xref: /xnu-12377.1.9/tests/cpu_counters/kpc_tests.c (revision f6217f891ac0bb64f3d375211650a4c1ff8ca1ea)

// Copyright (c) 2018-2023 Apple Inc.  All rights reserved.

#include <darwintest.h>
#include <ktrace/config.h>
#include <ktrace/session.h>
#include <inttypes.h>
#include <libproc.h>
#include <pthread.h>
#include <stdint.h>
#include <sys/resource.h>
#include <sys/sysctl.h>

#include <kperf/kpc.h>
#include <kperf/kperf.h>
#include <kperfdata/kpep.h>

#include "ktrace_helpers.h"
#include "kperf_helpers.h"
#include "test_utils.h"

T_GLOBAL_META(
	T_META_NAMESPACE("xnu.cpu_counters"),
	T_META_RADAR_COMPONENT_NAME("xnu"),
	T_META_RADAR_COMPONENT_VERSION("cpu counters"),
	T_META_OWNER("mwidmann"),
	T_META_ASROOT(true),
	T_META_CHECK_LEAKS(false));

struct machine {
	unsigned int ncpus;
	unsigned int nfixed;
	unsigned int nconfig;
	uint64_t selector;
};

#ifndef ABSV64
#define ABSV64(n) ((((int64_t)(n)) < 0) ? -((int64_t)(n)) : ((int64_t)(n)))
#endif

static void
skip_if_unsupported(void)
{
	int r;
	int supported = 0;
	size_t supported_size = sizeof(supported);

	r = sysctlbyname("kern.monotonic.supported", &supported, &supported_size,
	    NULL, 0);
	if (r < 0) {
		T_WITH_ERRNO;
		T_SKIP("could not find \"kern.monotonic.supported\" sysctl");
	}

	if (!supported) {
		T_SKIP("PMCs are not supported on this platform");
	}
}

static struct rusage_info_v4 pre_ru = {};

static void
start_kpc(void)
{
	T_SETUPBEGIN;

	kpc_classmask_t classes = KPC_CLASS_FIXED_MASK |
	    KPC_CLASS_CONFIGURABLE_MASK;
	int ret = kpc_set_counting(classes);
	T_ASSERT_POSIX_SUCCESS(ret, "started counting");

	ret = proc_pid_rusage(getpid(), RUSAGE_INFO_V4, (rusage_info_t *)&pre_ru);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "got rusage information");

	kpc_classmask_t classes_on = kpc_get_counting();
	T_QUIET;
	T_ASSERT_EQ(classes, classes_on, "classes counting is correct");

	T_SETUPEND;
}

static void kpc_reset_atend(void);

static void
_assert_kpep_ok(int kpep_err, const char *fmt, ...)
{
	char msg[1024] = "";
	va_list args;
	va_start(args, fmt);
	vsnprintf(msg, sizeof(msg), fmt, args);
	va_end(args);
	T_QUIET;
	T_ASSERT_EQ(kpep_err, KPEP_ERR_NONE, "%s: %s", msg, kpep_strerror(kpep_err));
}

static void
prepare_kpc(struct machine *mch, unsigned int n, const char *event_name,
    uint64_t period)
{
	T_SETUPBEGIN;

	T_ATEND(kpc_reset_atend);

	kpep_db_t db = NULL;
	int ret = kpep_db_create(NULL, &db);
	_assert_kpep_ok(ret, "get kpep database");
	kpep_config_t config = NULL;
	ret = kpep_config_create(db, &config);
	_assert_kpep_ok(ret, "creating event configuration");
	ret = kpep_config_force_counters(config);
	_assert_kpep_ok(ret, "forcing counters with configuration");
	kpep_event_t event = NULL;
	ret = kpep_db_event(db, event_name, &event);
	_assert_kpep_ok(ret, "finding event named %s", event_name);

	size_t ncpus_sz = sizeof(mch->ncpus);
	ret = sysctlbyname("hw.logicalcpu_max", &mch->ncpus, &ncpus_sz,
	    NULL, 0);
	T_QUIET;
	T_ASSERT_POSIX_SUCCESS(ret, "sysctlbyname(hw.logicalcpu_max)");
	T_QUIET;
	T_ASSERT_GT(mch->ncpus, 0, "must have some number of CPUs");

