Quelle  pmu_counters_test.c   Sprache: C

// SPDX-License-Identifier: GPL-2.0
/*
 * Copyright (C) 2023, Tencent, Inc.
 */
#include <x86intrin.h>

#include "pmu.h"
#include "processor.h"

/* Number of iterations of the loop for the guest measurement payload. */
#define NUM_LOOPS   10

/* Each iteration of the loop retires one branch instruction. */
#define NUM_BRANCH_INSNS_RETIRED (NUM_LOOPS)

/*
 * Number of instructions in each loop. 1 CLFLUSH/CLFLUSHOPT/NOP, 1 MFENCE,
 * 1 LOOP.
 */
#define NUM_INSNS_PER_LOOP  4

/*
 * Number of "extra" instructions that will be counted, i.e. the number of
 * instructions that are needed to set up the loop and then disable the
 * counter.  2 MOV, 2 XOR, 1 WRMSR.
 */
#define NUM_EXTRA_INSNS   5

/* Total number of instructions retired within the measured section. */
#define NUM_INSNS_RETIRED  (NUM_LOOPS * NUM_INSNS_PER_LOOP + NUM_EXTRA_INSNS)

/* Track which architectural events are supported by hardware. */
static uint32_t hardware_pmu_arch_events;

static uint8_t kvm_pmu_version;
static bool kvm_has_perf_caps;

#define X86_PMU_FEATURE_NULL      \
({         \
 struct kvm_x86_pmu_feature feature = {};   \
         \
 feature;       \
})

static bool pmu_is_null_feature(struct kvm_x86_pmu_feature event)
{
 return !(*(u64 *)&event);
}

struct kvm_intel_pmu_event {
 struct kvm_x86_pmu_feature gp_event;
 struct kvm_x86_pmu_feature fixed_event;
};

/*
 * Wrap the array to appease the compiler, as the macros used to construct each
 * kvm_x86_pmu_feature use syntax that's only valid in function scope, and the
 * compiler often thinks the feature definitions aren't compile-time constants.
 */
static struct kvm_intel_pmu_event intel_event_to_feature(uint8_t idx)
{
 const struct kvm_intel_pmu_event __intel_event_to_feature[] = {
  [INTEL_ARCH_CPU_CYCLES_INDEX]   = { X86_PMU_FEATURE_CPU_CYCLES, X86_PMU_FEATURE_CPU_CYCLES_FIXED },
  [INTEL_ARCH_INSTRUCTIONS_RETIRED_INDEX]  = { X86_PMU_FEATURE_INSNS_RETIRED, X86_PMU_FEATURE_INSNS_RETIRED_FIXED },
  /*
 * Note, the fixed counter for reference cycles is NOT the same as the
 * general purpose architectural event.  The fixed counter explicitly
 * counts at the same frequency as the TSC, whereas the GP event counts
 * at a fixed, but uarch specific, frequency.  Bundle them here for
 * simplicity.
 */
  [INTEL_ARCH_REFERENCE_CYCLES_INDEX]  = { X86_PMU_FEATURE_REFERENCE_CYCLES, X86_PMU_FEATURE_REFERENCE_TSC_CYCLES_FIXED },
  [INTEL_ARCH_LLC_REFERENCES_INDEX]  = { X86_PMU_FEATURE_LLC_REFERENCES, X86_PMU_FEATURE_NULL },
  [INTEL_ARCH_LLC_MISSES_INDEX]   = { X86_PMU_FEATURE_LLC_MISSES, X86_PMU_FEATURE_NULL },
  [INTEL_ARCH_BRANCHES_RETIRED_INDEX]  = { X86_PMU_FEATURE_BRANCH_INSNS_RETIRED, X86_PMU_FEATURE_NULL },
  [INTEL_ARCH_BRANCHES_MISPREDICTED_INDEX] = { X86_PMU_FEATURE_BRANCHES_MISPREDICTED, X86_PMU_FEATURE_NULL },
  [INTEL_ARCH_TOPDOWN_SLOTS_INDEX]  = { X86_PMU_FEATURE_TOPDOWN_SLOTS, X86_PMU_FEATURE_TOPDOWN_SLOTS_FIXED },
 };

 kvm_static_assert(ARRAY_SIZE(__intel_event_to_feature) == NR_INTEL_ARCH_EVENTS);

 return __intel_event_to_feature[idx];
}

static struct kvm_vm *pmu_vm_create_with_one_vcpu(struct kvm_vcpu **vcpu,
        void *guest_code,
        uint8_t pmu_version,
        uint64_t perf_capabilities)
{
 struct kvm_vm *vm;

 vm = vm_create_with_one_vcpu(vcpu, guest_code);
 sync_global_to_guest(vm, kvm_pmu_version);
 sync_global_to_guest(vm, hardware_pmu_arch_events);

