Re: [RFC 5/9] RISC-V: KVM: Add skeleton support for perf

From: Atish Patra
Date: Tue Nov 22 2022 - 20:35:10 EST


On Tue, Nov 22, 2022 at 4:46 PM Atish Patra <atishp@xxxxxxxxxxxxxx> wrote:
>
> at runtime.
>
> On Tue, Nov 1, 2022 at 7:13 AM Andrew Jones <ajones@xxxxxxxxxxxxxxxx> wrote:
> >
> > On Mon, Jul 18, 2022 at 10:02:01AM -0700, Atish Patra wrote:
> > > This patch only adds barebore structure of perf implementation. Most of
> > a bare bones ^ the
> >
> > > the function returns zero at this point and will be implemented
> > functions
> >
> > > fully in the future.
> >
> > s/the future/later patches/
> >
>
> Done.
>
> > >
> > > Signed-off-by: Atish Patra <atishp@xxxxxxxxxxxx>
> > > ---
> > > arch/riscv/include/asm/kvm_host.h | 3 +
> > > arch/riscv/include/asm/kvm_vcpu_pmu.h | 70 +++++++++++++
> > > arch/riscv/kvm/Makefile | 1 +
> > > arch/riscv/kvm/main.c | 3 +-
> > > arch/riscv/kvm/vcpu.c | 5 +
> > > arch/riscv/kvm/vcpu_insn.c | 3 +-
> > > arch/riscv/kvm/vcpu_pmu.c | 136 ++++++++++++++++++++++++++
> > > 7 files changed, 219 insertions(+), 2 deletions(-)
> > > 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 59a0cf2ca7b9..5d2312828bb2 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_fp.h>
> > > #include <asm/kvm_vcpu_insn.h>
> > > #include <asm/kvm_vcpu_timer.h>
> > > +#include <asm/kvm_vcpu_pmu.h>
> > >
> > > #define KVM_MAX_VCPUS 1024
> > >
> > > @@ -226,6 +227,8 @@ struct kvm_vcpu_arch {
> > >
> > > /* Don't run the VCPU (blocked) */
> > > bool pause;
> > > +
> > > + struct kvm_pmu pmu;
> > > };
> > >
> > > 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 000000000000..bffee052f2ae
> > > --- /dev/null
> > > +++ b/arch/riscv/include/asm/kvm_vcpu_pmu.h
> > > @@ -0,0 +1,70 @@
> > > +/* SPDX-License-Identifier: GPL-2.0-only */
> > > +/*
> > > + * Copyright (c) 2022 Rivos Inc
> > > + *
> > > + * Authors:
> > > + * Atish Patra <atishp@xxxxxxxxxxxx>
> > > + */
> > > +
> > > +#ifndef _KVM_VCPU_RISCV_PMU_H
> > > +#define _KVM_VCPU_RISCV_PMU_H
> >
> > The convention seems to be to leading underscores for these types of
> > defines, i.e. __KVM_VCPU_RISCV_PMU_H
> >
>
> Yes. It was a typo. Fixed.
>
> > > +
> > > +#include <linux/perf/riscv_pmu.h>
> > > +#include <asm/sbi.h>
> > > +
> > > +#ifdef CONFIG_RISCV_PMU_SBI
> > > +#define RISCV_KVM_MAX_FW_CTRS 32
> > > +
> > > +/* Per virtual pmu counter data */
> > > +struct kvm_pmc {
> > > + u8 idx;
> > > + struct kvm_vcpu *vcpu;
> >
> > I'm not sure we need a vcpu pointer here. If it's just to implement
> > pmc_to_pmu(), then we can instead implement a pmc_to_vcpu(), like
> > arm64's kvm_pmc_to_vcpu(). x86 might be able to do that too, since
> > it appears the conversion macros below originated there.
> >
>
> Yes. We can implement arm64 as well instead of x86.
> I just thought the x86 approach keeping a reference to vcpu is a bit
> simpler than computing
> the parent offset using container_of multiple times. This will be
> invoked once per overflow in the counter overflow path.
>
> If you feel strongly about it arm64 way, we can follow that. I have
> removed from this series for now.
> Depending on the conclusion, I will add it back in kvm sscofpmf
> support series if required.
>
> > > + struct perf_event *perf_event;
> > > + uint64_t counter_val;
> > > + union sbi_pmu_ctr_info cinfo;
> > > +};
> > > +
> > > +/* 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;
> > > + /* Bit map of all the virtual counter used */
> > counters
> >
> > > + DECLARE_BITMAP(used_pmc, RISCV_MAX_COUNTERS);
> >
> > How about naming this pmc_in_use like x86?
> >
>
> Done.
>
> > > +};
> > > +
> > > +#define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu)
> > > +#define pmu_to_vcpu(pmu) (container_of((pmu), struct kvm_vcpu, arch.pmu))
> > > +#define pmc_to_pmu(pmc) (&(pmc)->vcpu->arch.pmu)
> > > +
> > > +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, unsigned long *out_val);
> > > +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > + unsigned long *ctr_info);
> > > +
> >
> > nit: no need for this blank line
>
> Fixed.
>
> >
> > > +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);
> > > +int kvm_riscv_vcpu_pmu_ctr_stop(struct kvm_vcpu *vcpu, unsigned long ctr_base,
> > > + unsigned long ctr_mask, unsigned long flag);
> > > +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 edata);
> > > +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > + unsigned long *out_val);
> > > +int 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 int kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
> > > +{
> > > + return 0;
> > > +}
> > > +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
> > > +#endif
> > > diff --git a/arch/riscv/kvm/Makefile b/arch/riscv/kvm/Makefile
> > > index 019df9208bdd..342d7199e89d 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_sbi_pmu.o vcpu_pmu.o
> > > diff --git a/arch/riscv/kvm/main.c b/arch/riscv/kvm/main.c
> > > index 1549205fe5fe..d41ab6d1987d 100644
> > > --- a/arch/riscv/kvm/main.c
> > > +++ b/arch/riscv/kvm/main.c
> > > @@ -49,7 +49,8 @@ int kvm_arch_hardware_enable(void)
> > > hideleg |= (1UL << IRQ_VS_EXT);
> > > csr_write(CSR_HIDELEG, hideleg);
> > >
> > > - csr_write(CSR_HCOUNTEREN, -1UL);
> > > + /* VS should access only TM bit. Everything else should trap */
> > > + csr_write(CSR_HCOUNTEREN, 0x02);
> >
> > This looks like something that should be broken out into a separate patch
> > with a description of what happens now when guests try to access the newly
> > trapping counter registers. We should probably also create a TM define.
> >
>
> Done.
>

