sbi_pmu_ctr_info cinfo;
>> > +     /* Event monitoring status */
>> > +     bool started;
>
>
>

Atish Patra Feb. 3, 2023, 8:47 a.m. UTC | #6

On Thu, Feb 2, 2023 at 9:03 AM Andrew Jones <[e-mail]> wrote:
>
> On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> > This patch only adds barebone structure of perf implementation. Most of
> > the function returns zero at this point and will be implemented
> > fully in the future.
> >
> > Signed-off-by: Atish Patra <[e-mail]>
> > ---
> >  arch/riscv/include/asm/kvm_host.h     |   4 +
> >  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
> >  arch/riscv/kvm/Makefile               |   1 +
> >  arch/riscv/kvm/vcpu.c                 |   7 ++
> >  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
> >  5 files changed, 226 insertions(+)
> >  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
> >  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
> >
> > diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> > index 93f43a3..b90be9a 100644
> > --- a/arch/riscv/include/asm/kvm_host.h
> > +++ b/arch/riscv/include/asm/kvm_host.h
> > @@ -18,6 +18,7 @@
> >  #include <asm/kvm_vcpu_insn.h>
> >  #include <asm/kvm_vcpu_sbi.h>
> >  #include <asm/kvm_vcpu_timer.h>
> > +#include <asm/kvm_vcpu_pmu.h>
> >
> >  #define KVM_MAX_VCPUS                        1024
> >
> > @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
> >
> >       /* Don't run the VCPU (blocked) */
> >       bool pause;
> > +
> > +     /* Performance monitoring context */
> > +     struct kvm_pmu pmu_context;
> >  };
> >
> >  static inline void kvm_arch_hardware_unsetup(void) {}
> > diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > new file mode 100644
> > index 0000000..e2b4038
> > --- /dev/null
> > +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > @@ -0,0 +1,78 @@
> > +/* SPDX-License-Identifier: GPL-2.0-only */
> > +/*
> > + * Copyright (c) 2023 Rivos Inc
> > + *
> > + * Authors:
> > + *     Atish Patra <[e-mail]>
> > + */
> > +
> > +#ifndef __KVM_VCPU_RISCV_PMU_H
> > +#define __KVM_VCPU_RISCV_PMU_H
> > +
> > +#include <linux/perf/riscv_pmu.h>
> > +#include <asm/kvm_vcpu_sbi.h>
> > +#include <asm/sbi.h>
> > +
> > +#ifdef CONFIG_RISCV_PMU_SBI
> > +#define RISCV_KVM_MAX_FW_CTRS        32
> > +
> > +#if RISCV_KVM_MAX_FW_CTRS > 32
> > +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
>
> "The number of firmware counters cannot exceed 32 without increasing RISCV_MAX_COUNTERS"
>
> > +#endif
> > +
> > +#define RISCV_MAX_COUNTERS      64
>
> But instead of that message, what I think we need is something like
>
>  #define RISCV_KVM_MAX_HW_CTRS  32
>  #define RISCV_KVM_MAX_FW_CTRS  32
>  #define RISCV_MAX_COUNTERS     (RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
>
>  static_assert(RISCV_MAX_COUNTERS <= 64)
>
> And then in pmu_sbi_device_probe() should ensure
>
>   num_counters <= RISCV_MAX_COUNTERS
>
> and pmu_sbi_get_ctrinfo() should ensure
>
>   num_hw_ctr <= RISCV_KVM_MAX_HW_CTRS
>   num_fw_ctr <= RISCV_KVM_MAX_FW_CTRS
>
> which has to be done at runtime.
>

Sure. I will add the additional sanity checks.

> > +
> > +/* Per virtual pmu counter data */
> > +struct kvm_pmc {
> > +     u8 idx;
> > +     struct perf_event *perf_event;
> > +     uint64_t counter_val;
> > +     union sbi_pmu_ctr_info cinfo;
> > +     /* Event monitoring status */
> > +     bool started;
> > +};
> > +
> > +/* PMU data structure per vcpu */
> > +struct kvm_pmu {
> > +     struct kvm_pmc pmc[RISCV_MAX_COUNTERS];
> > +     /* Number of the virtual firmware counters available */
> > +     int num_fw_ctrs;
> > +     /* Number of the virtual hardware counters available */
> > +     int num_hw_ctrs;
> > +     /* A flag to indicate that pmu initialization is done */
> > +     bool init_done;
> > +     /* Bit map of all the virtual counter used */
> > +     DECLARE_BITMAP(pmc_in_use, RISCV_MAX_COUNTERS);
> > +};
> > +
> > +#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu_context)
> > +#define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu_context))
> > +
> > +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata);
> > +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> > +                             struct kvm_vcpu_sbi_return *retdata);
> > +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > +                              unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> > +                              struct kvm_vcpu_sbi_return *retdata);
> > +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > +                             unsigned long ctr_mask, unsigned long flag,
> > +                             struct kvm_vcpu_sbi_return *retdata);
> > +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > +                                  unsigned long ctr_mask, unsigned long flag,
> > +                                  unsigned long eidx, uint64_t evtdata,
> > +                                  struct kvm_vcpu_sbi_return *retdata);
>
> s/flag/flags/ for all the above prototypes and all the implementations
> below.
>

Fixed.