	ret = kpc_force_all_ctrs_set(1);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_force_all_ctrs_set(1)");

	int forcing = 0;
	ret = kpc_force_all_ctrs_get(&forcing);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_force_all_ctrs_get");
	T_QUIET; T_ASSERT_EQ(forcing, 1, "counters must be forced");

	mch->nfixed = kpc_get_counter_count(KPC_CLASS_FIXED_MASK);
	mch->nconfig = kpc_get_counter_count(KPC_CLASS_CONFIGURABLE_MASK);

	T_LOG("machine: ncpus = %d, nfixed = %d, nconfig = %d", mch->ncpus,
	    mch->nfixed, mch->nconfig);

	uint32_t nconfigs = kpc_get_config_count(KPC_CLASS_CONFIGURABLE_MASK);
	for (uint32_t i = 0; i < nconfigs; i++) {
		if (period != 0 && (n == 0 || i == 0)) {
			ret = kpep_config_add_event_trigger(config, &event, 0,
			    period + i * 1000, NULL);
		} else {
			ret = kpep_config_add_event(config, &event, 0, NULL);
		}
		if (ret == KPEP_ERR_CONFIG_CONFLICT) {
			T_LOG("configured %d counters with %s", i, event_name);
			break;
		}
		_assert_kpep_ok(ret, "adding %d event %s to configuration", i,
		    event_name);
	}

	uint64_t *configs = calloc(nconfigs, sizeof(*configs));
	T_QUIET; T_ASSERT_NOTNULL(configs, "allocated config words");
	ret = kpep_config_kpc(config, configs, nconfigs * sizeof(*configs));
	_assert_kpep_ok(ret, "get kpc configuration");
	for (uint32_t i = 0; i < nconfigs; i++) {
		if (configs[i] != 0) {
			mch->selector = configs[i];
			break;
		}
	}
	T_QUIET; T_ASSERT_NE(mch->selector, 0ULL, "found event selector to check");
	ret = kpc_set_config(KPC_CLASS_CONFIGURABLE_MASK, configs);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_config");

	ret = kpep_config_kpc_periods(config, configs, nconfigs * sizeof(*configs));
	_assert_kpep_ok(ret, "get kpc periods");
	ret = kpc_set_period(KPC_CLASS_CONFIGURABLE_MASK, configs);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_period");

	free(configs);

	T_SETUPEND;
}

static void
kpc_reset_atend(void)
{
	uint32_t nconfigs = kpc_get_config_count(KPC_CLASS_CONFIGURABLE_MASK);
	uint64_t *configs = calloc(nconfigs, sizeof(*configs));
	T_QUIET; T_ASSERT_NOTNULL(configs, "allocated config words");

	int ret = kpc_set_period(KPC_CLASS_CONFIGURABLE_MASK, configs);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_period");
	ret = kpc_set_config(KPC_CLASS_CONFIGURABLE_MASK, configs);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_config");

	free(configs);
}

static void *
spin(void *arg)
{
	while (*(volatile int *)arg == 0) {
		;
	}

	return NULL;
}

static pthread_t *
start_threads(const struct machine *mch, void *(*func)(void *), void *arg)
{
	T_SETUPBEGIN;

	pthread_t *threads = calloc((unsigned int)mch->ncpus,
	    sizeof(*threads));
	T_QUIET; T_ASSERT_NOTNULL(threads, "allocated array of threads");
	for (unsigned int i = 0; i < mch->ncpus; i++) {
		int error = pthread_create(&threads[i], NULL, func, arg);
		T_QUIET; T_ASSERT_POSIX_ZERO(error, "pthread_create");
	}