 /*
 * Set PERF_CAPABILITIES before PMU version as KVM disallows enabling
 * features via PERF_CAPABILITIES if the guest doesn't have a vPMU.
 */
 if (kvm_has_perf_caps)
  vcpu_set_msr(*vcpu, MSR_IA32_PERF_CAPABILITIES, perf_capabilities);

 vcpu_set_cpuid_property(*vcpu, X86_PROPERTY_PMU_VERSION, pmu_version);
 return vm;
}

static void run_vcpu(struct kvm_vcpu *vcpu)
{
 struct ucall uc;

 do {
  vcpu_run(vcpu);
  switch (get_ucall(vcpu, &uc)) {
  case UCALL_SYNC:
   break;
  case UCALL_ABORT:
   REPORT_GUEST_ASSERT(uc);
   break;
  case UCALL_PRINTF:
   pr_info("%s", uc.buffer);
   break;
  case UCALL_DONE:
   break;
  default:
   TEST_FAIL("Unexpected ucall: %lu", uc.cmd);
  }
 } while (uc.cmd != UCALL_DONE);
}

static uint8_t guest_get_pmu_version(void)
{
 /*
 * Return the effective PMU version, i.e. the minimum between what KVM
 * supports and what is enumerated to the guest.  The host deliberately
 * advertises a PMU version to the guest beyond what is actually
 * supported by KVM to verify KVM doesn't freak out and do something
 * bizarre with an architecturally valid, but unsupported, version.
 */
 return min_t(uint8_t, kvm_pmu_version, this_cpu_property(X86_PROPERTY_PMU_VERSION));
}

/*
 * If an architectural event is supported and guaranteed to generate at least
 * one "hit, assert that its count is non-zero.  If an event isn't supported or
 * the test can't guarantee the associated action will occur, then all bets are
 * off regarding the count, i.e. no checks can be done.
 *
 * Sanity check that in all cases, the event doesn't count when it's disabled,
 * and that KVM correctly emulates the write of an arbitrary value.
 */
static void guest_assert_event_count(uint8_t idx, uint32_t pmc, uint32_t pmc_msr)
{
 uint64_t count;

 count = _rdpmc(pmc);
 if (!(hardware_pmu_arch_events & BIT(idx)))
  goto sanity_checks;

 switch (idx) {
 case INTEL_ARCH_INSTRUCTIONS_RETIRED_INDEX:
  GUEST_ASSERT_EQ(count, NUM_INSNS_RETIRED);
  break;
 case INTEL_ARCH_BRANCHES_RETIRED_INDEX:
  GUEST_ASSERT_EQ(count, NUM_BRANCH_INSNS_RETIRED);
  break;
 case INTEL_ARCH_LLC_REFERENCES_INDEX:
 case INTEL_ARCH_LLC_MISSES_INDEX:
  if (!this_cpu_has(X86_FEATURE_CLFLUSHOPT) &&
      !this_cpu_has(X86_FEATURE_CLFLUSH))
   break;
  fallthrough;
 case INTEL_ARCH_CPU_CYCLES_INDEX:
 case INTEL_ARCH_REFERENCE_CYCLES_INDEX:
  GUEST_ASSERT_NE(count, 0);
  break;
 case INTEL_ARCH_TOPDOWN_SLOTS_INDEX:
  __GUEST_ASSERT(count >= NUM_INSNS_RETIRED,
          "Expected top-down slots >= %u, got count = %lu",
          NUM_INSNS_RETIRED, count);
  break;
 default:
  break;
 }

sanity_checks:
 __asm__ __volatile__("loop ." : "+c"((int){NUM_LOOPS}));
 GUEST_ASSERT_EQ(_rdpmc(pmc), count);

