xnu-4903.241.1.tar.gz

[apple/xnu.git] / tools / tests / zero-to-n / zero-to-n.c

/*
 * Copyright (c) 2009 Apple Inc. All rights reserved.
 *
 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
 * 
 * This file contains Original Code and/or Modifications of Original Code
 * as defined in and that are subject to the Apple Public Source License
 * Version 2.0 (the 'License'). You may not use this file except in
 * compliance with the License. The rights granted to you under the License
 * may not be used to create, or enable the creation or redistribution of,
 * unlawful or unlicensed copies of an Apple operating system, or to
 * circumvent, violate, or enable the circumvention or violation of, any
 * terms of an Apple operating system software license agreement.
 * 
 * Please obtain a copy of the License at
 * http://www.opensource.apple.com/apsl/ and read it before using this file.
 * 
 * The Original Code and all software distributed under the License are
 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
 * Please see the License for the specific language governing rights and
 * limitations under the License.
 * 
 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
 */
#include <unistd.h>
#include <stdio.h>
#include <math.h>
#include <sys/kdebug.h>
#include <stdlib.h>
#include <pthread.h>
#include <errno.h>
#include <err.h>
#include <string.h>
#include <assert.h>
#include <sysexits.h>
#include <sys/sysctl.h>
#include <getopt.h>

#include <spawn.h>
#include <spawn_private.h>
#include <sys/spawn_internal.h>
#include <mach-o/dyld.h>

#include <mach/mach_time.h>
#include <mach/mach.h>
#include <mach/task.h>
#include <mach/semaphore.h>

#include <pthread/qos_private.h>

#include <sys/resource.h>

#include <stdatomic.h>

typedef enum wake_type { WAKE_BROADCAST_ONESEM, WAKE_BROADCAST_PERTHREAD, WAKE_CHAIN, WAKE_HOP } wake_type_t;
typedef enum my_policy_type { MY_POLICY_REALTIME, MY_POLICY_TIMESHARE, MY_POLICY_FIXEDPRI } my_policy_type_t;

#define mach_assert_zero(error)        do { if ((error) != 0) { fprintf(stderr, "[FAIL] error %d (%s) ", (error), mach_error_string(error)); assert(error == 0); } } while (0)
#define mach_assert_zero_t(tid, error) do { if ((error) != 0) { fprintf(stderr, "[FAIL] Thread %d error %d (%s) ", (tid), (error), mach_error_string(error)); assert(error == 0); } } while (0)
#define assert_zero_t(tid, error)      do { if ((error) != 0) { fprintf(stderr, "[FAIL] Thread %d error %d ", (tid), (error)); assert(error == 0); } } while (0)

#define CONSTRAINT_NANOS        (20000000ll)    /* 20 ms */
#define COMPUTATION_NANOS       (10000000ll)    /* 10 ms */
#define TRACEWORTHY_NANOS       (10000000ll)    /* 10 ms */

#define DEBUG 0

#if DEBUG
#define debug_log(args...) printf(args)
#else
#define debug_log(args...) do { } while(0)
#endif

/* Declarations */
static void*                    worker_thread(void *arg);
static void                     usage();
static int                      thread_setup(uint32_t my_id);
static my_policy_type_t         parse_thread_policy(const char *str);
static void                     selfexec_with_apptype(int argc, char *argv[]);
static void                     parse_args(int argc, char *argv[]);

static __attribute__((aligned(128))) _Atomic uint32_t   g_done_threads;
static __attribute__((aligned(128))) _Atomic boolean_t  g_churn_stop = FALSE;
static __attribute__((aligned(128))) _Atomic uint64_t   g_churn_stopped_at = 0;

/* Global variables (general) */
static uint32_t                 g_numcpus;
static uint32_t                 g_numthreads;
static wake_type_t              g_waketype;
static policy_t                 g_policy;
static uint32_t                 g_iterations;
static struct mach_timebase_info g_mti;
static semaphore_t              g_main_sem;
static uint64_t                *g_thread_endtimes_abs;
static boolean_t                g_verbose       = FALSE;
static boolean_t                g_do_affinity   = FALSE;
static uint64_t                 g_starttime_abs;
static uint32_t                 g_iteration_sleeptime_us = 0;
static uint32_t                 g_priority = 0;
static uint32_t                 g_churn_pri = 0;
static uint32_t                 g_churn_count = 0;

static pthread_t*               g_churn_threads = NULL;

/* Threshold for dropping a 'bad run' tracepoint */
static uint64_t                 g_traceworthy_latency_ns = TRACEWORTHY_NANOS;