> > +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> > +                             struct kvm_vcpu_sbi_return *retdata);
> > +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu);
> > +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu);
> > +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu);
> > +
> > +#else
> > +struct kvm_pmu {
> > +};
> > +
> > +static inline void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu) {}
> > +static inline void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu) {}
> > +static inline void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu) {}
> > +#endif /* CONFIG_RISCV_PMU_SBI */
> > +#endif /* !__KVM_VCPU_RISCV_PMU_H */
> > diff --git a/arch/riscv/kvm/Makefile b/arch/riscv/kvm/Makefile
> > index 019df920..5de1053 100644
> > --- a/arch/riscv/kvm/Makefile
> > +++ b/arch/riscv/kvm/Makefile
> > @@ -25,3 +25,4 @@ kvm-y += vcpu_sbi_base.o
> >  kvm-y += vcpu_sbi_replace.o
> >  kvm-y += vcpu_sbi_hsm.o
> >  kvm-y += vcpu_timer.o
> > +kvm-$(CONFIG_RISCV_PMU_SBI) += vcpu_pmu.o
> > diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> > index 7c08567..7d010b0 100644
> > --- a/arch/riscv/kvm/vcpu.c
> > +++ b/arch/riscv/kvm/vcpu.c
> > @@ -138,6 +138,8 @@ static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu)
> >       WRITE_ONCE(vcpu->arch.irqs_pending, 0);
> >       WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0);
> >
> > +     kvm_riscv_vcpu_pmu_reset(vcpu);
> > +
> >       vcpu->arch.hfence_head = 0;
> >       vcpu->arch.hfence_tail = 0;
> >       memset(vcpu->arch.hfence_queue, 0, sizeof(vcpu->arch.hfence_queue));
> > @@ -194,6 +196,9 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
> >       /* Setup VCPU timer */
> >       kvm_riscv_vcpu_timer_init(vcpu);
> >
> > +     /* setup performance monitoring */
> > +     kvm_riscv_vcpu_pmu_init(vcpu);
> > +
> >       /* Reset VCPU */
> >       kvm_riscv_reset_vcpu(vcpu);
> >
> > @@ -216,6 +221,8 @@ void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
> >       /* Cleanup VCPU timer */
> >       kvm_riscv_vcpu_timer_deinit(vcpu);
> >
> > +     kvm_riscv_vcpu_pmu_deinit(vcpu);
> > +
> >       /* Free unused pages pre-allocated for G-stage page table mappings */
> >       kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
> >  }
> > diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
> > new file mode 100644
> > index 0000000..2dad37f
> > --- /dev/null
> > +++ b/arch/riscv/kvm/vcpu_pmu.c
> > @@ -0,0 +1,136 @@
> > +// SPDX-License-Identifier: GPL-2.0
> > +/*
> > + * Copyright (c) 2023 Rivos Inc
> > + *
> > + * Authors:
> > + *     Atish Patra <[e-mail]>
> > + */
> > +
> > +#include <linux/errno.h>
> > +#include <linux/err.h>
> > +#include <linux/kvm_host.h>
> > +#include <linux/perf/riscv_pmu.h>
> > +#include <asm/csr.h>
> > +#include <asm/kvm_vcpu_sbi.h>
> > +#include <asm/kvm_vcpu_pmu.h>
> > +#include <linux/kvm_host.h>
> > +
> > +#define kvm_pmu_num_counters(pmu) ((pmu)->num_hw_ctrs + (pmu)->num_fw_ctrs)
> > +
> > +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata)
> > +{
> > +     struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > +
> > +     retdata->out_val = kvm_pmu_num_counters(kvpmu);
> > +
> > +     return 0;
> > +}
> > +
> > +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> > +                             struct kvm_vcpu_sbi_return *retdata)
> > +{
> > +     struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > +
> > +     if (cidx > RISCV_MAX_COUNTERS || cidx == 1) {
> > +             retdata->err_val = SBI_ERR_INVALID_PARAM;
> > +             return 0;
> > +     }
> > +
> > +     retdata->out_val = kvpmu->pmc[cidx].cinfo.value;
> > +
> > +     return 0;
> > +}
> > +
> > +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > +                              unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> > +                              struct kvm_vcpu_sbi_return *retdata)
> > +{
> > +     /* TODO */
> > +     return 0;
> > +}
> > +
> > +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > +                             unsigned long ctr_mask, unsigned long flag,
> > +                             struct kvm_vcpu_sbi_return *retdata)
> > +{
> > +     /* TODO */
> > +     return 0;
> > +}
> > +
> > +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > +                                  unsigned long ctr_mask, unsigned long flag,
> > +                                  unsigned long eidx, uint64_t evtdata,
> > +                                  struct kvm_vcpu_sbi_return *retdata)
> > +{
> > +     /* TODO */
> > +     return 0;
> > +}
> > +
> > +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> > +                             struct kvm_vcpu_sbi_return *retdata)
> > +{
> > +     /* TODO */
> > +     return 0;
> > +}
> > +
> > +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
> > +{
> > +     int i = 0, ret, num_hw_ctrs = 0, hpm_width = 0;
> > +     struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > +     struct kvm_pmc *pmc;
> > +
> > +     ret = riscv_pmu_get_hpm_info(&hpm_width, &num_hw_ctrs);
> > +     if (ret < 0 || !hpm_width || !num_hw_ctrs)
> > +             return;
> > +
> > +     /*
> > +      * It is guranteed that RISCV_KVM_MAX_FW_CTRS can't exceed 32 as
> > +      * that may exceed total number of counters more than RISCV_MAX_COUNTERS
> > +      */
> > +     kvpmu->num_hw_ctrs = num_hw_ctrs;
> > +     kvpmu->num_fw_ctrs = RISCV_KVM_MAX_FW_CTRS;
>
> If we sanity check that num_hw_ctrs <= 32 and num_fw_ctrs <= 32 at sbi_pmu
> probe time, then we can also return num_fw_ctrs (or num_ctrs) along with
> num_hw_ctrs from riscv_pmu_get_hpm_info(). Then, we can put the exact
> number here into kvmpmu->num_fw_ctrs, rather than using its max.
>

The firmware counter information retrieved from PMU driver will be the
number of firmware
counter host supports (i.e. M-mode firmware supports). The number of
counters supported for a
guest is entirely up to the hypervisor. There shouldn't be any
relation with the host's firmware counter.

Looking at it again, we should probably set kvpmu->num_fw_ctrs to
SBI_PMU_FW_MAX instead of RISCV_KVM_MAX_FW_CTRS.
We already have a sanity check for SBI_PMU_FW_MAX in the code.
> > +
> > +     /*
> > +      * There is no correlation between the logical hardware counter and virtual counters.
> > +      * However, we need to encode a hpmcounter CSR in the counter info field so that
> > +      * KVM can trap n emulate the read. This works well in the migration use case as
> > +      * KVM doesn't care if the actual hpmcounter is available in the hardware or not.
> > +      */
> > +     for (i = 0; i < kvm_pmu_num_counters(kvpmu); i++) {
> > +             /* TIME CSR shouldn't be read from perf interface */
> > +             if (i == 1)
> > +                     continue;
> > +             pmc = &kvpmu->pmc[i];
> > +             pmc->idx = i;
> > +             if (i < kvpmu->num_hw_ctrs) {
> > +                     pmc->cinfo.type = SBI_PMU_CTR_TYPE_HW;
> > +                     if (i < 3)
> > +                             /* CY, IR counters */
> > +                             pmc->cinfo.width = 63;
> > +                     else
> > +                             pmc->cinfo.width = hpm_width;
> > +                     /*
> > +                      * The CSR number doesn't have any relation with the logical
> > +                      * hardware counters. The CSR numbers are encoded sequentially
> > +                      * to avoid maintaining a map between the virtual counter
> > +                      * and CSR number.
> > +                      */
> > +                     pmc->cinfo.csr = CSR_CYCLE + i;
> > +             } else {
> > +                     pmc->cinfo.type = SBI_PMU_CTR_TYPE_FW;
> > +                     pmc->cinfo.width = BITS_PER_LONG - 1;
> > +             }
> > +     }
> > +
> > +     kvpmu->init_done = true;
> > +}
> > +
> > +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu)
> > +{
> > +     /* TODO */
> > +}
> > +
> > +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu)
> > +{
> > +     kvm_riscv_vcpu_pmu_deinit(vcpu);
> > +}
> > --
> > 2.25.1
> >
>
> Thanks,
> drew