 wrmsr(pmc_msr, 0xdead);
 GUEST_ASSERT_EQ(_rdpmc(pmc), 0xdead);
}

/*
 * Enable and disable the PMC in a monolithic asm blob to ensure that the
 * compiler can't insert _any_ code into the measured sequence.  Note, ECX
 * doesn't need to be clobbered as the input value, @pmc_msr, is restored
 * before the end of the sequence.
 *
 * If CLFUSH{,OPT} is supported, flush the cacheline containing (at least) the
 * CLFUSH{,OPT} instruction on each loop iteration to force LLC references and
 * misses, i.e. to allow testing that those events actually count.
 *
 * If forced emulation is enabled (and specified), force emulation on a subset
 * of the measured code to verify that KVM correctly emulates instructions and
 * branches retired events in conjunction with hardware also counting said
 * events.
 */
#define GUEST_MEASURE_EVENT(_msr, _value, clflush, FEP)    \
do {          \
 __asm__ __volatile__("wrmsr\n\t"     \
        " mov $" __stringify(NUM_LOOPS) ", %%ecx\n\t" \
        "1:\n\t"      \
        clflush "\n\t"     \
        "mfence\n\t"     \
        "mov %[m], %%eax\n\t"    \
        FEP "loop 1b\n\t"     \
        FEP "mov %%edi, %%ecx\n\t"    \
        FEP "xor %%eax, %%eax\n\t"    \
        FEP "xor %%edx, %%edx\n\t"    \
        "wrmsr\n\t"     \
        :: "a"((uint32_t)_value), "d"(_value >> 32), \
    "c"(_msr), "D"(_msr), [m]"m"(kvm_pmu_version) \
 );         \
} while (0)

#define GUEST_TEST_EVENT(_idx, _pmc, _pmc_msr, _ctrl_msr, _value, FEP)  \
do {          \
 wrmsr(_pmc_msr, 0);       \
          \
 if (this_cpu_has(X86_FEATURE_CLFLUSHOPT))    \
  GUEST_MEASURE_EVENT(_ctrl_msr, _value, "clflushopt %[m]", FEP); \
 else if (this_cpu_has(X86_FEATURE_CLFLUSH))    \
  GUEST_MEASURE_EVENT(_ctrl_msr, _value, "clflush %[m]", FEP); \
 else         \
  GUEST_MEASURE_EVENT(_ctrl_msr, _value, "nop", FEP);  \
          \
 guest_assert_event_count(_idx, _pmc, _pmc_msr);    \
} while (0)

static void __guest_test_arch_event(uint8_t idx, uint32_t pmc, uint32_t pmc_msr,
        uint32_t ctrl_msr, uint64_t ctrl_msr_value)
{
 GUEST_TEST_EVENT(idx, pmc, pmc_msr, ctrl_msr, ctrl_msr_value, "");

 if (is_forced_emulation_enabled)
  GUEST_TEST_EVENT(idx, pmc, pmc_msr, ctrl_msr, ctrl_msr_value, KVM_FEP);
}

static void guest_test_arch_event(uint8_t idx)
{
 uint32_t nr_gp_counters = this_cpu_property(X86_PROPERTY_PMU_NR_GP_COUNTERS);
 uint32_t pmu_version = guest_get_pmu_version();
 /* PERF_GLOBAL_CTRL exists only for Architectural PMU Version 2+. */
 bool guest_has_perf_global_ctrl = pmu_version >= 2;
 struct kvm_x86_pmu_feature gp_event, fixed_event;
 uint32_t base_pmc_msr;
 unsigned int i;

 /* The host side shouldn't invoke this without a guest PMU. */
 GUEST_ASSERT(pmu_version);

 if (this_cpu_has(X86_FEATURE_PDCM) &&
     rdmsr(MSR_IA32_PERF_CAPABILITIES) & PMU_CAP_FW_WRITES)
  base_pmc_msr = MSR_IA32_PMC0;
 else
  base_pmc_msr = MSR_IA32_PERFCTR0;

 gp_event = intel_event_to_feature(idx).gp_event;
 GUEST_ASSERT_EQ(idx, gp_event.f.bit);