	T_SETUPEND;

	return threads;
}

static void
end_threads(const struct machine *mch, pthread_t *threads)
{
	for (unsigned int i = 0; i < mch->ncpus; i++) {
		int error = pthread_join(threads[i], NULL);
		T_QUIET; T_ASSERT_POSIX_ZERO(error, "joined thread %d", i);
	}
	free(threads);
}

struct tally {
	uint64_t firstvalue;
	uint64_t lastvalue;
	uint64_t nchecks;
	uint64_t nzero;
	uint64_t nstuck;
	uint64_t ndecrease;
};

static void
check_counters(unsigned int ncpus, unsigned int nctrs, struct tally *tallies,
		uint64_t *counts)
{
	for (unsigned int i = 0; i < ncpus; i++) {
		for (unsigned int j = 0; j < nctrs; j++) {
			unsigned int ctr = i * nctrs + j;
			struct tally *tly = &tallies[ctr];
			uint64_t count = counts[ctr];

			if (counts[ctr] == 0) {
				tly->nzero++;
			}
			if (tly->lastvalue == count) {
				tly->nstuck++;
			}
			if (tly->lastvalue > count) {
				tly->ndecrease++;
			}
			tly->lastvalue = count;
			if (tly->nchecks == 0) {
				tly->firstvalue = count;
			}
			tly->nchecks++;
		}
	}
}

static void
check_tally(unsigned int ncpus, unsigned int nctrs, struct tally *tallies)
{
	uint64_t nstuck = 0;
	uint64_t nchecks = 0;
	uint64_t nzero = 0;
	uint64_t ndecrease = 0;

	for (unsigned int i = 0; i < ncpus; i++) {
		for (unsigned int j = 0; j < nctrs; j++) {
			unsigned int ctr = i * nctrs + j;
			struct tally *tly = &tallies[ctr];

			T_LOG("CPU %2u PMC %u: nchecks = %llu, last value = %llx, "
				"delta = %llu, nstuck = %llu", i, j,
			    tly->nchecks, tly->lastvalue, tly->lastvalue - tly->firstvalue,
			    tly->nstuck);

			nchecks += tly->nchecks;
			nstuck += tly->nstuck;
			nzero += tly->nzero;
			ndecrease += tly->ndecrease;
		}
	}

	T_EXPECT_GT(nchecks, 0ULL, "checked 0x%" PRIx64 " counter values", nchecks);
	T_EXPECT_EQ(nzero, 0ULL, "found 0x%" PRIx64 " zero values", nzero);
	T_EXPECT_EQ(nstuck, 0ULL, "found 0x%" PRIx64 " stuck values", nstuck);
	T_EXPECT_EQ(ndecrease, 0ULL,
	    "found 0x%" PRIx64 " decreasing values", ndecrease);
}

#define TESTDUR_NS (5 * NSEC_PER_SEC)

T_DECL(kpc_cpu_direct_configurable,
    "test that configurable counters return monotonically increasing values",
    XNU_T_META_SOC_SPECIFIC,
    T_META_BOOTARGS_SET("enable_skstb=1"),
    T_META_TAG_VM_NOT_ELIGIBLE,
    T_META_ENABLED(false) /* rdar://134505531 */)
{
	skip_if_unsupported();

	struct machine mch = {};
	prepare_kpc(&mch, 0, "CORE_ACTIVE_CYCLE", 0);

	int until = 0;
	pthread_t *threads = start_threads(&mch, spin, &until);
	start_kpc();

	T_SETUPBEGIN;

	uint64_t startns = clock_gettime_nsec_np(CLOCK_MONOTONIC);
	uint64_t *counts = kpc_counterbuf_alloc();
	T_QUIET; T_ASSERT_NOTNULL(counts, "allocated space for counter values");
	memset(counts, 0, sizeof(*counts) * mch.ncpus * (mch.nfixed + mch.nconfig));
	struct tally *tly = calloc(mch.ncpus * mch.nconfig, sizeof(*tly));
	T_QUIET; T_ASSERT_NOTNULL(tly, "allocated space for tallies");