/* Have we re-execed to set apptype? */
static boolean_t                g_seen_apptype = FALSE;

/* usleep in betweeen iterations */
static boolean_t                g_do_sleep      = TRUE;

/* Every thread spins until all threads have checked in */
static boolean_t                g_do_all_spin = FALSE;

/* Every thread backgrounds temporarily before parking */
static boolean_t                g_drop_priority = FALSE;

/* One randomly chosen thread holds up the train for a certain duration. */
static boolean_t                g_do_one_long_spin = FALSE;
static uint32_t                 g_one_long_spin_id = 0;
static uint64_t                 g_one_long_spin_length_abs = 0;
static uint64_t                 g_one_long_spin_length_ns = 0;

/* Each thread spins for a certain duration after waking up before blocking again. */
static boolean_t                g_do_each_spin = FALSE;
static uint64_t                 g_each_spin_duration_abs = 0;
static uint64_t                 g_each_spin_duration_ns = 0;

/* Global variables (broadcast) */
static semaphore_t              g_broadcastsem;
static semaphore_t              g_leadersem;
static semaphore_t              g_readysem;
static semaphore_t              g_donesem;

/* Global variables (chain) */
static semaphore_t             *g_semarr;

static uint64_t
abs_to_nanos(uint64_t abstime)
{
        return (uint64_t)(abstime * (((double)g_mti.numer) / ((double)g_mti.denom)));
}

static uint64_t
nanos_to_abs(uint64_t ns)
{
        return (uint64_t)(ns * (((double)g_mti.denom) / ((double)g_mti.numer)));
}

inline static void
yield(void)
{
#if defined(__arm__) || defined(__arm64__)
        asm volatile("yield");
#elif defined(__x86_64__) || defined(__i386__)
        asm volatile("pause");
#else
#error Unrecognized architecture
#endif
}

static void *
churn_thread(__unused void *arg)
{
        uint64_t spin_count = 0;

        /*
         * As a safety measure to avoid wedging, we will bail on the spin if
         * it's been more than 1s after the most recent run start
         */

        while (g_churn_stop == FALSE &&
               mach_absolute_time() < (g_starttime_abs + NSEC_PER_SEC)) {
                spin_count++;
                yield();
        }

        /* This is totally racy, but only here to detect if anyone stops early */
        atomic_fetch_add_explicit(&g_churn_stopped_at, spin_count, memory_order_relaxed);

        return NULL;
}

static void
create_churn_threads()
{
        if (g_churn_count == 0)
                g_churn_count = g_numcpus - 1;

        errno_t err;

        struct sched_param param = { .sched_priority = (int)g_churn_pri };
        pthread_attr_t attr;

        /* Array for churn threads */
        g_churn_threads = (pthread_t*) valloc(sizeof(pthread_t) * g_churn_count);
        assert(g_churn_threads);

        if ((err = pthread_attr_init(&attr)))
                errc(EX_OSERR, err, "pthread_attr_init");

        if ((err = pthread_attr_setschedparam(&attr, &param)))
                errc(EX_OSERR, err, "pthread_attr_setschedparam");

        if ((err = pthread_attr_setschedpolicy(&attr, SCHED_RR)))
                errc(EX_OSERR, err, "pthread_attr_setschedpolicy");

        for (uint32_t i = 0 ; i < g_churn_count ; i++) {
                pthread_t new_thread;

                if ((err = pthread_create(&new_thread, &attr, churn_thread, NULL)))
                        errc(EX_OSERR, err, "pthread_create");
                g_churn_threads[i] = new_thread;
        }

        if ((err = pthread_attr_destroy(&attr)))
                errc(EX_OSERR, err, "pthread_attr_destroy");
}

static void
join_churn_threads(void)
{
        if (atomic_load_explicit(&g_churn_stopped_at, memory_order_seq_cst) != 0)
                printf("Warning: Some of the churn threads may have stopped early: %lld\n",
                       g_churn_stopped_at);

        atomic_store_explicit(&g_churn_stop, TRUE, memory_order_seq_cst);

        /* Rejoin churn threads */
        for (uint32_t i = 0; i < g_churn_count; i++) {
                errno_t err = pthread_join(g_churn_threads[i], NULL);
                if (err) errc(EX_OSERR, err, "pthread_join %d", i);
        }
}

/*
 * Figure out what thread policy to use 
 */
static my_policy_type_t
parse_thread_policy(const char *str)
{
        if (strcmp(str, "timeshare") == 0) {
                return MY_POLICY_TIMESHARE;
        } else if (strcmp(str, "realtime") == 0) {
                return MY_POLICY_REALTIME;
        } else if (strcmp(str, "fixed") == 0) {
                return MY_POLICY_FIXEDPRI;
        } else {
                errx(EX_USAGE, "Invalid thread policy \"%s\"", str);
        }
}