 GUEST_ASSERT(nr_gp_counters);

 for (i = 0; i < nr_gp_counters; i++) {
  uint64_t eventsel = ARCH_PERFMON_EVENTSEL_OS |
        ARCH_PERFMON_EVENTSEL_ENABLE |
        intel_pmu_arch_events[idx];

  wrmsr(MSR_P6_EVNTSEL0 + i, 0);
  if (guest_has_perf_global_ctrl)
   wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, BIT_ULL(i));

  __guest_test_arch_event(idx, i, base_pmc_msr + i,
     MSR_P6_EVNTSEL0 + i, eventsel);
 }

 if (!guest_has_perf_global_ctrl)
  return;

 fixed_event = intel_event_to_feature(idx).fixed_event;
 if (pmu_is_null_feature(fixed_event) || !this_pmu_has(fixed_event))
  return;

 i = fixed_event.f.bit;

 wrmsr(MSR_CORE_PERF_FIXED_CTR_CTRL, FIXED_PMC_CTRL(i, FIXED_PMC_KERNEL));

 __guest_test_arch_event(idx, i | INTEL_RDPMC_FIXED,
    MSR_CORE_PERF_FIXED_CTR0 + i,
    MSR_CORE_PERF_GLOBAL_CTRL,
    FIXED_PMC_GLOBAL_CTRL_ENABLE(i));
}

static void guest_test_arch_events(void)
{
 uint8_t i;

 for (i = 0; i < NR_INTEL_ARCH_EVENTS; i++)
  guest_test_arch_event(i);

 GUEST_DONE();
}

static void test_arch_events(uint8_t pmu_version, uint64_t perf_capabilities,
        uint8_t length, uint8_t unavailable_mask)
{
 struct kvm_vcpu *vcpu;
 struct kvm_vm *vm;

 /* Testing arch events requires a vPMU (there are no negative tests). */
 if (!pmu_version)
  return;

 vm = pmu_vm_create_with_one_vcpu(&vcpu, guest_test_arch_events,
      pmu_version, perf_capabilities);

 vcpu_set_cpuid_property(vcpu, X86_PROPERTY_PMU_EBX_BIT_VECTOR_LENGTH,
    length);
 vcpu_set_cpuid_property(vcpu, X86_PROPERTY_PMU_EVENTS_MASK,
    unavailable_mask);

 run_vcpu(vcpu);

 kvm_vm_free(vm);
}

/*
 * Limit testing to MSRs that are actually defined by Intel (in the SDM).  MSRs
 * that aren't defined counter MSRs *probably* don't exist, but there's no
 * guarantee that currently undefined MSR indices won't be used for something
 * other than PMCs in the future.
 */
#define MAX_NR_GP_COUNTERS 8
#define MAX_NR_FIXED_COUNTERS 3

#define GUEST_ASSERT_PMC_MSR_ACCESS(insn, msr, expect_gp, vector)  \
__GUEST_ASSERT(expect_gp ? vector == GP_VECTOR : !vector,   \
        "Expected %s on " #insn "(0x%x), got vector %u",   \
        expect_gp ? "#GP" : "no fault", msr, vector)   \

#define GUEST_ASSERT_PMC_VALUE(insn, msr, val, expected)   \
 __GUEST_ASSERT(val == expected,     \
         "Expected " #insn "(0x%x) to yield 0x%lx, got 0x%lx", \
         msr, expected, val);

static void guest_test_rdpmc(uint32_t rdpmc_idx, bool expect_success,
        uint64_t expected_val)
{
 uint8_t vector;
 uint64_t val;

 vector = rdpmc_safe(rdpmc_idx, &val);
 GUEST_ASSERT_PMC_MSR_ACCESS(RDPMC, rdpmc_idx, !expect_success, vector);
 if (expect_success)
  GUEST_ASSERT_PMC_VALUE(RDPMC, rdpmc_idx, val, expected_val);

 if (!is_forced_emulation_enabled)
  return;

 vector = rdpmc_safe_fep(rdpmc_idx, &val);
 GUEST_ASSERT_PMC_MSR_ACCESS(RDPMC, rdpmc_idx, !expect_success, vector);
 if (expect_success)
  GUEST_ASSERT_PMC_VALUE(RDPMC, rdpmc_idx, val, expected_val);
}

static void guest_rd_wr_counters(uint32_t base_msr, uint8_t nr_possible_counters,
     uint8_t nr_counters, uint32_t or_mask)
{
 const bool pmu_has_fast_mode = !guest_get_pmu_version();
 uint8_t i;

 for (i = 0; i < nr_possible_counters; i++) {
  /*
 * TODO: Test a value that validates full-width writes and the
 * width of the counters.
 */
  const uint64_t test_val = 0xffff;
  const uint32_t msr = base_msr + i;