Atish Patra Feb. 5, 2023, 7:37 a.m. UTC | #7

On Fri, Feb 3, 2023 at 12:47 AM Atish Patra <[e-mail]> wrote:
>
> On Thu, Feb 2, 2023 at 9:03 AM Andrew Jones <[e-mail]> wrote:
> >
> > On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> > > This patch only adds barebone structure of perf implementation. Most of
> > > the function returns zero at this point and will be implemented
> > > fully in the future.
> > >
> > > Signed-off-by: Atish Patra <[e-mail]>
> > > ---
> > >  arch/riscv/include/asm/kvm_host.h     |   4 +
> > >  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
> > >  arch/riscv/kvm/Makefile               |   1 +
> > >  arch/riscv/kvm/vcpu.c                 |   7 ++
> > >  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
> > >  5 files changed, 226 insertions(+)
> > >  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
> > >  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
> > >
> > > diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> > > index 93f43a3..b90be9a 100644
> > > --- a/arch/riscv/include/asm/kvm_host.h
> > > +++ b/arch/riscv/include/asm/kvm_host.h
> > > @@ -18,6 +18,7 @@
> > >  #include <asm/kvm_vcpu_insn.h>
> > >  #include <asm/kvm_vcpu_sbi.h>
> > >  #include <asm/kvm_vcpu_timer.h>
> > > +#include <asm/kvm_vcpu_pmu.h>
> > >
> > >  #define KVM_MAX_VCPUS                        1024
> > >
> > > @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
> > >
> > >       /* Don't run the VCPU (blocked) */
> > >       bool pause;
> > > +
> > > +     /* Performance monitoring context */
> > > +     struct kvm_pmu pmu_context;
> > >  };
> > >
> > >  static inline void kvm_arch_hardware_unsetup(void) {}
> > > diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > new file mode 100644
> > > index 0000000..e2b4038
> > > --- /dev/null
> > > +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > @@ -0,0 +1,78 @@
> > > +/* SPDX-License-Identifier: GPL-2.0-only */
> > > +/*
> > > + * Copyright (c) 2023 Rivos Inc
> > > + *
> > > + * Authors:
> > > + *     Atish Patra <[e-mail]>
> > > + */
> > > +
> > > +#ifndef __KVM_VCPU_RISCV_PMU_H
> > > +#define __KVM_VCPU_RISCV_PMU_H
> > > +
> > > +#include <linux/perf/riscv_pmu.h>
> > > +#include <asm/kvm_vcpu_sbi.h>
> > > +#include <asm/sbi.h>
> > > +
> > > +#ifdef CONFIG_RISCV_PMU_SBI
> > > +#define RISCV_KVM_MAX_FW_CTRS        32
> > > +
> > > +#if RISCV_KVM_MAX_FW_CTRS > 32
> > > +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
> >
> > "The number of firmware counters cannot exceed 32 without increasing RISCV_MAX_COUNTERS"
> >
> > > +#endif
> > > +
> > > +#define RISCV_MAX_COUNTERS      64
> >
> > But instead of that message, what I think we need is something like
> >
> >  #define RISCV_KVM_MAX_HW_CTRS  32
> >  #define RISCV_KVM_MAX_FW_CTRS  32
> >  #define RISCV_MAX_COUNTERS     (RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
> >
> >  static_assert(RISCV_MAX_COUNTERS <= 64)
> >
> > And then in pmu_sbi_device_probe() should ensure
> >
> >   num_counters <= RISCV_MAX_COUNTERS
> >
> > and pmu_sbi_get_ctrinfo() should ensure
> >
> >   num_hw_ctr <= RISCV_KVM_MAX_HW_CTRS
> >   num_fw_ctr <= RISCV_KVM_MAX_FW_CTRS
> >
> > which has to be done at runtime.
> >
>
> Sure. I will add the additional sanity checks.
>

As explained above, I feel we shouldn't mix the firmware number of
counters that the host gets and it exposes to a guest.
So I have not included this suggestion in the v5.
I have changed the num_fw_ctrs to PMU_FW_MAX though to accurately
reflect the firmware counters KVM is actually using.
I don't know if there is any benefit of static_assert over #error.
Please let me know if you feel strongly about that.