/*
 * Figure out what wakeup pattern to use
 */
static wake_type_t
parse_wakeup_pattern(const char *str) 
{
        if (strcmp(str, "chain") == 0) {
                return WAKE_CHAIN;
        } else if (strcmp(str, "hop") == 0) {
                return WAKE_HOP;
        } else if (strcmp(str, "broadcast-single-sem") == 0) {
                return WAKE_BROADCAST_ONESEM;
        } else if (strcmp(str, "broadcast-per-thread") == 0) {
                return WAKE_BROADCAST_PERTHREAD;
        } else {
                errx(EX_USAGE, "Invalid wakeup pattern \"%s\"", str);
        }
}

/*
 * Set policy
 */
static int
thread_setup(uint32_t my_id)
{
        kern_return_t kr;
        errno_t ret;
        thread_time_constraint_policy_data_t pol;

        if (g_priority) {
                int policy = SCHED_OTHER;
                if (g_policy == MY_POLICY_FIXEDPRI)
                        policy = SCHED_RR;

                struct sched_param param = {.sched_priority = (int)g_priority};
                if ((ret = pthread_setschedparam(pthread_self(), policy, &param)))
                        errc(EX_OSERR, ret, "pthread_setschedparam: %d", my_id);
        }

        switch (g_policy) {
                case MY_POLICY_TIMESHARE:
                        break;
                case MY_POLICY_REALTIME:
                        /* Hard-coded realtime parameters (similar to what Digi uses) */
                        pol.period      = 100000;
                        pol.constraint  = (uint32_t) nanos_to_abs(CONSTRAINT_NANOS);
                        pol.computation = (uint32_t) nanos_to_abs(COMPUTATION_NANOS);
                        pol.preemptible = 0; /* Ignored by OS */

                        kr = thread_policy_set(mach_thread_self(), THREAD_TIME_CONSTRAINT_POLICY,
                                               (thread_policy_t) &pol, THREAD_TIME_CONSTRAINT_POLICY_COUNT);
                        mach_assert_zero_t(my_id, kr);
                        break;
                case MY_POLICY_FIXEDPRI:
                        ret = pthread_set_fixedpriority_self();
                        if (ret) errc(EX_OSERR, ret, "pthread_set_fixedpriority_self");
                        break;
                default:
                        errx(EX_USAGE, "invalid policy type %d", g_policy);
        }

        if (g_do_affinity) {
                thread_affinity_policy_data_t affinity;

                affinity.affinity_tag = my_id % 2;

                kr = thread_policy_set(mach_thread_self(), THREAD_AFFINITY_POLICY,
                                       (thread_policy_t)&affinity, THREAD_AFFINITY_POLICY_COUNT);
                mach_assert_zero_t(my_id, kr);
        }

        return 0;
}

/*
 * Wait for a wakeup, potentially wake up another of the "0-N" threads,
 * and notify the main thread when done.
 */
static void*
worker_thread(void *arg)
{
        uint32_t my_id = (uint32_t)(uintptr_t)arg;
        kern_return_t kr;

        volatile double x = 0.0;
        volatile double y = 0.0;

        /* Set policy and so forth */
        thread_setup(my_id);

        for (uint32_t i = 0; i < g_iterations; i++) {
                if (my_id == 0) {
                        /*
                         * Leader thread either wakes everyone up or starts the chain going.
                         */

                        /* Give the worker threads undisturbed time to finish before waiting on them */
                        if (g_do_sleep)
                                usleep(g_iteration_sleeptime_us);

                        debug_log("%d Leader thread wait for ready\n", i);

                        /*
                         * Wait for everyone else to declare ready
                         * Is there a better way to do this that won't interfere with the rest of the chain?
                         * TODO: Invent 'semaphore wait for N signals'
                         */

                        for (uint32_t j = 0 ; j < g_numthreads - 1; j++) {
                                kr = semaphore_wait(g_readysem);
                                mach_assert_zero_t(my_id, kr);
                        }

                        debug_log("%d Leader thread wait\n", i);

                        /* Signal main thread and wait for start of iteration */

                        kr = semaphore_wait_signal(g_leadersem, g_main_sem);
                        mach_assert_zero_t(my_id, kr);

                        g_thread_endtimes_abs[my_id] = mach_absolute_time();

                        debug_log("%d Leader thread go\n", i);