  /*
 * Fixed counters are supported if the counter is less than the
 * number of enumerated contiguous counters *or* the counter is
 * explicitly enumerated in the supported counters mask.
 */
  const bool expect_success = i < nr_counters || (or_mask & BIT(i));

  /*
 * KVM drops writes to MSR_P6_PERFCTR[0|1] if the counters are
 * unsupported, i.e. doesn't #GP and reads back '0'.
 */
  const uint64_t expected_val = expect_success ? test_val : 0;
  const bool expect_gp = !expect_success && msr != MSR_P6_PERFCTR0 &&
           msr != MSR_P6_PERFCTR1;
  uint32_t rdpmc_idx;
  uint8_t vector;
  uint64_t val;

  vector = wrmsr_safe(msr, test_val);
  GUEST_ASSERT_PMC_MSR_ACCESS(WRMSR, msr, expect_gp, vector);

  vector = rdmsr_safe(msr, &val);
  GUEST_ASSERT_PMC_MSR_ACCESS(RDMSR, msr, expect_gp, vector);

  /* On #GP, the result of RDMSR is undefined. */
  if (!expect_gp)
   GUEST_ASSERT_PMC_VALUE(RDMSR, msr, val, expected_val);

  /*
 * Redo the read tests with RDPMC, which has different indexing
 * semantics and additional capabilities.
 */
  rdpmc_idx = i;
  if (base_msr == MSR_CORE_PERF_FIXED_CTR0)
   rdpmc_idx |= INTEL_RDPMC_FIXED;

  guest_test_rdpmc(rdpmc_idx, expect_success, expected_val);

  /*
 * KVM doesn't support non-architectural PMUs, i.e. it should
 * impossible to have fast mode RDPMC.  Verify that attempting
 * to use fast RDPMC always #GPs.
 */
  GUEST_ASSERT(!expect_success || !pmu_has_fast_mode);
  rdpmc_idx |= INTEL_RDPMC_FAST;
  guest_test_rdpmc(rdpmc_idx, false, -1ull);

  vector = wrmsr_safe(msr, 0);
  GUEST_ASSERT_PMC_MSR_ACCESS(WRMSR, msr, expect_gp, vector);
 }
}

static void guest_test_gp_counters(void)
{
 uint8_t pmu_version = guest_get_pmu_version();
 uint8_t nr_gp_counters = 0;
 uint32_t base_msr;

 if (pmu_version)
  nr_gp_counters = this_cpu_property(X86_PROPERTY_PMU_NR_GP_COUNTERS);

 /*
 * For v2+ PMUs, PERF_GLOBAL_CTRL's architectural post-RESET value is
 * "Sets bits n-1:0 and clears the upper bits", where 'n' is the number
 * of GP counters.  If there are no GP counters, require KVM to leave
 * PERF_GLOBAL_CTRL '0'.  This edge case isn't covered by the SDM, but
 * follow the spirit of the architecture and only globally enable GP
 * counters, of which there are none.
 */
 if (pmu_version > 1) {
  uint64_t global_ctrl = rdmsr(MSR_CORE_PERF_GLOBAL_CTRL);

  if (nr_gp_counters)
   GUEST_ASSERT_EQ(global_ctrl, GENMASK_ULL(nr_gp_counters - 1, 0));
  else
   GUEST_ASSERT_EQ(global_ctrl, 0);
 }

 if (this_cpu_has(X86_FEATURE_PDCM) &&
     rdmsr(MSR_IA32_PERF_CAPABILITIES) & PMU_CAP_FW_WRITES)
  base_msr = MSR_IA32_PMC0;
 else
  base_msr = MSR_IA32_PERFCTR0;

 guest_rd_wr_counters(base_msr, MAX_NR_GP_COUNTERS, nr_gp_counters, 0);
 GUEST_DONE();
}