> > > +
> > > +/* Per virtual pmu counter data */
> > > +struct kvm_pmc {
> > > +     u8 idx;
> > > +     struct perf_event *perf_event;
> > > +     uint64_t counter_val;
> > > +     union sbi_pmu_ctr_info cinfo;
> > > +     /* Event monitoring status */
> > > +     bool started;
> > > +};
> > > +
> > > +/* PMU data structure per vcpu */
> > > +struct kvm_pmu {
> > > +     struct kvm_pmc pmc[RISCV_MAX_COUNTERS];
> > > +     /* Number of the virtual firmware counters available */
> > > +     int num_fw_ctrs;
> > > +     /* Number of the virtual hardware counters available */
> > > +     int num_hw_ctrs;
> > > +     /* A flag to indicate that pmu initialization is done */
> > > +     bool init_done;
> > > +     /* Bit map of all the virtual counter used */
> > > +     DECLARE_BITMAP(pmc_in_use, RISCV_MAX_COUNTERS);
> > > +};
> > > +
> > > +#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu_context)
> > > +#define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu_context))
> > > +
> > > +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata);
> > > +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > +                             struct kvm_vcpu_sbi_return *retdata);
> > > +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > +                              unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> > > +                              struct kvm_vcpu_sbi_return *retdata);
> > > +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > +                             unsigned long ctr_mask, unsigned long flag,
> > > +                             struct kvm_vcpu_sbi_return *retdata);
> > > +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > +                                  unsigned long ctr_mask, unsigned long flag,
> > > +                                  unsigned long eidx, uint64_t evtdata,
> > > +                                  struct kvm_vcpu_sbi_return *retdata);
> >
> > s/flag/flags/ for all the above prototypes and all the implementations
> > below.
> >
>
> Fixed.
>
> > > +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > +                             struct kvm_vcpu_sbi_return *retdata);
> > > +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu);
> > > +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu);
> > > +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu);
> > > +
> > > +#else
> > > +struct kvm_pmu {
> > > +};
> > > +
> > > +static inline void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu) {}
> > > +static inline void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu) {}
> > > +static inline void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu) {}
> > > +#endif /* CONFIG_RISCV_PMU_SBI */
> > > +#endif /* !__KVM_VCPU_RISCV_PMU_H */
> > > diff --git a/arch/riscv/kvm/Makefile b/arch/riscv/kvm/Makefile
> > > index 019df920..5de1053 100644
> > > --- a/arch/riscv/kvm/Makefile
> > > +++ b/arch/riscv/kvm/Makefile
> > > @@ -25,3 +25,4 @@ kvm-y += vcpu_sbi_base.o
> > >  kvm-y += vcpu_sbi_replace.o
> > >  kvm-y += vcpu_sbi_hsm.o
> > >  kvm-y += vcpu_timer.o
> > > +kvm-$(CONFIG_RISCV_PMU_SBI) += vcpu_pmu.o
> > > diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> > > index 7c08567..7d010b0 100644
> > > --- a/arch/riscv/kvm/vcpu.c
> > > +++ b/arch/riscv/kvm/vcpu.c
> > > @@ -138,6 +138,8 @@ static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu)
> > >       WRITE_ONCE(vcpu->arch.irqs_pending, 0);
> > >       WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0);
> > >
> > > +     kvm_riscv_vcpu_pmu_reset(vcpu);
> > > +
> > >       vcpu->arch.hfence_head = 0;
> > >       vcpu->arch.hfence_tail = 0;
> > >       memset(vcpu->arch.hfence_queue, 0, sizeof(vcpu->arch.hfence_queue));
> > > @@ -194,6 +196,9 @@ int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
> > >       /* Setup VCPU timer */
> > >       kvm_riscv_vcpu_timer_init(vcpu);
> > >
> > > +     /* setup performance monitoring */
> > > +     kvm_riscv_vcpu_pmu_init(vcpu);
> > > +
> > >       /* Reset VCPU */
> > >       kvm_riscv_reset_vcpu(vcpu);
> > >
> > > @@ -216,6 +221,8 @@ void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
> > >       /* Cleanup VCPU timer */
> > >       kvm_riscv_vcpu_timer_deinit(vcpu);
> > >
> > > +     kvm_riscv_vcpu_pmu_deinit(vcpu);
> > > +
> > >       /* Free unused pages pre-allocated for G-stage page table mappings */
> > >       kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
> > >  }
> > > diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
> > > new file mode 100644
> > > index 0000000..2dad37f
> > > --- /dev/null
> > > +++ b/arch/riscv/kvm/vcpu_pmu.c
> > > @@ -0,0 +1,136 @@
> > > +// SPDX-License-Identifier: GPL-2.0
> > > +/*
> > > + * Copyright (c) 2023 Rivos Inc
> > > + *
> > > + * Authors:
> > > + *     Atish Patra <[e-mail]>
> > > + */
> > > +
> > > +#include <linux/errno.h>
> > > +#include <linux/err.h>
> > > +#include <linux/kvm_host.h>
> > > +#include <linux/perf/riscv_pmu.h>
> > > +#include <asm/csr.h>
> > > +#include <asm/kvm_vcpu_sbi.h>
> > > +#include <asm/kvm_vcpu_pmu.h>
> > > +#include <linux/kvm_host.h>
> > > +
> > > +#define kvm_pmu_num_counters(pmu) ((pmu)->num_hw_ctrs + (pmu)->num_fw_ctrs)
> > > +
> > > +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata)
> > > +{
> > > +     struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > > +
> > > +     retdata->out_val = kvm_pmu_num_counters(kvpmu);
> > > +
> > > +     return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > +                             struct kvm_vcpu_sbi_return *retdata)
> > > +{
> > > +     struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > > +
> > > +     if (cidx > RISCV_MAX_COUNTERS || cidx == 1) {
> > > +             retdata->err_val = SBI_ERR_INVALID_PARAM;
> > > +             return 0;
> > > +     }
> > > +
> > > +     retdata->out_val = kvpmu->pmc[cidx].cinfo.value;
> > > +
> > > +     return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > +                              unsigned long ctr_mask, unsigned long flag, uint64_t ival,
> > > +                              struct kvm_vcpu_sbi_return *retdata)
> > > +{
> > > +     /* TODO */
> > > +     return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > +                             unsigned long ctr_mask, unsigned long flag,
> > > +                             struct kvm_vcpu_sbi_return *retdata)
> > > +{
> > > +     /* TODO */
> > > +     return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > +                                  unsigned long ctr_mask, unsigned long flag,
> > > +                                  unsigned long eidx, uint64_t evtdata,
> > > +                                  struct kvm_vcpu_sbi_return *retdata)
> > > +{
> > > +     /* TODO */
> > > +     return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > +                             struct kvm_vcpu_sbi_return *retdata)
> > > +{
> > > +     /* TODO */
> > > +     return 0;
> > > +}
> > > +
> > > +void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
> > > +{
> > > +     int i = 0, ret, num_hw_ctrs = 0, hpm_width = 0;
> > > +     struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > > +     struct kvm_pmc *pmc;
> > > +
> > > +     ret = riscv_pmu_get_hpm_info(&hpm_width, &num_hw_ctrs);
> > > +     if (ret < 0 || !hpm_width || !num_hw_ctrs)
> > > +             return;
> > > +
> > > +     /*
> > > +      * It is guranteed that RISCV_KVM_MAX_FW_CTRS can't exceed 32 as
> > > +      * that may exceed total number of counters more than RISCV_MAX_COUNTERS
> > > +      */
> > > +     kvpmu->num_hw_ctrs = num_hw_ctrs;
> > > +     kvpmu->num_fw_ctrs = RISCV_KVM_MAX_FW_CTRS;
> >
> > If we sanity check that num_hw_ctrs <= 32 and num_fw_ctrs <= 32 at sbi_pmu
> > probe time, then we can also return num_fw_ctrs (or num_ctrs) along with
> > num_hw_ctrs from riscv_pmu_get_hpm_info(). Then, we can put the exact
> > number here into kvmpmu->num_fw_ctrs, rather than using its max.
> >
>
> The firmware counter information retrieved from PMU driver will be the
> number of firmware
> counter host supports (i.e. M-mode firmware supports). The number of
> counters supported for a
> guest is entirely up to the hypervisor. There shouldn't be any
> relation with the host's firmware counter.
>
> Looking at it again, we should probably set kvpmu->num_fw_ctrs to
> SBI_PMU_FW_MAX instead of RISCV_KVM_MAX_FW_CTRS.
> We already have a sanity check for SBI_PMU_FW_MAX in the code.
> > > +
> > > +     /*
> > > +      * There is no correlation between the logical hardware counter and virtual counters.
> > > +      * However, we need to encode a hpmcounter CSR in the counter info field so that
> > > +      * KVM can trap n emulate the read. This works well in the migration use case as
> > > +      * KVM doesn't care if the actual hpmcounter is available in the hardware or not.
> > > +      */
> > > +     for (i = 0; i < kvm_pmu_num_counters(kvpmu); i++) {
> > > +             /* TIME CSR shouldn't be read from perf interface */
> > > +             if (i == 1)
> > > +                     continue;
> > > +             pmc = &kvpmu->pmc[i];
> > > +             pmc->idx = i;
> > > +             if (i < kvpmu->num_hw_ctrs) {
> > > +                     pmc->cinfo.type = SBI_PMU_CTR_TYPE_HW;
> > > +                     if (i < 3)
> > > +                             /* CY, IR counters */
> > > +                             pmc->cinfo.width = 63;
> > > +                     else
> > > +                             pmc->cinfo.width = hpm_width;
> > > +                     /*
> > > +                      * The CSR number doesn't have any relation with the logical
> > > +                      * hardware counters. The CSR numbers are encoded sequentially
> > > +                      * to avoid maintaining a map between the virtual counter
> > > +                      * and CSR number.
> > > +                      */
> > > +                     pmc->cinfo.csr = CSR_CYCLE + i;
> > > +             } else {
> > > +                     pmc->cinfo.type = SBI_PMU_CTR_TYPE_FW;
> > > +                     pmc->cinfo.width = BITS_PER_LONG - 1;
> > > +             }
> > > +     }
> > > +
> > > +     kvpmu->init_done = true;
> > > +}
> > > +
> > > +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu)
> > > +{
> > > +     /* TODO */
> > > +}
> > > +
> > > +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu)
> > > +{
> > > +     kvm_riscv_vcpu_pmu_deinit(vcpu);
> > > +}
> > > --
> > > 2.25.1
> > >
> >
> > Thanks,
> > drew
>
>
>
> --
> Regards,
> Atish