As we allow cycles & instret for host user space now [1], should we do the same
for guests as well ? I would prefer not to but same user space
software will start to break
they will run inside a guest.

https://lore.kernel.org/all/20220928131807.30386-1-palmer@xxxxxxxxxxxx/

> > >
> > > csr_write(CSR_HVIP, 0);
> > >
> > > diff --git a/arch/riscv/kvm/vcpu.c b/arch/riscv/kvm/vcpu.c
> > > index 3c95924d38c7..4cc964aaf2ad 100644
> > > --- a/arch/riscv/kvm/vcpu.c
> > > +++ b/arch/riscv/kvm/vcpu.c
> > > @@ -122,6 +122,7 @@ 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;
> > > @@ -174,6 +175,9 @@ 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);
> > >
> > > @@ -196,6 +200,7 @@ 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_insn.c b/arch/riscv/kvm/vcpu_insn.c
> > > index 7eb90a47b571..0aa334f853c8 100644
> > > --- a/arch/riscv/kvm/vcpu_insn.c
> > > +++ b/arch/riscv/kvm/vcpu_insn.c
> > > @@ -214,7 +214,8 @@ struct csr_func {
> > > unsigned long wr_mask);
> > > };
> > >
> > > -static const struct csr_func csr_funcs[] = { };
> > > +static const struct csr_func csr_funcs[] = {
> > > +};
> >
> > stray change
> >
>
> Fixed
>
> > >
> > > /**
> > > * kvm_riscv_vcpu_csr_return -- Handle CSR read/write after user space
> > > diff --git a/arch/riscv/kvm/vcpu_pmu.c b/arch/riscv/kvm/vcpu_pmu.c
> > > new file mode 100644
> > > index 000000000000..3168ed740bdd
> > > --- /dev/null
> > > +++ b/arch/riscv/kvm/vcpu_pmu.c
> > > @@ -0,0 +1,136 @@
> > > +// SPDX-License-Identifier: GPL-2.0
> > > +/*
> > > + * Copyright (c) 2022 Rivos Inc
> > > + *
> > > + * Authors:
> > > + * Atish Patra <atishp@xxxxxxxxxxxx>
> > > + */
> > > +
> > > +#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_pmu.h>
> > > +#include <linux/kvm_host.h>
> > > +
> > > +int kvm_riscv_vcpu_pmu_num_ctrs(struct kvm_vcpu *vcpu, unsigned long *out_val)
> > > +{
> > > + struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > > +
> > > + if (!kvpmu)
> > > + return -EINVAL;
> >
> > kvpmu can never be null because arch.pmu isn't a pointer. We probably
> > shouldn't be making calls to kvm_riscv_vcpu_pmu_num_ctrs() without knowing
> > we have an initialized pmu anyway, though.
> >
>
> Yes. I have added an init_done flag to do that sanity check.
> I can change it based on the conclusion on PATCH 6.
>
> > > +
> > > + *out_val = kvpmu->num_fw_ctrs + kvpmu->num_hw_ctrs;
> > > + return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_info(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > + unsigned long *ctr_info)
> > > +{
> > > + struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > > +
> > > + if (!kvpmu || (cidx > RISCV_MAX_COUNTERS) || (cidx == 1))
> >
> > nit: unnecessary ()
> >
> > > + return -EINVAL;
> > > +
> > > + *ctr_info = 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)
> > > +{
> > > + /* 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)
> > > +{
> > > + /* 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 edata)
> > > +{
> > > + /* TODO */
> > > + return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_ctr_read(struct kvm_vcpu *vcpu, unsigned long cidx,
> > > + unsigned long *out_val)
> > > +{
> > > + /* TODO */
> > > + return 0;
> > > +}
> > > +
> > > +int kvm_riscv_vcpu_pmu_init(struct kvm_vcpu *vcpu)
> > > +{
> > > + int i = 0, num_hw_ctrs, num_fw_ctrs, hpm_width;