                        assert_zero_t(my_id, atomic_load_explicit(&g_done_threads, memory_order_relaxed));

                        switch (g_waketype) {
                        case WAKE_BROADCAST_ONESEM:
                                kr = semaphore_signal_all(g_broadcastsem);
                                mach_assert_zero_t(my_id, kr);
                                break;
                        case WAKE_BROADCAST_PERTHREAD:
                                for (uint32_t j = 1; j < g_numthreads; j++) {
                                        kr = semaphore_signal(g_semarr[j]);
                                        mach_assert_zero_t(my_id, kr);
                                }
                                break;
                        case WAKE_CHAIN:
                                kr = semaphore_signal(g_semarr[my_id + 1]);
                                mach_assert_zero_t(my_id, kr);
                                break;
                        case WAKE_HOP:
                                kr = semaphore_wait_signal(g_donesem, g_semarr[my_id + 1]);
                                mach_assert_zero_t(my_id, kr);
                                break;
                        }
                } else {
                        /*
                         * Everyone else waits to be woken up,
                         * records when she wakes up, and possibly
                         * wakes up a friend.
                         */
                        switch(g_waketype)  {
                        case WAKE_BROADCAST_ONESEM:
                                kr = semaphore_wait_signal(g_broadcastsem, g_readysem);
                                mach_assert_zero_t(my_id, kr);

                                g_thread_endtimes_abs[my_id] = mach_absolute_time();
                                break;

                        case WAKE_BROADCAST_PERTHREAD:
                                kr = semaphore_wait_signal(g_semarr[my_id], g_readysem);
                                mach_assert_zero_t(my_id, kr);

                                g_thread_endtimes_abs[my_id] = mach_absolute_time();
                                break;

                        case WAKE_CHAIN:
                                kr = semaphore_wait_signal(g_semarr[my_id], g_readysem);
                                mach_assert_zero_t(my_id, kr);

                                /* Signal the next thread *after* recording wake time */

                                g_thread_endtimes_abs[my_id] = mach_absolute_time();

                                if (my_id < (g_numthreads - 1)) {
                                        kr = semaphore_signal(g_semarr[my_id + 1]);
                                        mach_assert_zero_t(my_id, kr);
                                }

                                break;

                        case WAKE_HOP:
                                kr = semaphore_wait_signal(g_semarr[my_id], g_readysem);
                                mach_assert_zero_t(my_id, kr);

                                /* Signal the next thread *after* recording wake time */

                                g_thread_endtimes_abs[my_id] = mach_absolute_time();

                                if (my_id < (g_numthreads - 1)) {
                                        kr = semaphore_wait_signal(g_donesem, g_semarr[my_id + 1]);
                                        mach_assert_zero_t(my_id, kr);
                                } else {
                                        kr = semaphore_signal_all(g_donesem);
                                        mach_assert_zero_t(my_id, kr);
                                }

                                break;
                        }
                }

                debug_log("Thread %p woke up for iteration %d.\n", pthread_self(), i);

                if (g_do_one_long_spin && g_one_long_spin_id == my_id) {
                        /* One randomly chosen thread holds up the train for a while. */

                        uint64_t endspin = g_starttime_abs + g_one_long_spin_length_abs;
                        while (mach_absolute_time() < endspin) {
                                y = y + 1.5 + x;
                                x = sqrt(y);
                        }
                }

                if (g_do_each_spin) {
                        /* Each thread spins for a certain duration after waking up before blocking again. */

                        uint64_t endspin = mach_absolute_time() + g_each_spin_duration_abs;
                        while (mach_absolute_time() < endspin) {
                                y = y + 1.5 + x;
                                x = sqrt(y);
                        }
                }

                uint32_t done_threads;
                done_threads = atomic_fetch_add_explicit(&g_done_threads, 1, memory_order_relaxed) + 1;

                debug_log("Thread %p new value is %d, iteration %d\n", pthread_self(), done_threads, i);

                if (g_drop_priority) {
                        /* Drop priority to BG momentarily */
                        errno_t ret = setpriority(PRIO_DARWIN_THREAD, 0, PRIO_DARWIN_BG);
                        if (ret) errc(EX_OSERR, ret, "setpriority PRIO_DARWIN_BG");
                }

                if (g_do_all_spin) {
                        /* Everyone spins until the last thread checks in. */