static void test_gp_counters(uint8_t pmu_version, uint64_t perf_capabilities,
        uint8_t nr_gp_counters)
{
 struct kvm_vcpu *vcpu;
 struct kvm_vm *vm;

 vm = pmu_vm_create_with_one_vcpu(&vcpu, guest_test_gp_counters,
      pmu_version, perf_capabilities);

 vcpu_set_cpuid_property(vcpu, X86_PROPERTY_PMU_NR_GP_COUNTERS,
    nr_gp_counters);

 run_vcpu(vcpu);

 kvm_vm_free(vm);
}

static void guest_test_fixed_counters(void)
{
 uint64_t supported_bitmask = 0;
 uint8_t nr_fixed_counters = 0;
 uint8_t i;

 /* Fixed counters require Architectural vPMU Version 2+. */
 if (guest_get_pmu_version() >= 2)
  nr_fixed_counters = this_cpu_property(X86_PROPERTY_PMU_NR_FIXED_COUNTERS);

 /*
 * The supported bitmask for fixed counters was introduced in PMU
 * version 5.
 */
 if (guest_get_pmu_version() >= 5)
  supported_bitmask = this_cpu_property(X86_PROPERTY_PMU_FIXED_COUNTERS_BITMASK);

 guest_rd_wr_counters(MSR_CORE_PERF_FIXED_CTR0, MAX_NR_FIXED_COUNTERS,
        nr_fixed_counters, supported_bitmask);

 for (i = 0; i < MAX_NR_FIXED_COUNTERS; i++) {
  uint8_t vector;
  uint64_t val;

  if (i >= nr_fixed_counters && !(supported_bitmask & BIT_ULL(i))) {
   vector = wrmsr_safe(MSR_CORE_PERF_FIXED_CTR_CTRL,
         FIXED_PMC_CTRL(i, FIXED_PMC_KERNEL));
   __GUEST_ASSERT(vector == GP_VECTOR,
           "Expected #GP for counter %u in FIXED_CTR_CTRL", i);

   vector = wrmsr_safe(MSR_CORE_PERF_GLOBAL_CTRL,
         FIXED_PMC_GLOBAL_CTRL_ENABLE(i));
   __GUEST_ASSERT(vector == GP_VECTOR,
           "Expected #GP for counter %u in PERF_GLOBAL_CTRL", i);
   continue;
  }

  wrmsr(MSR_CORE_PERF_FIXED_CTR0 + i, 0);
  wrmsr(MSR_CORE_PERF_FIXED_CTR_CTRL, FIXED_PMC_CTRL(i, FIXED_PMC_KERNEL));
  wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, FIXED_PMC_GLOBAL_CTRL_ENABLE(i));
  __asm__ __volatile__("loop ." : "+c"((int){NUM_LOOPS}));
  wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, 0);
  val = rdmsr(MSR_CORE_PERF_FIXED_CTR0 + i);

  GUEST_ASSERT_NE(val, 0);
 }
 GUEST_DONE();
}

static void test_fixed_counters(uint8_t pmu_version, uint64_t perf_capabilities,
    uint8_t nr_fixed_counters,
    uint32_t supported_bitmask)
{
 struct kvm_vcpu *vcpu;
 struct kvm_vm *vm;

 vm = pmu_vm_create_with_one_vcpu(&vcpu, guest_test_fixed_counters,
      pmu_version, perf_capabilities);

 vcpu_set_cpuid_property(vcpu, X86_PROPERTY_PMU_FIXED_COUNTERS_BITMASK,
    supported_bitmask);
 vcpu_set_cpuid_property(vcpu, X86_PROPERTY_PMU_NR_FIXED_COUNTERS,
    nr_fixed_counters);

 run_vcpu(vcpu);

 kvm_vm_free(vm);
}

static void test_intel_counters(void)
{
 uint8_t nr_fixed_counters = kvm_cpu_property(X86_PROPERTY_PMU_NR_FIXED_COUNTERS);
 uint8_t nr_gp_counters = kvm_cpu_property(X86_PROPERTY_PMU_NR_GP_COUNTERS);
 uint8_t pmu_version = kvm_cpu_property(X86_PROPERTY_PMU_VERSION);
 unsigned int i;
 uint8_t v, j;
 uint32_t k;

 const uint64_t perf_caps[] = {
  0,
  PMU_CAP_FW_WRITES,
 };

