tools/tests/darwintests/contextswitch.c

   1 #include <unistd.h>
   2 #include <stdio.h>
   3 #include <stdlib.h>
   4 #include <pthread.h>
   5 #include <errno.h>
   6 #include <err.h>
   7 #include <string.h>
   8 #include <assert.h>
   9 #include <sysexits.h>
  10 #include <getopt.h>
  11 #include <spawn.h>
  12 #include <stdbool.h>
  13 #include <sys/sysctl.h>
  14 #include <mach/mach_time.h>
  15 #include <mach/mach.h>
  16 #include <mach/semaphore.h>
  17 #include <TargetConditionals.h>
  18
  19 #ifdef T_NAMESPACE
  20 #undef T_NAMESPACE
  21 #endif
  22
  23 #include <darwintest.h>
  24 #include <stdatomic.h>
  25
  26 #define MAX_THREADS     32
  27 #define SPIN_SECS       6
  28 #define THR_SPINNER_PRI 63
  29 #define THR_MANAGER_PRI 62
  30 #define WARMUP_ITERATIONS 100
  31 #define POWERCTRL_SUCCESS_STR "Factor1: 1.000000"
  32
  33 static mach_timebase_info_data_t timebase_info;
  34 static semaphore_t semaphore;
  35 static semaphore_t worker_sem;
  36 static uint32_t g_numcpus;
  37 static _Atomic uint32_t keep_going = 1;
  38 static dt_stat_time_t s;
  39
  40 static struct {
  41     pthread_t thread;
  42     bool measure_thread;
  43 } threads[MAX_THREADS];
  44
  45 static uint64_t
  46 nanos_to_abs(uint64_t nanos)
  47 {
  48     return nanos * timebase_info.denom / timebase_info.numer;
  49 }
  50
  51 extern char **environ;
  52
  53 static void
  54 csw_perf_test_init(void)
  55 {
  56     int spawn_ret, pid;
  57     char *const clpcctrl_args[] = {"/usr/local/bin/clpcctrl", "-f", "5000", NULL};
  58     spawn_ret = posix_spawn(&pid, clpcctrl_args[0], NULL, NULL, clpcctrl_args, environ);
  59     waitpid(pid, &spawn_ret, 0);
  60 }
  61
  62 static void
  63 csw_perf_test_cleanup(void)
  64 {
  65     int spawn_ret, pid;
  66     char *const clpcctrl_args[] = {"/usr/local/bin/clpcctrl", "-d", NULL};
  67     spawn_ret = posix_spawn(&pid, clpcctrl_args[0], NULL, NULL, clpcctrl_args, environ);
  68     waitpid(pid, &spawn_ret, 0);
  69 }
  70
  71 static pthread_t
  72 create_thread(uint32_t thread_id, uint32_t priority, bool fixpri,
  73         void *(*start_routine)(void *))
  74 {
  75     int rv;
  76     pthread_t new_thread;
  77     struct sched_param param = { .sched_priority = (int)priority };
  78     pthread_attr_t attr;
  79
  80     T_ASSERT_POSIX_ZERO(pthread_attr_init(&attr), "pthread_attr_init");
  81
  82     T_ASSERT_POSIX_ZERO(pthread_attr_setschedparam(&attr, &param),
  83             "pthread_attr_setschedparam");
  84
  85     if (fixpri) {
  86         T_ASSERT_POSIX_ZERO(pthread_attr_setschedpolicy(&attr, SCHED_RR),
  87                 "pthread_attr_setschedpolicy");
  88     }
  89
  90     T_ASSERT_POSIX_ZERO(pthread_create(&new_thread, &attr, start_routine,
  91             (void*)(uintptr_t)thread_id), "pthread_create");
  92
  93     T_ASSERT_POSIX_ZERO(pthread_attr_destroy(&attr), "pthread_attr_destroy");
  94
  95     threads[thread_id].thread = new_thread;
  96
  97     return new_thread;
  98 }
  99
 100 /* Spin until a specified number of seconds elapses */
 101 static void
 102 spin_for_duration(uint32_t seconds)
 103 {
 104     uint64_t duration       = nanos_to_abs((uint64_t)seconds * NSEC_PER_SEC);
 105     uint64_t current_time   = mach_absolute_time();
 106     uint64_t timeout        = duration + current_time;
 107
 108     uint64_t spin_count = 0;
 109
 110     while (mach_absolute_time() < timeout && atomic_load_explicit(&keep_going,
 111                 memory_order_relaxed)) {
 112         spin_count++;
 113     }
 114 }
 115
 116 static void *
 117 spin_thread(void *arg)
 118 {
 119     uint32_t thread_id = (uint32_t) arg;
 120     char name[30] = "";
 121
 122     snprintf(name, sizeof(name), "spin thread %2d", thread_id);
 123     pthread_setname_np(name);
 124     T_ASSERT_MACH_SUCCESS(semaphore_wait_signal(semaphore, worker_sem),
 125             "semaphore_wait_signal");
 126     spin_for_duration(SPIN_SECS);
 127     return NULL;
 128 }
 129
 130 static void *
 131 thread(void *arg)
 132 {
 133     uint32_t thread_id = (uint32_t) arg;
 134     char name[30] = "";
 135
 136     snprintf(name, sizeof(name), "thread %2d", thread_id);
 137     pthread_setname_np(name);