                        while (atomic_load_explicit(&g_done_threads, memory_order_relaxed) < g_numthreads) {
                                y = y + 1.5 + x;
                                x = sqrt(y);
                        }
                }

                if (g_drop_priority) {
                        /* Restore normal priority */
                        errno_t ret = setpriority(PRIO_DARWIN_THREAD, 0, 0);
                        if (ret) errc(EX_OSERR, ret, "setpriority 0");
                }

                debug_log("Thread %p done spinning, iteration %d\n", pthread_self(), i);
        }

        if (my_id == 0) {
                /* Give the worker threads undisturbed time to finish before waiting on them */
                if (g_do_sleep)
                        usleep(g_iteration_sleeptime_us);

                /* Wait for the worker threads to finish */
                for (uint32_t i = 0 ; i < g_numthreads - 1; i++) {
                        kr = semaphore_wait(g_readysem);
                        mach_assert_zero_t(my_id, kr);
                }

                /* Tell everyone and the main thread that the last iteration is done */
                debug_log("%d Leader thread done\n", i);

                kr = semaphore_signal_all(g_main_sem);
                mach_assert_zero_t(my_id, kr);
        } else {
                /* Hold up thread teardown so it doesn't affect the last iteration */
                kr = semaphore_wait_signal(g_main_sem, g_readysem);
                mach_assert_zero_t(my_id, kr);
        }

        return 0;
}

/*
 * Given an array of uint64_t values, compute average, max, min, and standard deviation
 */
static void
compute_stats(uint64_t *values, uint64_t count, float *averagep, uint64_t *maxp, uint64_t *minp, float *stddevp)
{
        uint32_t i;
        uint64_t _sum = 0;
        uint64_t _max = 0;
        uint64_t _min = UINT64_MAX;
        float    _avg = 0;
        float    _dev = 0;

        for (i = 0; i < count; i++) {
                _sum += values[i];
                _max = values[i] > _max ? values[i] : _max;
                _min = values[i] < _min ? values[i] : _min;
        }

        _avg = ((float)_sum) / ((float)count);
        
        _dev = 0;
        for (i = 0; i < count; i++) {
                _dev += powf((((float)values[i]) - _avg), 2);
        }
        
        _dev /= count;
        _dev = sqrtf(_dev);

        *averagep = _avg;
        *maxp = _max;
        *minp = _min;
        *stddevp = _dev;
}

int
main(int argc, char **argv)
{
        errno_t ret;
        kern_return_t kr;

        pthread_t       *threads;
        uint64_t        *worst_latencies_ns;
        uint64_t        *worst_latencies_from_first_ns;
        uint64_t        max, min;
        float           avg, stddev;

        for (int i = 0; i < argc; i++)
                if (strcmp(argv[i], "--switched_apptype") == 0)
                        g_seen_apptype = TRUE;

        if (!g_seen_apptype)
                selfexec_with_apptype(argc, argv);

        parse_args(argc, argv);

        srand((unsigned int)time(NULL));

        mach_timebase_info(&g_mti);

        size_t ncpu_size = sizeof(g_numcpus);
        ret = sysctlbyname("hw.ncpu", &g_numcpus, &ncpu_size, NULL, 0);
        if (ret) err(EX_OSERR, "Failed sysctlbyname(hw.ncpu)");

        if (g_do_each_spin)
                g_each_spin_duration_abs = nanos_to_abs(g_each_spin_duration_ns);

        /* Configure the long-spin thread to take up half of its computation */
        if (g_do_one_long_spin) {
                g_one_long_spin_length_ns = COMPUTATION_NANOS / 2;
                g_one_long_spin_length_abs = nanos_to_abs(g_one_long_spin_length_ns);
        }

        /* Estimate the amount of time the cleanup phase needs to back off */
        g_iteration_sleeptime_us = g_numthreads * 20;

        uint32_t threads_per_core = (g_numthreads / g_numcpus) + 1;
        if (g_do_each_spin)
                g_iteration_sleeptime_us += threads_per_core * (g_each_spin_duration_ns / NSEC_PER_USEC);
        if (g_do_one_long_spin)
                g_iteration_sleeptime_us += g_one_long_spin_length_ns / NSEC_PER_USEC;

        /* Arrays for threads and their wakeup times */
        threads = (pthread_t*) valloc(sizeof(pthread_t) * g_numthreads);
        assert(threads);

        size_t endtimes_size = sizeof(uint64_t) * g_numthreads;

        g_thread_endtimes_abs = (uint64_t*) valloc(endtimes_size);
        assert(g_thread_endtimes_abs);

        /* Ensure the allocation is pre-faulted */
        ret = memset_s(g_thread_endtimes_abs, endtimes_size, 0, endtimes_size);
        if (ret) errc(EX_OSERR, ret, "memset_s endtimes");

        size_t latencies_size = sizeof(uint64_t) * g_iterations;

        worst_latencies_ns = (uint64_t*) valloc(latencies_size);
        assert(worst_latencies_ns);