> > > + struct kvm_pmu *kvpmu = vcpu_to_pmu(vcpu);
> > > +
> > > + if (!kvpmu)
> > > + return -EINVAL;
> > > +
> > > + num_hw_ctrs = riscv_pmu_sbi_get_num_hw_ctrs();
> > > + if ((num_hw_ctrs + RISCV_KVM_MAX_FW_CTRS) > RISCV_MAX_COUNTERS)
> > > + num_fw_ctrs = RISCV_MAX_COUNTERS - num_hw_ctrs;
> > > + else
> > > + num_fw_ctrs = RISCV_KVM_MAX_FW_CTRS;
> >
> > Why do we need RISCV_KVM_MAX_FW_CTRS? Can't we just always get the number
> > with RISCV_MAX_COUNTERS - num_hw_ctrs ?
> >
> We can. But we have to allocate fw_event at runtime. As most platforms
> don't implement
> more than all 29 hpmcounters, you end up having more firmware counters
> than needed.
> Current, SBI spec only define 21 firmware counter anyways.
>
> Thus I felt it is unnecessary to do runtime allocation. But it's just
> few bytes. So I don't feel
> strongly about it.
>
> > > +
> > > + hpm_width = riscv_pmu_sbi_hpmc_width();
> > > + if (hpm_width <= 0) {
> > > + pr_err("Can not initialize PMU for vcpu as hpmcounter width is not available\n");
> > > + return -EINVAL;
> > > + }
> > > +
> > > + kvpmu->num_hw_ctrs = num_hw_ctrs;
> > > + kvpmu->num_fw_ctrs = num_fw_ctrs;
> >
> > Maybe it's coming later, but we need to give KVM userspace control over
> > the number of counters to allow it to migrate to a larger set of hosts.
> > Also, a previous patch said the virtual width must be the same as the
> > host width for the hw counters, so we need userspace to know what that
> > is in order to determine to which hosts it can migrate a guest.
> >
>
> Yes. The entire user space access control needs to be sketched out.
> We probably need another one reg interface to set/get the number of
> counters/width.
>
> However, Is it a common to migrate a guest between different hosts
> with different PMU capabilities ?
>
> > > + /*
> > > + * There is no corelation betwen 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 migraiton usecase as well
> >
> > s/well//
> >
> > > + * KVM doesn't care if the actual hpmcounter is available in the hardware or not.
> > > + */
> > > + for (i = 0; i < num_hw_ctrs + num_fw_ctrs; i++) {
> >
> > Maybe we need a helper macro like
> >
> > #define kvm_pmu_num_counters(pmu) ((pmu)->num_hw_ctrs + (pmu)->num_fw_ctrs)
> >
> > if we're going to loop over all counters frequently.
> >
>
> Done.
>
>
> > > + /* TIME CSR shouldn't be read from perf interface */
> > > + if (i == 1)
> > > + continue;
> > > + kvpmu->pmc[i].idx = i;
> > > + kvpmu->pmc[i].vcpu = vcpu;
> > > + if (i < kvpmu->num_hw_ctrs) {
> > > + kvpmu->pmc[i].cinfo.type = SBI_PMU_CTR_TYPE_HW;
> > > + if (i < 3)
> > > + /* CY, IR counters */
> > > + kvpmu->pmc[i].cinfo.width = 63;
> > > + else
> > > + kvpmu->pmc[i].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.
> > > + */
> > > + kvpmu->pmc[i].cinfo.csr = CSR_CYCLE + i;
> > > + } else {
> > > + kvpmu->pmc[i].cinfo.type = SBI_PMU_CTR_TYPE_FW;
> > > + kvpmu->pmc[i].cinfo.width = BITS_PER_LONG - 1;
> > > + }
> > > + }
> > > +
> > > + return 0;
> > > +}
> > > +
> > > +void kvm_riscv_vcpu_pmu_reset(struct kvm_vcpu *vcpu)
> > > +{
> > > + /* TODO */
> > > +}
> > > +
> > > +void kvm_riscv_vcpu_pmu_deinit(struct kvm_vcpu *vcpu)
> > > +{
> > > + /* TODO */
> > > +}
> > > +
> > > --
> > > 2.25.1
> > >
> >
> > Thanks,
> > drew
>
>
>
> --
> Regards,
> Atish



--
Regards,
Atish