--
Regards,
Atish

Andrew Jones Feb. 6, 2023, 9:22 a.m. UTC | #8

On Sat, Feb 04, 2023 at 11:37:47PM -0800, Atish Patra wrote:
> On Fri, Feb 3, 2023 at 12:47 AM Atish Patra <[e-mail]> wrote:
> >
> > On Thu, Feb 2, 2023 at 9:03 AM Andrew Jones <[e-mail]> wrote:
> > >
> > > On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> > > > This patch only adds barebone structure of perf implementation. Most of
> > > > the function returns zero at this point and will be implemented
> > > > fully in the future.
> > > >
> > > > Signed-off-by: Atish Patra <[e-mail]>
> > > > ---
> > > >  arch/riscv/include/asm/kvm_host.h     |   4 +
> > > >  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
> > > >  arch/riscv/kvm/Makefile               |   1 +
> > > >  arch/riscv/kvm/vcpu.c                 |   7 ++
> > > >  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
> > > >  5 files changed, 226 insertions(+)
> > > >  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > >  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
> > > >
> > > > diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> > > > index 93f43a3..b90be9a 100644
> > > > --- a/arch/riscv/include/asm/kvm_host.h
> > > > +++ b/arch/riscv/include/asm/kvm_host.h
> > > > @@ -18,6 +18,7 @@
> > > >  #include <asm/kvm_vcpu_insn.h>
> > > >  #include <asm/kvm_vcpu_sbi.h>
> > > >  #include <asm/kvm_vcpu_timer.h>
> > > > +#include <asm/kvm_vcpu_pmu.h>
> > > >
> > > >  #define KVM_MAX_VCPUS                        1024
> > > >
> > > > @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
> > > >
> > > >       /* Don't run the VCPU (blocked) */
> > > >       bool pause;
> > > > +
> > > > +     /* Performance monitoring context */
> > > > +     struct kvm_pmu pmu_context;
> > > >  };
> > > >
> > > >  static inline void kvm_arch_hardware_unsetup(void) {}
> > > > diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > new file mode 100644
> > > > index 0000000..e2b4038
> > > > --- /dev/null
> > > > +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > @@ -0,0 +1,78 @@
> > > > +/* SPDX-License-Identifier: GPL-2.0-only */
> > > > +/*
> > > > + * Copyright (c) 2023 Rivos Inc
> > > > + *
> > > > + * Authors:
> > > > + *     Atish Patra <[e-mail]>
> > > > + */
> > > > +
> > > > +#ifndef __KVM_VCPU_RISCV_PMU_H
> > > > +#define __KVM_VCPU_RISCV_PMU_H
> > > > +
> > > > +#include <linux/perf/riscv_pmu.h>
> > > > +#include <asm/kvm_vcpu_sbi.h>
> > > > +#include <asm/sbi.h>
> > > > +
> > > > +#ifdef CONFIG_RISCV_PMU_SBI
> > > > +#define RISCV_KVM_MAX_FW_CTRS        32
> > > > +
> > > > +#if RISCV_KVM_MAX_FW_CTRS > 32
> > > > +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
> > >
> > > "The number of firmware counters cannot exceed 32 without increasing RISCV_MAX_COUNTERS"
> > >
> > > > +#endif
> > > > +
> > > > +#define RISCV_MAX_COUNTERS      64
> > >
> > > But instead of that message, what I think we need is something like
> > >
> > >  #define RISCV_KVM_MAX_HW_CTRS  32
> > >  #define RISCV_KVM_MAX_FW_CTRS  32
> > >  #define RISCV_MAX_COUNTERS     (RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
> > >
> > >  static_assert(RISCV_MAX_COUNTERS <= 64)
> > >
> > > And then in pmu_sbi_device_probe() should ensure
> > >
> > >   num_counters <= RISCV_MAX_COUNTERS
> > >
> > > and pmu_sbi_get_ctrinfo() should ensure
> > >
> > >   num_hw_ctr <= RISCV_KVM_MAX_HW_CTRS
> > >   num_fw_ctr <= RISCV_KVM_MAX_FW_CTRS
> > >
> > > which has to be done at runtime.
> > >
> >
> > Sure. I will add the additional sanity checks.
> >
> 
> As explained above, I feel we shouldn't mix the firmware number of
> counters that the host gets and it exposes to a guest.
> So I have not included this suggestion in the v5.
> I have changed the num_fw_ctrs to PMU_FW_MAX though to accurately
> reflect the firmware counters KVM is actually using.

Sounds good

> I don't know if there is any benefit of static_assert over #error.
> Please let me know if you feel strongly about that.

One "normal" line vs. three #-lines?

Thanks,
drew

Andrew Jones Feb. 6, 2023, 11:39 a.m. UTC | #9

On Mon, Feb 06, 2023 at 10:22:04AM +0100, Andrew Jones wrote:
> On Sat, Feb 04, 2023 at 11:37:47PM -0800, Atish Patra wrote:
> > On Fri, Feb 3, 2023 at 12:47 AM Atish Patra <[e-mail]> wrote:
> > >
> > > On Thu, Feb 2, 2023 at 9:03 AM Andrew Jones <[e-mail]> wrote:
> > > >
> > > > On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> > > > > This patch only adds barebone structure of perf implementation. Most of
> > > > > the function returns zero at this point and will be implemented
> > > > > fully in the future.
> > > > >
> > > > > Signed-off-by: Atish Patra <[e-mail]>
> > > > > ---
> > > > >  arch/riscv/include/asm/kvm_host.h     |   4 +
> > > > >  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
> > > > >  arch/riscv/kvm/Makefile               |   1 +
> > > > >  arch/riscv/kvm/vcpu.c                 |   7 ++
> > > > >  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
> > > > >  5 files changed, 226 insertions(+)
> > > > >  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > >  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
> > > > >
> > > > > diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> > > > > index 93f43a3..b90be9a 100644
> > > > > --- a/arch/riscv/include/asm/kvm_host.h
> > > > > +++ b/arch/riscv/include/asm/kvm_host.h
> > > > > @@ -18,6 +18,7 @@
> > > > >  #include <asm/kvm_vcpu_insn.h>
> > > > >  #include <asm/kvm_vcpu_sbi.h>
> > > > >  #include <asm/kvm_vcpu_timer.h>
> > > > > +#include <asm/kvm_vcpu_pmu.h>
> > > > >
> > > > >  #define KVM_MAX_VCPUS                        1024
> > > > >
> > > > > @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
> > > > >
> > > > >       /* Don't run the VCPU (blocked) */
> > > > >       bool pause;
> > > > > +
> > > > > +     /* Performance monitoring context */
> > > > > +     struct kvm_pmu pmu_context;
> > > > >  };
> > > > >
> > > > >  static inline void kvm_arch_hardware_unsetup(void) {}
> > > > > diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > > new file mode 100644
> > > > > index 0000000..e2b4038
> > > > > --- /dev/null
> > > > > +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > > @@ -0,0 +1,78 @@
> > > > > +/* SPDX-License-Identifier: GPL-2.0-only */
> > > > > +/*
> > > > > + * Copyright (c) 2023 Rivos Inc
> > > > > + *
> > > > > + * Authors:
> > > > > + *     Atish Patra <[e-mail]>
> > > > > + */
> > > > > +
> > > > > +#ifndef __KVM_VCPU_RISCV_PMU_H
> > > > > +#define __KVM_VCPU_RISCV_PMU_H
> > > > > +
> > > > > +#include <linux/perf/riscv_pmu.h>
> > > > > +#include <asm/kvm_vcpu_sbi.h>
> > > > > +#include <asm/sbi.h>
> > > > > +
> > > > > +#ifdef CONFIG_RISCV_PMU_SBI
> > > > > +#define RISCV_KVM_MAX_FW_CTRS        32
> > > > > +
> > > > > +#if RISCV_KVM_MAX_FW_CTRS > 32
> > > > > +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
> > > >
> > > > "The number of firmware counters cannot exceed 32 without increasing RISCV_MAX_COUNTERS"
> > > >
> > > > > +#endif
> > > > > +
> > > > > +#define RISCV_MAX_COUNTERS      64
> > > >
> > > > But instead of that message, what I think we need is something like
> > > >
> > > >  #define RISCV_KVM_MAX_HW_CTRS  32
> > > >  #define RISCV_KVM_MAX_FW_CTRS  32
> > > >  #define RISCV_MAX_COUNTERS     (RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
> > > >
> > > >  static_assert(RISCV_MAX_COUNTERS <= 64)
> > > >
> > > > And then in pmu_sbi_device_probe() should ensure
> > > >
> > > >   num_counters <= RISCV_MAX_COUNTERS
> > > >
> > > > and pmu_sbi_get_ctrinfo() should ensure
> > > >
> > > >   num_hw_ctr <= RISCV_KVM_MAX_HW_CTRS
> > > >   num_fw_ctr <= RISCV_KVM_MAX_FW_CTRS
> > > >
> > > > which has to be done at runtime.
> > > >
> > >
> > > Sure. I will add the additional sanity checks.
> > >
> > 
> > As explained above, I feel we shouldn't mix the firmware number of
> > counters that the host gets and it exposes to a guest.
> > So I have not included this suggestion in the v5.
> > I have changed the num_fw_ctrs to PMU_FW_MAX though to accurately
> > reflect the firmware counters KVM is actually using.
> 
> Sounds good