        /* Ensure the allocation is pre-faulted */
        ret = memset_s(worst_latencies_ns, latencies_size, 0, latencies_size);
        if (ret) errc(EX_OSERR, ret, "memset_s latencies");

        worst_latencies_from_first_ns = (uint64_t*) valloc(latencies_size);
        assert(worst_latencies_from_first_ns);

        /* Ensure the allocation is pre-faulted */
        ret = memset_s(worst_latencies_from_first_ns, latencies_size, 0, latencies_size);
        if (ret) errc(EX_OSERR, ret, "memset_s latencies_from_first");

        kr = semaphore_create(mach_task_self(), &g_main_sem, SYNC_POLICY_FIFO, 0);
        mach_assert_zero(kr);

        /* Either one big semaphore or one per thread */
        if (g_waketype == WAKE_CHAIN ||
            g_waketype == WAKE_BROADCAST_PERTHREAD ||
            g_waketype == WAKE_HOP) {

                g_semarr = valloc(sizeof(semaphore_t) * g_numthreads);
                assert(g_semarr);

                for (uint32_t i = 0; i < g_numthreads; i++) {
                        kr = semaphore_create(mach_task_self(), &g_semarr[i], SYNC_POLICY_FIFO, 0);
                        mach_assert_zero(kr);
                }

                g_leadersem = g_semarr[0];
        } else {
                kr = semaphore_create(mach_task_self(), &g_broadcastsem, SYNC_POLICY_FIFO, 0);
                mach_assert_zero(kr);
                kr = semaphore_create(mach_task_self(), &g_leadersem, SYNC_POLICY_FIFO, 0);
                mach_assert_zero(kr);
        }

        if (g_waketype == WAKE_HOP) {
                kr = semaphore_create(mach_task_self(), &g_donesem, SYNC_POLICY_FIFO, 0);
                mach_assert_zero(kr);
        }

        kr = semaphore_create(mach_task_self(), &g_readysem, SYNC_POLICY_FIFO, 0);
        mach_assert_zero(kr);

        atomic_store_explicit(&g_done_threads, 0, memory_order_relaxed);

        /* Create the threads */
        for (uint32_t i = 0; i < g_numthreads; i++) {
                ret = pthread_create(&threads[i], NULL, worker_thread, (void*)(uintptr_t)i);
                if (ret) errc(EX_OSERR, ret, "pthread_create %d", i);
        }

        ret = setpriority(PRIO_DARWIN_ROLE, 0, PRIO_DARWIN_ROLE_UI_FOCAL);
        if (ret) errc(EX_OSERR, ret, "setpriority");

        thread_setup(0);

        g_starttime_abs = mach_absolute_time();

        if (g_churn_pri)
                create_churn_threads();

        /* Let everyone get settled */
        kr = semaphore_wait(g_main_sem);
        mach_assert_zero(kr);

        /* Give the system a bit more time to settle */
        if (g_do_sleep)
                usleep(g_iteration_sleeptime_us);

        /* Go! */
        for (uint32_t i = 0; i < g_iterations; i++) {
                uint32_t j;
                uint64_t worst_abs = 0, best_abs = UINT64_MAX;

                if (g_do_one_long_spin)
                        g_one_long_spin_id = (uint32_t)rand() % g_numthreads;

                debug_log("%d Main thread reset\n", i);

                atomic_store_explicit(&g_done_threads, 0, memory_order_seq_cst);

                g_starttime_abs = mach_absolute_time();

                /* Fire them off and wait for worker threads to finish */
                kr = semaphore_wait_signal(g_main_sem, g_leadersem);
                mach_assert_zero(kr);

                debug_log("%d Main thread return\n", i);

                assert(atomic_load_explicit(&g_done_threads, memory_order_relaxed) == g_numthreads);

                /*
                 * We report the worst latencies relative to start time
                 * and relative to the lead worker thread.
                 */
                for (j = 0; j < g_numthreads; j++) {
                        uint64_t latency_abs;

                        latency_abs = g_thread_endtimes_abs[j] - g_starttime_abs;
                        worst_abs = worst_abs < latency_abs ? latency_abs : worst_abs;
                }
        
                worst_latencies_ns[i] = abs_to_nanos(worst_abs);

                worst_abs = 0;
                for (j = 1; j < g_numthreads; j++) {
                        uint64_t latency_abs;
                