	T_SETUPEND;

	int n = 0;
	while (clock_gettime_nsec_np(CLOCK_MONOTONIC) - startns < TESTDUR_NS) {
		int ret = kpc_get_cpu_counters(true,
		    KPC_CLASS_CONFIGURABLE_MASK, NULL, counts);
		T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_get_cpu_counters");

		check_counters(mch.ncpus, mch.nconfig, tly, counts);

		usleep(10000);
		n++;
		if (n % 100 == 0) {
			T_LOG("checked 100 times");
		}
	}

	check_tally(mch.ncpus, mch.nconfig, tly);

	until = 1;
	end_threads(&mch, threads);
}

T_DECL(kpc_thread_direct_instrs_cycles,
    "test that fixed thread counters return monotonically increasing values",
    XNU_T_META_SOC_SPECIFIC, T_META_TAG_VM_NOT_ELIGIBLE)
{
	int err;
	uint32_t ctrs_cnt;
	uint64_t *ctrs_a;
	uint64_t *ctrs_b;

	skip_if_unsupported();

	T_SETUPBEGIN;

	ctrs_cnt = kpc_get_counter_count(KPC_CLASS_FIXED_MASK);
	if (ctrs_cnt == 0) {
		T_SKIP("no fixed counters available");
	}
	T_LOG("device has %" PRIu32 " fixed counters", ctrs_cnt);

	T_QUIET; T_ASSERT_POSIX_SUCCESS(kpc_force_all_ctrs_set(1), NULL);
	T_ASSERT_POSIX_SUCCESS(kpc_set_counting(KPC_CLASS_FIXED_MASK),
	    "kpc_set_counting");
	T_ASSERT_POSIX_SUCCESS(kpc_set_thread_counting(KPC_CLASS_FIXED_MASK),
	    "kpc_set_thread_counting");

	T_SETUPEND;

	ctrs_a = malloc(ctrs_cnt * sizeof(uint64_t));
	T_QUIET; T_ASSERT_NOTNULL(ctrs_a, NULL);

	err = kpc_get_thread_counters(0, ctrs_cnt, ctrs_a);
	T_ASSERT_POSIX_SUCCESS(err, "kpc_get_thread_counters");

	for (uint32_t i = 0; i < ctrs_cnt; i++) {
		T_LOG("checking counter %d with value %" PRIu64 " > 0", i, ctrs_a[i]);
		T_QUIET;
		T_EXPECT_GT(ctrs_a[i], UINT64_C(0), "counter %d is non-zero", i);
	}

	ctrs_b = malloc(ctrs_cnt * sizeof(uint64_t));
	T_QUIET; T_ASSERT_NOTNULL(ctrs_b, NULL);

	err = kpc_get_thread_counters(0, ctrs_cnt, ctrs_b);
	T_ASSERT_POSIX_SUCCESS(err, "kpc_get_thread_counters");

	for (uint32_t i = 0; i < ctrs_cnt; i++) {
		T_LOG("checking counter %d with value %" PRIu64
		    " > previous value %" PRIu64, i, ctrs_b[i], ctrs_a[i]);
		T_QUIET;
		T_EXPECT_GT(ctrs_b[i], UINT64_C(0), "counter %d is non-zero", i);
		T_QUIET; T_EXPECT_LT(ctrs_a[i], ctrs_b[i],
		    "counter %d is increasing", i);
	}

	free(ctrs_a);
	free(ctrs_b);
}

#define PMI_TEST_DURATION_NS (15 * NSEC_PER_SEC)
#define PERIODIC_CPU_COUNT_MS (250)
#define NTIMESLICES (72)
#define PMI_PERIOD (50ULL * 1000 * 1000)
#define END_EVENT KDBG_EVENTID(0xfe, 0xfe, 0)

struct cpu {
	uint64_t prev_count, max_skid;
	unsigned int scheduled_outside_slice;
	unsigned int pmi_timeslices[NTIMESLICES];
	unsigned int scheduled_timeslices[NTIMESLICES];
};

T_DECL(kpc_pmi_configurable,
    "test that PMIs don't interfere with sampling counters in kperf",
    XNU_T_META_SOC_SPECIFIC,
    T_META_BOOTARGS_SET("enable_skstb=1"),
    T_META_TAG_VM_NOT_ELIGIBLE,
    T_META_ENABLED(false) /* rdar://134505531 */)
{
	skip_if_unsupported();

	start_controlling_ktrace();
	struct machine mch = {};
	prepare_kpc(&mch, 1, "CORE_ACTIVE_CYCLE", PMI_PERIOD);