I just looked at v5. IMO, much of what I proposed above still makes
sense, since what I'm proposing is that the relationship between
RISCV_KVM_MAX_HW_CTRS, RISCV_KVM_MAX_FW_CTRS, RISCV_MAX_COUNTERS, and 64
(our current max bitmap size) be explicitly checked. So, even if we want
RISCV_KVM_MAX_FW_CTRS to be SBI_PMU_FW_MAX, it'd be good to have

#define RISCV_KVM_MAX_HW_CTRS  32

(And a runtime check confirming num_hw_ctrs + 1 <= RISCV_KVM_MAX_HW_CTRS,
 and then either silently capping or issuing a warning and capping)

And, to be sure the sum of RISCV_KVM_MAX_FW_CTRS and RISCV_KVM_MAX_HW_CTRS
doesn't exceed the size of the bitmap

#define RISCV_KVM_MAX_FW_CTRS	SBI_PMU_FW_MAX
#define RISCV_MAX_COUNTERS	(RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
static_assert(RISCV_MAX_COUNTERS <= 64)

Thanks,
drew

> 
> > I don't know if there is any benefit of static_assert over #error.
> > Please let me know if you feel strongly about that.
> 
> One "normal" line vs. three #-lines?
> 
> Thanks,
> drew

Atish Patra Feb. 7, 2023, 9:20 a.m. UTC | #10

On Mon, Feb 6, 2023 at 3:39 AM Andrew Jones <[e-mail]> wrote:
>
> On Mon, Feb 06, 2023 at 10:22:04AM +0100, Andrew Jones wrote:
> > On Sat, Feb 04, 2023 at 11:37:47PM -0800, Atish Patra wrote:
> > > On Fri, Feb 3, 2023 at 12:47 AM Atish Patra <[e-mail]> wrote:
> > > >
> > > > On Thu, Feb 2, 2023 at 9:03 AM Andrew Jones <[e-mail]> wrote:
> > > > >
> > > > > On Wed, Feb 01, 2023 at 03:12:43PM -0800, Atish Patra wrote:
> > > > > > This patch only adds barebone structure of perf implementation. Most of
> > > > > > the function returns zero at this point and will be implemented
> > > > > > fully in the future.
> > > > > >
> > > > > > Signed-off-by: Atish Patra <[e-mail]>
> > > > > > ---
> > > > > >  arch/riscv/include/asm/kvm_host.h     |   4 +
> > > > > >  arch/riscv/include/asm/kvm_vcpu_pmu.h |  78 +++++++++++++++
> > > > > >  arch/riscv/kvm/Makefile               |   1 +
> > > > > >  arch/riscv/kvm/vcpu.c                 |   7 ++
> > > > > >  arch/riscv/kvm/vcpu_pmu.c             | 136 ++++++++++++++++++++++++++
> > > > > >  5 files changed, 226 insertions(+)
> > > > > >  create mode 100644 arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > > >  create mode 100644 arch/riscv/kvm/vcpu_pmu.c
> > > > > >
> > > > > > diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
> > > > > > index 93f43a3..b90be9a 100644
> > > > > > --- a/arch/riscv/include/asm/kvm_host.h
> > > > > > +++ b/arch/riscv/include/asm/kvm_host.h
> > > > > > @@ -18,6 +18,7 @@
> > > > > >  #include <asm/kvm_vcpu_insn.h>
> > > > > >  #include <asm/kvm_vcpu_sbi.h>
> > > > > >  #include <asm/kvm_vcpu_timer.h>
> > > > > > +#include <asm/kvm_vcpu_pmu.h>
> > > > > >
> > > > > >  #define KVM_MAX_VCPUS                        1024
> > > > > >
> > > > > > @@ -228,6 +229,9 @@ struct kvm_vcpu_arch {
> > > > > >
> > > > > >       /* Don't run the VCPU (blocked) */
> > > > > >       bool pause;
> > > > > > +
> > > > > > +     /* Performance monitoring context */
> > > > > > +     struct kvm_pmu pmu_context;
> > > > > >  };
> > > > > >
> > > > > >  static inline void kvm_arch_hardware_unsetup(void) {}
> > > > > > diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > > > new file mode 100644
> > > > > > index 0000000..e2b4038
> > > > > > --- /dev/null
> > > > > > +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > > > > @@ -0,0 +1,78 @@
> > > > > > +/* SPDX-License-Identifier: GPL-2.0-only */
> > > > > > +/*
> > > > > > + * Copyright (c) 2023 Rivos Inc
> > > > > > + *
> > > > > > + * Authors:
> > > > > > + *     Atish Patra <[e-mail]>
> > > > > > + */
> > > > > > +
> > > > > > +#ifndef __KVM_VCPU_RISCV_PMU_H
> > > > > > +#define __KVM_VCPU_RISCV_PMU_H
> > > > > > +
> > > > > > +#include <linux/perf/riscv_pmu.h>
> > > > > > +#include <asm/kvm_vcpu_sbi.h>
> > > > > > +#include <asm/sbi.h>
> > > > > > +
> > > > > > +#ifdef CONFIG_RISCV_PMU_SBI
> > > > > > +#define RISCV_KVM_MAX_FW_CTRS        32
> > > > > > +
> > > > > > +#if RISCV_KVM_MAX_FW_CTRS > 32
> > > > > > +#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
> > > > >
> > > > > "The number of firmware counters cannot exceed 32 without increasing RISCV_MAX_COUNTERS"
> > > > >
> > > > > > +#endif
> > > > > > +
> > > > > > +#define RISCV_MAX_COUNTERS      64
> > > > >
> > > > > But instead of that message, what I think we need is something like
> > > > >
> > > > >  #define RISCV_KVM_MAX_HW_CTRS  32
> > > > >  #define RISCV_KVM_MAX_FW_CTRS  32
> > > > >  #define RISCV_MAX_COUNTERS     (RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
> > > > >
> > > > >  static_assert(RISCV_MAX_COUNTERS <= 64)
> > > > >
> > > > > And then in pmu_sbi_device_probe() should ensure
> > > > >
> > > > >   num_counters <= RISCV_MAX_COUNTERS
> > > > >
> > > > > and pmu_sbi_get_ctrinfo() should ensure
> > > > >
> > > > >   num_hw_ctr <= RISCV_KVM_MAX_HW_CTRS
> > > > >   num_fw_ctr <= RISCV_KVM_MAX_FW_CTRS
> > > > >
> > > > > which has to be done at runtime.
> > > > >
> > > >
> > > > Sure. I will add the additional sanity checks.
> > > >
> > >
> > > As explained above, I feel we shouldn't mix the firmware number of
> > > counters that the host gets and it exposes to a guest.
> > > So I have not included this suggestion in the v5.
> > > I have changed the num_fw_ctrs to PMU_FW_MAX though to accurately
> > > reflect the firmware counters KVM is actually using.
> >
> > Sounds good
>
> I just looked at v5. IMO, much of what I proposed above still makes
> sense, since what I'm proposing is that the relationship between
> RISCV_KVM_MAX_HW_CTRS, RISCV_KVM_MAX_FW_CTRS, RISCV_MAX_COUNTERS, and 64
> (our current max bitmap size) be explicitly checked. So, even if we want
> RISCV_KVM_MAX_FW_CTRS to be SBI_PMU_FW_MAX, it'd be good to have
>
> #define RISCV_KVM_MAX_HW_CTRS  32
>
> (And a runtime check confirming num_hw_ctrs + 1 <= RISCV_KVM_MAX_HW_CTRS,
>  and then either silently capping or issuing a warning and capping)
>
> And, to be sure the sum of RISCV_KVM_MAX_FW_CTRS and RISCV_KVM_MAX_HW_CTRS
> doesn't exceed the size of the bitmap
>
> #define RISCV_KVM_MAX_FW_CTRS   SBI_PMU_FW_MAX
> #define RISCV_MAX_COUNTERS      (RISCV_KVM_MAX_HW_CTRS + RISCV_KVM_MAX_FW_CTRS)
> static_assert(RISCV_MAX_COUNTERS <= 64)
>