                        latency_abs = g_thread_endtimes_abs[j] - g_thread_endtimes_abs[0];
                        worst_abs = worst_abs < latency_abs ? latency_abs : worst_abs;
                        best_abs = best_abs > latency_abs ? latency_abs : best_abs;
                }

                worst_latencies_from_first_ns[i] = abs_to_nanos(worst_abs);

                /*
                 * In the event of a bad run, cut a trace point.
                 */
                if (worst_latencies_from_first_ns[i] > g_traceworthy_latency_ns) {
                        /* Ariadne's ad-hoc test signpost */
                        kdebug_trace(ARIADNEDBG_CODE(0, 0), worst_latencies_from_first_ns[i], g_traceworthy_latency_ns, 0, 0);

                        if (g_verbose)
                                printf("Worst on this round was %.2f us.\n", ((float)worst_latencies_from_first_ns[i]) / 1000.0);
                }

                /* Give the system a bit more time to settle */
                if (g_do_sleep)
                        usleep(g_iteration_sleeptime_us);
        }

        /* Rejoin threads */
        for (uint32_t i = 0; i < g_numthreads; i++) {
                ret = pthread_join(threads[i], NULL);
                if (ret) errc(EX_OSERR, ret, "pthread_join %d", i);
        }

        if (g_churn_pri)
                join_churn_threads();

        compute_stats(worst_latencies_ns, g_iterations, &avg, &max, &min, &stddev);
        printf("Results (from a stop):\n");
        printf("Max:\t\t%.2f us\n", ((float)max) / 1000.0);
        printf("Min:\t\t%.2f us\n", ((float)min) / 1000.0);
        printf("Avg:\t\t%.2f us\n", avg / 1000.0);
        printf("Stddev:\t\t%.2f us\n", stddev / 1000.0);

        putchar('\n');

        compute_stats(worst_latencies_from_first_ns, g_iterations, &avg, &max, &min, &stddev);
        printf("Results (relative to first thread):\n");
        printf("Max:\t\t%.2f us\n", ((float)max) / 1000.0);
        printf("Min:\t\t%.2f us\n", ((float)min) / 1000.0);
        printf("Avg:\t\t%.2f us\n", avg / 1000.0);
        printf("Stddev:\t\t%.2f us\n", stddev / 1000.0);

#if 0
        for (uint32_t i = 0; i < g_iterations; i++) {
                printf("Iteration %d: %f us\n", i, worst_latencies_ns[i] / 1000.0);
        }
#endif

        free(threads);
        free(g_thread_endtimes_abs);
        free(worst_latencies_ns);
        free(worst_latencies_from_first_ns);

        return 0;
}

/*
 * WARNING: This is SPI specifically intended for use by launchd to start UI
 * apps. We use it here for a test tool only to opt into QoS using the same
 * policies. Do not use this outside xnu or libxpc/launchd.
 */
static void
selfexec_with_apptype(int argc, char *argv[])
{
        int ret;
        posix_spawnattr_t attr;
        extern char **environ;
        char *new_argv[argc + 1 + 1 /* NULL */];
        int i;
        char prog[PATH_MAX];
        uint32_t prog_size = PATH_MAX;

        ret = _NSGetExecutablePath(prog, &prog_size);
        if (ret) err(EX_OSERR, "_NSGetExecutablePath");

        for (i=0; i < argc; i++) {
                new_argv[i] = argv[i];
        }

        new_argv[i]   = "--switched_apptype";
        new_argv[i+1] = NULL;

        ret = posix_spawnattr_init(&attr);
        if (ret) errc(EX_OSERR, ret, "posix_spawnattr_init");

        ret = posix_spawnattr_setflags(&attr, POSIX_SPAWN_SETEXEC);
        if (ret) errc(EX_OSERR, ret, "posix_spawnattr_setflags");

        ret = posix_spawnattr_setprocesstype_np(&attr, POSIX_SPAWN_PROC_TYPE_APP_DEFAULT);
        if (ret) errc(EX_OSERR, ret, "posix_spawnattr_setprocesstype_np");

        ret = posix_spawn(NULL, prog, NULL, &attr, new_argv, environ);
        if (ret) errc(EX_OSERR, ret, "posix_spawn");
}