	T_SETUPBEGIN;

	int32_t *actions = calloc(mch.nconfig, sizeof(*actions));
	actions[0] = 1;
	int ret = kpc_set_actionid(KPC_CLASS_CONFIGURABLE_MASK, actions);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_actionid");
	free(actions);

	(void)kperf_action_count_set(1);
	ret = kperf_action_samplers_set(1,
	    KPERF_SAMPLER_TINFO | KPERF_SAMPLER_KSTACK);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kperf_action_samplers_set");

	ktrace_config_t ktconfig = ktrace_config_create_current();
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(ktconfig, "create current config");
	ret = ktrace_config_print_description(ktconfig, stdout);
	T_QUIET; T_ASSERT_POSIX_ZERO(ret, "print config description");

	struct cpu *cpus = calloc(mch.ncpus, sizeof(*cpus));
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(cpus, "allocate CPUs array");

	__block unsigned int sample_count = 0;
	__block unsigned int pmi_count = 0;
	__block unsigned int callstack_count = 0;
	__block uint64_t first_ns = 0;
	__block uint64_t last_ns = 0;

	ktrace_session_t sess = ktrace_session_create();
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(sess, "ktrace_session_create");

	ktrace_events_single(sess, PERF_KPC_PMI, ^(struct trace_point *tp) {
		if (tp->debugid & DBG_FUNC_END) {
			return;
		}

		uint64_t cur_ns = 0;
		int cret = ktrace_convert_timestamp_to_nanoseconds(sess,
		    tp->timestamp, &cur_ns);
		T_QUIET; T_ASSERT_POSIX_ZERO(cret, "convert timestamp");

		uint64_t desc = tp->arg1;
		uint64_t config = desc & UINT32_MAX;
		T_QUIET; T_EXPECT_EQ(config & UINT16_MAX,
				mch.selector & UINT16_MAX,
				"PMI argument matches configuration");
		__unused uint64_t counter = (desc >> 32) & UINT16_MAX;
		__unused uint64_t flags = desc >> 48;

		uint64_t count = tp->arg2;
		if (first_ns == 0) {
			first_ns = cur_ns;
		}
		struct cpu *cpu = &cpus[tp->cpuid];

		if (cpu->prev_count != 0) {
			uint64_t delta = count - cpu->prev_count;
			uint64_t skid = delta - PMI_PERIOD;
			if (skid > cpu->max_skid) {
				cpu->max_skid = skid;
			}
		}
		cpu->prev_count = count;

		__unused uint64_t pc = tp->arg3;

		double slice = (double)(cur_ns - first_ns) / PMI_TEST_DURATION_NS *
		    NTIMESLICES;
		if (slice < NTIMESLICES) {
			cpu->pmi_timeslices[(unsigned int)slice] += 1;
		}

		pmi_count++;
	});

	void (^sched_handler)(struct trace_point *tp) =
	    ^(struct trace_point *tp) {
		uint64_t cur_ns = 0;
		int cret = ktrace_convert_timestamp_to_nanoseconds(sess,
		    tp->timestamp, &cur_ns);
		T_QUIET; T_ASSERT_POSIX_ZERO(cret, "convert timestamp");
		if (first_ns == 0) {
			first_ns = cur_ns;
		}

		struct cpu *cpu = &cpus[tp->cpuid];
		double slice = (double)(cur_ns - first_ns) / PMI_TEST_DURATION_NS *
		    NTIMESLICES;
		if (slice < NTIMESLICES) {
			cpu->scheduled_timeslices[(unsigned int)slice] += 1;
		} else {
			cpu->scheduled_outside_slice += 1;
		}
	};
	ktrace_events_single(sess, MACH_SCHED, sched_handler);
	ktrace_events_single(sess, MACH_STACK_HANDOFF, sched_handler);

	ktrace_events_single(sess, PERF_SAMPLE, ^(struct trace_point * tp) {
		if (tp->debugid & DBG_FUNC_START) {
			sample_count++;
		}
	});
	ktrace_events_single(sess, PERF_STK_KHDR,
	    ^(struct trace_point * __unused tp) {
		callstack_count++;
	});