ok. I have added those changes but I have modified RISCV_MAX_COUNTERS
to RISCV_KVM_MAX_COUNTERS,
to avoid overlapping with the RISCV_MAX_COUNTERS defined in the host.
Logically, the host and the guest can have separate counters anyways.

> Thanks,
> drew
>
> >
> > > I don't know if there is any benefit of static_assert over #error.
> > > Please let me know if you feel strongly about that.
> >
> > One "normal" line vs. three #-lines?
> >

Fair enough. Fixed.

> > Thanks,
> > drew

Patch

diff --git a/arch/riscv/include/asm/kvm_host.h b/arch/riscv/include/asm/kvm_host.h
index 93f43a3..b90be9a 100644
--- a/arch/riscv/include/asm/kvm_host.h
+++ b/arch/riscv/include/asm/kvm_host.h
@@ -18,6 +18,7 @@ 
 #include <asm/kvm_vcpu_insn.h>
 #include <asm/kvm_vcpu_sbi.h>
 #include <asm/kvm_vcpu_timer.h>
+#include <asm/kvm_vcpu_pmu.h>
 
 #define KVM_MAX_VCPUS			1024
 
@@ -228,6 +229,9 @@  struct kvm_vcpu_arch {
 
 	/* Don't run the VCPU (blocked) */
 	bool pause;
+
+	/* Performance monitoring context */
+	struct kvm_pmu pmu_context;
 };
 
 static inline void kvm_arch_hardware_unsetup(void) {}
diff --git a/arch/riscv/include/asm/kvm_vcpu_pmu.h b/arch/riscv/include/asm/kvm_vcpu_pmu.h
new file mode 100644
index 0000000..e2b4038
--- /dev/null
+++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
@@ -0,0 +1,78 @@ 
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * Copyright (c) 2023 Rivos Inc
+ *
+ * Authors:
+ *     Atish Patra <[e-mail]>
+ */
+
+#ifndef __KVM_VCPU_RISCV_PMU_H
+#define __KVM_VCPU_RISCV_PMU_H
+
+#include <linux/perf/riscv_pmu.h>
+#include <asm/kvm_vcpu_sbi.h>
+#include <asm/sbi.h>
+
+#ifdef CONFIG_RISCV_PMU_SBI
+#define RISCV_KVM_MAX_FW_CTRS	32
+
+#if RISCV_KVM_MAX_FW_CTRS > 32
+#error "Maximum firmware counter can't exceed 32 without increasing the RISCV_MAX_COUNTERS"
+#endif
+
+#define RISCV_MAX_COUNTERS      64
+
+/* Per virtual pmu counter data */
+struct kvm_pmc {
+	u8 idx;
+	struct perf_event *perf_event;
+	uint64_t counter_val;
+	union sbi_pmu_ctr_info cinfo;
+	/* Event monitoring status */
+	bool started;
+};
+
+/* PMU data structure per vcpu */
+struct kvm_pmu {
+	struct kvm_pmc pmc[RISCV_MAX_COUNTERS];
+	/* Number of the virtual firmware counters available */
+	int num_fw_ctrs;
+	/* Number of the virtual hardware counters available */
+	int num_hw_ctrs;
+	/* A flag to indicate that pmu initialization is done */
+	bool init_done;
+	/* Bit map of all the virtual counter used */
+	DECLARE_BITMAP(pmc_in_use, RISCV_MAX_COUNTERS);
+};
+
+#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu_context)
+#define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu_context))
+
+int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata);
+int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
+				struct kvm_vcpu_sbi_return *retdata);
+int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
+				 unsigned long ctr_mask, unsigned long flag, uint64_t ival,
+				 struct kvm_vcpu_sbi_return *retdata);
+int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
+				unsigned long ctr_mask, unsigned long flag,
+				struct kvm_vcpu_sbi_return *retdata);
+int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
+				     unsigned long ctr_mask, unsigned long flag,
+				     unsigned long eidx, uint64_t evtdata,
+				     struct kvm_vcpu_sbi_return *retdata);
+int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
+				struct kvm_vcpu_sbi_return *retdata);
+void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu);
+void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu);
+void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu);
+
+#else
+struct kvm_pmu {
+};
+
+static inline void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu) {}
+static inline void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu) {}
+static inline void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu) {}
+#endif /* CONFIG_RISCV_PMU_SBI */
+#endif /* !__KVM_VCPU_RISCV_PMU_H */
diff --git a/arch/riscv/kvm/Makefile b/arch/riscv/kvm/Makefile
index 019df920..5de1053 100644
--- a/arch/riscv/kvm/Makefile
+++ b/arch/riscv/kvm/Makefile
@@ -25,3 +25,4 @@  kvm-y += vcpu_sbi_base.o
 kvm-y += vcpu_sbi_replace.o
 kvm-y += vcpu_sbi_hsm.o
 kvm-y += vcpu_timer.o
+kvm-$(CONFIG_RISCV_PMU_SBI) += vcpu_pmu.o
diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
index 7c08567..7d010b0 100644
--- a/arch/riscv/kvm/vcpu.c
+++ b/arch/riscv/kvm/vcpu.c
@@ -138,6 +138,8 @@  static void kvm_riscv_reset_vcpu(struct kvm_vcpu *vcpu)
 	WRITE_ONCE(vcpu->arch.irqs_pending, 0);
 	WRITE_ONCE(vcpu->arch.irqs_pending_mask, 0);
 