 /*
 * Test up to PMU v5, which is the current maximum version defined by
 * Intel, i.e. is the last version that is guaranteed to be backwards
 * compatible with KVM's existing behavior.
 */
 uint8_t max_pmu_version = max_t(typeof(pmu_version), pmu_version, 5);

 /*
 * Detect the existence of events that aren't supported by selftests.
 * This will (obviously) fail any time hardware adds support for a new
 * event, but it's worth paying that price to keep the test fresh.
 */
 TEST_ASSERT(this_cpu_property(X86_PROPERTY_PMU_EBX_BIT_VECTOR_LENGTH) <= NR_INTEL_ARCH_EVENTS,
      "New architectural event(s) detected; please update this test (length = %u, mask = %x)",
      this_cpu_property(X86_PROPERTY_PMU_EBX_BIT_VECTOR_LENGTH),
      this_cpu_property(X86_PROPERTY_PMU_EVENTS_MASK));

 /*
 * Iterate over known arch events irrespective of KVM/hardware support
 * to verify that KVM doesn't reject programming of events just because
 * the *architectural* encoding is unsupported.  Track which events are
 * supported in hardware; the guest side will validate supported events
 * count correctly, even if *enumeration* of the event is unsupported
 * by KVM and/or isn't exposed to the guest.
 */
 for (i = 0; i < NR_INTEL_ARCH_EVENTS; i++) {
  if (this_pmu_has(intel_event_to_feature(i).gp_event))
   hardware_pmu_arch_events |= BIT(i);
 }

 for (v = 0; v <= max_pmu_version; v++) {
  for (i = 0; i < ARRAY_SIZE(perf_caps); i++) {
   if (!kvm_has_perf_caps && perf_caps[i])
    continue;

   pr_info("Testing arch events, PMU version %u, perf_caps = %lx\n",
    v, perf_caps[i]);
   /*
 * To keep the total runtime reasonable, test every
 * possible non-zero, non-reserved bitmap combination
 * only with the native PMU version and the full bit
 * vector length.
 */
   if (v == pmu_version) {
    for (k = 1; k < (BIT(NR_INTEL_ARCH_EVENTS) - 1); k++)
     test_arch_events(v, perf_caps[i], NR_INTEL_ARCH_EVENTS, k);
   }
   /*
 * Test single bits for all PMU version and lengths up
 * the number of events +1 (to verify KVM doesn't do
 * weird things if the guest length is greater than the
 * host length).  Explicitly test a mask of '0' and all
 * ones i.e. all events being available and unavailable.
 */
   for (j = 0; j <= NR_INTEL_ARCH_EVENTS + 1; j++) {
    test_arch_events(v, perf_caps[i], j, 0);
    test_arch_events(v, perf_caps[i], j, 0xff);

    for (k = 0; k < NR_INTEL_ARCH_EVENTS; k++)
     test_arch_events(v, perf_caps[i], j, BIT(k));
   }

   pr_info("Testing GP counters, PMU version %u, perf_caps = %lx\n",
    v, perf_caps[i]);
   for (j = 0; j <= nr_gp_counters; j++)
    test_gp_counters(v, perf_caps[i], j);

   pr_info("Testing fixed counters, PMU version %u, perf_caps = %lx\n",
    v, perf_caps[i]);
   for (j = 0; j <= nr_fixed_counters; j++) {
    for (k = 0; k <= (BIT(nr_fixed_counters) - 1); k++)
     test_fixed_counters(v, perf_caps[i], j, k);
   }
  }
 }
}

int main(int argc, char *argv[])
{
 TEST_REQUIRE(kvm_is_pmu_enabled());

 TEST_REQUIRE(host_cpu_is_intel);
 TEST_REQUIRE(kvm_cpu_has_p(X86_PROPERTY_PMU_VERSION));
 TEST_REQUIRE(kvm_cpu_property(X86_PROPERTY_PMU_VERSION) > 0);

 kvm_pmu_version = kvm_cpu_property(X86_PROPERTY_PMU_VERSION);
 kvm_has_perf_caps = kvm_cpu_has(X86_FEATURE_PDCM);

 test_intel_counters();

 return 0;
}