	ktrace_events_single(sess, END_EVENT, ^(struct trace_point *tp) {
		int cret = ktrace_convert_timestamp_to_nanoseconds(sess,
		    tp->timestamp, &last_ns);
		T_QUIET; T_ASSERT_POSIX_ZERO(cret, "convert timestamp");

		ktrace_end(sess, 1);
	});

	uint64_t *counts = kpc_counterbuf_alloc();
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(counts,
			"allocated counter values array");
	memset(counts, 0, sizeof(*counts) * mch.ncpus * (mch.nfixed + mch.nconfig));
	struct tally *tly = calloc(mch.ncpus * (mch.nconfig + mch.nfixed),
			sizeof(*tly));
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(tly, "allocated tallies array");

	dispatch_source_t cpu_count_timer = dispatch_source_create(
			DISPATCH_SOURCE_TYPE_TIMER, 0, 0, dispatch_get_main_queue());
    dispatch_source_set_timer(cpu_count_timer, dispatch_time(DISPATCH_TIME_NOW,
        PERIODIC_CPU_COUNT_MS * NSEC_PER_MSEC),
        PERIODIC_CPU_COUNT_MS * NSEC_PER_MSEC, 0);
    dispatch_source_set_cancel_handler(cpu_count_timer, ^{
        dispatch_release(cpu_count_timer);
    });

    __block uint64_t first_check_ns = 0;
    __block uint64_t last_check_ns = 0;

    dispatch_source_set_event_handler(cpu_count_timer, ^{
		int cret = kpc_get_cpu_counters(true,
		    KPC_CLASS_FIXED_MASK | KPC_CLASS_CONFIGURABLE_MASK, NULL, counts);
		T_QUIET; T_ASSERT_POSIX_SUCCESS(cret, "kpc_get_cpu_counters");

		if (!first_check_ns) {
			first_check_ns = clock_gettime_nsec_np(CLOCK_MONOTONIC);
		} else {
			last_check_ns = clock_gettime_nsec_np(CLOCK_MONOTONIC);
		}
		check_counters(mch.ncpus, mch.nfixed + mch.nconfig, tly, counts);
	});
	ktrace_events_class(sess, DBG_PERF, ^(struct trace_point * __unused tp) {});

	int stop = 0;
	(void)start_threads(&mch, spin, &stop);

	ktrace_set_completion_handler(sess, ^{
		dispatch_cancel(cpu_count_timer);

		check_tally(mch.ncpus, mch.nfixed + mch.nconfig, tly);

		struct rusage_info_v4 post_ru = {};
		int ruret = proc_pid_rusage(getpid(), RUSAGE_INFO_V4,
				(rusage_info_t *)&post_ru);
		T_QUIET; T_ASSERT_POSIX_SUCCESS(ruret, "got rusage information");
		T_LOG("saw %llu cycles in process",
				post_ru.ri_cycles - pre_ru.ri_cycles);
		uint64_t total_cycles = 0;

		T_LOG("saw pmis = %u, samples = %u, stacks = %u", pmi_count, sample_count,
		    callstack_count);
		// Allow some slop in case the trace is cut-off midway through a
		// sample.
		const unsigned int cutoff_leeway = 32;
		T_EXPECT_GE(sample_count + cutoff_leeway, pmi_count,
		    "saw as many samples as PMIs");
		T_EXPECT_GE(callstack_count + cutoff_leeway, pmi_count,
		    "saw as many stacks as PMIs");

		unsigned int cpu_sample_count = 0;
		char sample_slices[NTIMESLICES + 1];
		sample_slices[NTIMESLICES] = '\0';
		for (unsigned int i = 0; i < mch.ncpus; i++) {
			memset(sample_slices, '-', sizeof(sample_slices) - 1);