+	kvm_riscv_vcpu_pmu_reset(vcpu);
+
 	vcpu->arch.hfence_head = 0;
 	vcpu->arch.hfence_tail = 0;
 	memset(vcpu->arch.hfence_queue, 0, sizeof(vcpu->arch.hfence_queue));
@@ -194,6 +196,9 @@  int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 	/* Setup VCPU timer */
 	kvm_riscv_vcpu_timer_init(vcpu);
 
+	/* setup performance monitoring */
+	kvm_riscv_vcpu_pmu_init(vcpu);
+
 	/* Reset VCPU */
 	kvm_riscv_reset_vcpu(vcpu);
 
@@ -216,6 +221,8 @@  void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 	/* Cleanup VCPU timer */
 	kvm_riscv_vcpu_timer_deinit(vcpu);
 
+	kvm_riscv_vcpu_pmu_deinit(vcpu);
+
 	/* Free unused pages pre-allocated for G-stage page table mappings */
 	kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 }
diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
new file mode 100644
index 0000000..2dad37f
--- /dev/null
+++ b/arch/riscv/kvm/vcpu_pmu.c
@@ -0,0 +1,136 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (c) 2023 Rivos Inc
+ *
+ * Authors:
+ *     Atish Patra <[e-mail]>
+ */
+
+#include <linux/errno.h>
+#include <linux/err.h>
+#include <linux/kvm_host.h>
+#include <linux/perf/riscv_pmu.h>
+#include <asm/csr.h>
+#include <asm/kvm_vcpu_sbi.h>
+#include <asm/kvm_vcpu_pmu.h>
+#include <linux/kvm_host.h>
+
+#define kvm_pmu_num_counters(pmu) ((pmu)->num_hw_ctrs + (pmu)->num_fw_ctrs)
+
+int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, struct kvm_vcpu_sbi_return *retdata)
+{
+	struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
+
+	retdata->out_val = kvm_pmu_num_counters(kvpmu);
+
+	return 0;
+}
+
+int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
+				struct kvm_vcpu_sbi_return *retdata)
+{
+	struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
+
+	if (cidx > RISCV_MAX_COUNTERS || cidx == 1) {
+		retdata->err_val = SBI_ERR_INVALID_PARAM;
+		return 0;
+	}
+
+	retdata->out_val = kvpmu->pmc[cidx].cinfo.value;
+
+	return 0;
+}
+
+int kvm_riscv_vcpu_pmu_ctr_start(struct kvm_vcpu *vcpu, unsigned long ctr_base,
+				 unsigned long ctr_mask, unsigned long flag, uint64_t ival,
+				 struct kvm_vcpu_sbi_return *retdata)
+{
+	/* TODO */
+	return 0;
+}
+
+int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
+				unsigned long ctr_mask, unsigned long flag,
+				struct kvm_vcpu_sbi_return *retdata)
+{
+	/* TODO */
+	return 0;
+}
+
+int kvm_riscv_vcpu_pmu_ctr_cfg_match(struct kvm_vcpu *vcpu, unsigned long ctr_base,
+				     unsigned long ctr_mask, unsigned long flag,
+				     unsigned long eidx, uint64_t evtdata,
+				     struct kvm_vcpu_sbi_return *retdata)
+{
+	/* TODO */
+	return 0;
+}
+
+int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
+				struct kvm_vcpu_sbi_return *retdata)
+{
+	/* TODO */
+	return 0;
+}
+
+void kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
+{
+	int i = 0, ret, num_hw_ctrs = 0, hpm_width = 0;
+	struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
+	struct kvm_pmc *pmc;
+
+	ret = riscv_pmu_get_hpm_info(&hpm_width, &num_hw_ctrs);
+	if (ret < 0 || !hpm_width || !num_hw_ctrs)
+		return;
+
+	/*
+	 * It is guranteed that RISCV_KVM_MAX_FW_CTRS can't exceed 32 as
+	 * that may exceed total number of counters more than RISCV_MAX_COUNTERS
+	 */
+	kvpmu->num_hw_ctrs = num_hw_ctrs;
+	kvpmu->num_fw_ctrs = RISCV_KVM_MAX_FW_CTRS;
+
+	/*
+	 * There is no correlation between the logical hardware counter and virtual counters.
+	 * However, we need to encode a hpmcounter CSR in the counter info field so that
+	 * KVM can trap n emulate the read. This works well in the migration use case as
+	 * KVM doesn't care if the actual hpmcounter is available in the hardware or not.
+	 */
+	for (i = 0; i < kvm_pmu_num_counters(kvpmu); i++) {
+		/* TIME CSR shouldn't be read from perf interface */
+		if (i == 1)
+			continue;
+		pmc = &kvpmu->pmc[i];
+		pmc->idx = i;
+		if (i < kvpmu->num_hw_ctrs) {
+			pmc->cinfo.type = SBI_PMU_CTR_TYPE_HW;
+			if (i < 3)
+				/* CY, IR counters */
+				pmc->cinfo.width = 63;
+			else
+				pmc->cinfo.width = hpm_width;
+			/*
+			 * The CSR number doesn't have any relation with the logical
+			 * hardware counters. The CSR numbers are encoded sequentially
+			 * to avoid maintaining a map between the virtual counter
+			 * and CSR number.
+			 */
+			pmc->cinfo.csr = CSR_CYCLE + i;
+		} else {
+			pmc->cinfo.type = SBI_PMU_CTR_TYPE_FW;
+			pmc->cinfo.width = BITS_PER_LONG - 1;
+		}
+	}
+
+	kvpmu->init_done = true;
+}
+
+void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu)
+{
+	/* TODO */
+}
+
+void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu)
+{
+	kvm_riscv_vcpu_pmu_deinit(vcpu);
+}