/*
 * Admittedly not very attractive.
 */
static void __attribute__((noreturn))
usage()
{
        errx(EX_USAGE, "Usage: %s <threads> <chain | hop | broadcast-single-sem | broadcast-per-thread> "
             "<realtime | timeshare | fixed> <iterations>\n\t\t"
             "[--trace <traceworthy latency in ns>] "
             "[--verbose] [--spin-one] [--spin-all] [--spin-time <nanos>] [--affinity]\n\t\t"
             "[--no-sleep] [--drop-priority] [--churn-pri <pri>] [--churn-count <n>]",
             getprogname());
}

static struct option* g_longopts;
static int option_index;

static uint32_t
read_dec_arg()
{
        char *cp;
        /* char* optarg is a magic global */

        uint32_t arg_val = (uint32_t)strtoull(optarg, &cp, 10);

        if (cp == optarg || *cp)
                errx(EX_USAGE, "arg --%s requires a decimal number, found \"%s\"",
                     g_longopts[option_index].name, optarg);

        return arg_val;
}

static void
parse_args(int argc, char *argv[])
{
        enum {
                OPT_GETOPT = 0,
                OPT_SPIN_TIME,
                OPT_TRACE,
                OPT_PRIORITY,
                OPT_CHURN_PRI,
                OPT_CHURN_COUNT,
        };

        static struct option longopts[] = {
                { "spin-time",          required_argument,      NULL,                           OPT_SPIN_TIME },
                { "trace",              required_argument,      NULL,                           OPT_TRACE     },
                { "priority",           required_argument,      NULL,                           OPT_PRIORITY  },
                { "churn-pri",          required_argument,      NULL,                           OPT_CHURN_PRI },
                { "churn-count",        required_argument,      NULL,                           OPT_CHURN_COUNT },
                { "switched_apptype",   no_argument,            (int*)&g_seen_apptype,          TRUE },
                { "spin-one",           no_argument,            (int*)&g_do_one_long_spin,      TRUE },
                { "spin-all",           no_argument,            (int*)&g_do_all_spin,           TRUE },
                { "affinity",           no_argument,            (int*)&g_do_affinity,           TRUE },
                { "no-sleep",           no_argument,            (int*)&g_do_sleep,              FALSE },
                { "drop-priority",      no_argument,            (int*)&g_drop_priority,         TRUE },
                { "verbose",            no_argument,            (int*)&g_verbose,               TRUE },
                { "help",               no_argument,            NULL,                           'h' },
                { NULL,                 0,                      NULL,                           0 }
        };

        g_longopts = longopts;
        int ch = 0;

        while ((ch = getopt_long(argc, argv, "h", longopts, &option_index)) != -1) {
                switch (ch) {
                case OPT_GETOPT:
                        /* getopt_long set a variable */
                        break;
                case OPT_SPIN_TIME:
                        g_do_each_spin = TRUE;
                        g_each_spin_duration_ns = read_dec_arg();
                        break;
                case OPT_TRACE:
                        g_traceworthy_latency_ns = read_dec_arg();
                        break;
                case OPT_PRIORITY:
                        g_priority = read_dec_arg();
                        break;
                case OPT_CHURN_PRI:
                        g_churn_pri = read_dec_arg();
                        break;
                case OPT_CHURN_COUNT:
                        g_churn_count = read_dec_arg();
                        break;
                case '?':
                case 'h':
                default:
                        usage();
                        /* NORETURN */
                }
        }

        /*
         * getopt_long reorders all the options to the beginning of the argv array.
         * Jump past them to the non-option arguments.
         */

        argc -= optind;
        argv += optind;

        if (argc > 4) {
                warnx("Too many non-option arguments passed");
                usage();
        }

        if (argc != 4) {
                warnx("Missing required <threads> <waketype> <policy> <iterations> arguments");
                usage();
        }

        char *cp;

        /* How many threads? */
        g_numthreads = (uint32_t)strtoull(argv[0], &cp, 10);

        if (cp == argv[0] || *cp)
                errx(EX_USAGE, "numthreads requires a decimal number, found \"%s\"", argv[0]);

        if (g_numthreads < 1)
                errx(EX_USAGE, "Must use at least one thread");

        /* What wakeup pattern? */
        g_waketype = parse_wakeup_pattern(argv[1]);

        /* Policy */
        g_policy = parse_thread_policy(argv[2]);

        /* Iterations */
        g_iterations = (uint32_t)strtoull(argv[3], &cp, 10);

        if (cp == argv[3] || *cp)
                errx(EX_USAGE, "numthreads requires a decimal number, found \"%s\"", argv[3]);

        if (g_iterations < 1)
                errx(EX_USAGE, "Must have at least one iteration");

        if (g_numthreads == 1 && g_waketype == WAKE_CHAIN)
                errx(EX_USAGE, "chain mode requires more than one thread");

        if (g_numthreads == 1 && g_waketype == WAKE_HOP)
                errx(EX_USAGE, "hop mode requires more than one thread");
}