			struct cpu *cpu = &cpus[i];
			unsigned int pmi_slice_count = 0, no_sched_slice_count = 0,
					cpu_pmi_count = 0, last_contiguous = 0;
			bool seen_empty = false;
			for (unsigned int j = 0; j < NTIMESLICES; j++) {
				unsigned int slice_pmi_count = cpu->pmi_timeslices[j];
				unsigned int slice_sched_count = cpu->scheduled_timeslices[j];
				cpu_pmi_count += slice_pmi_count;
				if (slice_pmi_count > 0) {
					pmi_slice_count++;
					sample_slices[j] = '*';
				} else if (slice_sched_count == 0) {
					no_sched_slice_count++;
					sample_slices[j] = '.';
				} else {
					seen_empty = true;
				}
				if (!seen_empty) {
					last_contiguous = j;
				}
			}
			unsigned int ctr = i * (mch.nfixed + mch.nconfig) + mch.nfixed;
			uint64_t delta = tly[ctr].lastvalue - tly[ctr].firstvalue;
			T_LOG("%g GHz", (double)delta / (last_check_ns - first_check_ns));
			total_cycles += delta;
			uint64_t abs_max_skid = (uint64_t)ABSV64(cpu->max_skid);
			T_LOG("CPU %2u: %4up:%4un/%u, %6u/%llu, max skid = %llu (%.4f%%), "
					"last contiguous = %u, scheduled outside = %u", i,
					pmi_slice_count, no_sched_slice_count, NTIMESLICES,
					sample_count, delta / PMI_PERIOD, abs_max_skid,
					(double)abs_max_skid / PMI_PERIOD * 100, last_contiguous,
					cpu->scheduled_outside_slice);
			T_LOG("%s", sample_slices);
			if (cpu_pmi_count > 0) {
				cpu_sample_count++;
			}
			T_EXPECT_EQ(last_contiguous, NTIMESLICES - 1,
					"CPU %2u: saw samples in each time slice", i);
		}
		T_LOG("kpc reported %llu total cycles", total_cycles);
		T_LOG("saw %u sample events, across %u/%u cpus", sample_count,
				cpu_sample_count, mch.ncpus);
		T_EXPECT_EQ(cpu_sample_count, mch.ncpus,
				"should see PMIs on every CPU");
		T_END;
	});

	int dbglvl = 3;
	ret = sysctlbyname("kperf.debug_level", NULL, NULL, &dbglvl,
	    sizeof(dbglvl));
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "set kperf debug level");
	ret = kperf_sample_set(1);
	T_ASSERT_POSIX_SUCCESS(ret, "kperf_sample_set");

	start_kpc();

	int error = ktrace_start(sess, dispatch_get_main_queue());
	T_ASSERT_POSIX_ZERO(error, "started tracing");

	dispatch_after(dispatch_time(DISPATCH_TIME_NOW, PMI_TEST_DURATION_NS),
			dispatch_get_main_queue(), ^{
		T_LOG("ending tracing after timeout");
		kdebug_trace(END_EVENT, 0, 0, 0, 0);
	});

	dispatch_activate(cpu_count_timer);

	T_SETUPEND;

	dispatch_main();
}

#if defined(__arm64__)
#define IS_ARM64 true
#else // defined(__arm64__)
#define IS_ARM64 false
#endif // !defined(__arm64__)

T_DECL(kpc_pmu_config, "ensure PMU can be configured",
    XNU_T_META_SOC_SPECIFIC,
    T_META_ENABLED(IS_ARM64), T_META_TAG_VM_NOT_ELIGIBLE)
{
	T_SETUPBEGIN;
	int ret = kpc_force_all_ctrs_set(1);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret,
			"force all counters to allow raw PMU configuration");
	uint32_t nconfigs = kpc_get_config_count(KPC_CLASS_RAWPMU_MASK);
	T_LOG("found %u raw PMU configuration words", nconfigs);
	uint64_t *configs = calloc(nconfigs, sizeof(*configs));
	T_QUIET; T_ASSERT_NOTNULL(configs, "allocated config words");
	T_SETUPEND;

	ret = kpc_set_config(KPC_CLASS_RAWPMU_MASK, configs);
	T_ASSERT_POSIX_SUCCESS(ret, "should set PMU configuration");
}

T_DECL(pmi_pc_capture, "ensure PC capture works for PMCs 5, 6, and 7",
    XNU_T_META_SOC_SPECIFIC,
    T_META_REQUIRES_SYSCTL_EQ("kpc.pc_capture_supported", 1), T_META_TAG_VM_NOT_ELIGIBLE)
{
	start_controlling_ktrace();
	struct machine mch = {};
	prepare_kpc(&mch, 0, "INST_BRANCH_TAKEN", PMI_PERIOD);

	T_SETUPBEGIN;

	uint64_t *periods = calloc(mch.nconfig, sizeof(*periods));
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(periods, "allocate periods array");
	for (unsigned int i = 0; i < mch.nconfig; i++) {
		/*
		 * Offset the periods so the PMIs don't alias to the same PC capture.
		 * Since there's only one PC capture register, they will clobber each
		 * other.
		 */
		periods[i] = PMI_PERIOD / 1000 + (i * 1000);
	}

	int ret = kpc_set_period(KPC_CLASS_CONFIGURABLE_MASK, periods);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_period");
	free(periods);

	int32_t *actions = calloc(mch.nconfig, sizeof(*actions));
	for (unsigned int i = 0; i < mch.nconfig; i++) {
		actions[i] = 1;
	}
	ret = kpc_set_actionid(KPC_CLASS_CONFIGURABLE_MASK, actions);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kpc_set_actionid");
	free(actions);

	(void)kperf_action_count_set(1);
	ret = kperf_action_samplers_set(1, KPERF_SAMPLER_TINFO);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "kperf_action_samplers_set");

	ktrace_session_t sess = ktrace_session_create();
	T_QUIET; T_WITH_ERRNO; T_ASSERT_NOTNULL(sess, "ktrace_session_create");

	uint64_t pc_captured_arr[3] = {};
	uint64_t *pc_captured = pc_captured_arr;
	uint64_t pmi_event_arr[3] = {};
	uint64_t *pmi_event = pmi_event_arr;
	ktrace_events_single(sess, PERF_KPC_PMI, ^(struct trace_point *tp) {
		if (tp->debugid & DBG_FUNC_END) {
			return;
		}

		uint64_t desc = tp->arg1;

#define KPC_DESC_COUNTER(DESC) (((DESC) >> 32) & 0xffff)
#define KPC_DESC_CONFIG(DESC) ((DESC) & 0xffff)
#define KPC_DESC_FLAGS(DESC) ((DESC) >> 48)
#define KPC_FLAG_PC_CAPTURED (0x08)

		uint64_t counter = KPC_DESC_COUNTER(desc);
		uint64_t flags = KPC_DESC_FLAGS(desc);
		if (counter >= 5 && counter <= 7) {
			pmi_event[counter - 5]++;
			if (flags & KPC_FLAG_PC_CAPTURED) {
				pc_captured[counter - 5]++;
			}
		}
		T_QUIET;
		T_ASSERT_EQ(KPC_DESC_CONFIG(desc), mch.selector,
		    "correct counter configuration");
	});

	ktrace_events_single(sess, END_EVENT, ^(struct trace_point *tp __unused) {
		ktrace_config_t config = ktrace_config_create_current();
		ktrace_config_print_description(config, stdout);
		ktrace_config_destroy(config);
		T_LOG("saw ending event");
		ktrace_end(sess, 1);
	});

	ktrace_set_completion_handler(sess, ^{
		ktrace_session_destroy(sess);
		for (unsigned int i = 0; i < 3; i++) {
			T_LOG("PMC%u: saw %llu/%llu (%g%%) PMIs with PC capture", i + 5,
			    pc_captured[i], pmi_event[i],
			    (double)pc_captured[i] / (double)pmi_event[i] * 100.0);
			T_EXPECT_GT(pc_captured[i], 0ULL, "saw PC capture for counter %u",
			    i + 5);
		}
		T_END;
	});

	ret = kperf_sample_set(1);
	T_ASSERT_POSIX_SUCCESS(ret, "kperf_sample_set");

	start_kpc();

	int error = ktrace_start(sess, dispatch_get_main_queue());
	T_ASSERT_POSIX_ZERO(error, "started tracing");

	dispatch_after(dispatch_time(DISPATCH_TIME_NOW, PMI_TEST_DURATION_NS),
			dispatch_get_main_queue(), ^{
		T_LOG("ending tracing after timeout");
		kdebug_trace(END_EVENT, 0, 0, 0, 0);
	});

	T_SETUPEND;

	dispatch_main();
}