 138     T_ASSERT_MACH_SUCCESS(semaphore_wait_signal(semaphore, worker_sem), "semaphore_wait");
 139
 140     if (threads[thread_id].measure_thread) {
 141         for (int i = 0; i < WARMUP_ITERATIONS; i++) {
 142             thread_switch(THREAD_NULL, SWITCH_OPTION_NONE, 0);
 143         }
 144         T_STAT_MEASURE_LOOP(s) {
 145             if(thread_switch(THREAD_NULL, SWITCH_OPTION_NONE, 0))
 146                 T_ASSERT_FAIL("thread_switch");
 147         }
 148         atomic_store_explicit(&keep_going, 0, memory_order_relaxed);
 149     } else {
 150         while (atomic_load_explicit(&keep_going, memory_order_relaxed)) {
 151             if (thread_switch(THREAD_NULL, SWITCH_OPTION_NONE, 0))
 152                 T_ASSERT_FAIL("thread_switch");
 153         }
 154     }
 155     return NULL;
 156 }
 157
 158 void check_device_temperature(void)
 159 {
 160     char buffer[256];
 161     FILE *pipe = popen("powerctrl Factor1", "r");
 162
 163     if (pipe == NULL) {
 164         T_FAIL("Failed to check device temperature");
 165         T_END;
 166     }
 167
 168     fgets(buffer, sizeof(buffer), pipe);
 169
 170     if (strncmp(POWERCTRL_SUCCESS_STR, buffer, strlen(POWERCTRL_SUCCESS_STR))) {
 171         T_PERF("temperature", 0.0, "factor", "device temperature");
 172     } else {
 173         T_PASS("Device temperature check pass");
 174         T_PERF("temperature", 1.0, "factor", "device temperature");
 175     }
 176     pclose(pipe);
 177 }
 178
 179 void record_perfcontrol_stats(const char *sysctlname, const char *units, const char *info)
 180 {
 181     int data = 0;
 182     size_t data_size = sizeof(data);
 183     T_ASSERT_POSIX_ZERO(sysctlbyname(sysctlname,
 184             &data, &data_size, NULL, 0),
 185             "%s", sysctlname);
 186     T_PERF(info, data, units, info);
 187 }
 188
 189
 190 T_GLOBAL_META(T_META_NAMESPACE("xnu.scheduler"));
 191
 192 /* Disable the test on MacOS for now */
 193 T_DECL(perf_csw, "context switch performance", T_META_TYPE_PERF, T_META_CHECK_LEAKS(NO), T_META_ASROOT(YES))
 194 {
 195
 196 #if !CONFIG_EMBEDDED
 197     T_SKIP("Not supported on MacOS");
 198     return;
 199 #endif /* CONFIG_EMBEDDED */
 200     check_device_temperature();
 201
 202     T_ATEND(csw_perf_test_cleanup);
 203
 204     csw_perf_test_init();
 205     pthread_setname_np("main thread");
 206
 207     T_ASSERT_MACH_SUCCESS(mach_timebase_info(&timebase_info), "mach_timebase_info");
 208
 209     struct sched_param param = {.sched_priority = 48};
 210
 211     T_ASSERT_POSIX_ZERO(pthread_setschedparam(pthread_self(), SCHED_FIFO, &param),
 212             "pthread_setschedparam");
 213
 214     T_ASSERT_MACH_SUCCESS(semaphore_create(mach_task_self(), &semaphore,
 215             SYNC_POLICY_FIFO, 0), "semaphore_create");
 216
 217     T_ASSERT_MACH_SUCCESS(semaphore_create(mach_task_self(), &worker_sem,
 218             SYNC_POLICY_FIFO, 0), "semaphore_create");
 219
 220     size_t ncpu_size = sizeof(g_numcpus);
 221     T_ASSERT_POSIX_ZERO(sysctlbyname("hw.ncpu", &g_numcpus, &ncpu_size, NULL, 0),
 222             "sysctlbyname hw.ncpu");
 223
 224     printf("hw.ncpu: %d\n", g_numcpus);
 225     uint32_t n_spinners = g_numcpus - 1;
 226
 227     int mt_supported = 0;
 228     size_t mt_supported_size = sizeof(mt_supported);
 229     T_ASSERT_POSIX_ZERO(sysctlbyname("kern.monotonic.supported", &mt_supported,
 230             &mt_supported_size, NULL, 0), "sysctlbyname kern.monotonic.supported");
 231
 232     for (uint32_t thread_id = 0; thread_id < n_spinners; thread_id++) {
 233         threads[thread_id].thread = create_thread(thread_id, THR_SPINNER_PRI,
 234                 true, &spin_thread);
 235     }
 236
 237     s = dt_stat_time_create("context switch time");
 238
 239     create_thread(n_spinners, THR_MANAGER_PRI, true, &thread);
 240     threads[n_spinners].measure_thread = true;
 241     create_thread(n_spinners + 1, THR_MANAGER_PRI, true, &thread);
 242
 243     /* Allow the context switch threads to get into sem_wait() */
 244     for (uint32_t thread_id = 0; thread_id < n_spinners + 2; thread_id++) {
 245         T_ASSERT_MACH_SUCCESS(semaphore_wait(worker_sem), "semaphore_wait");
 246     }
 247
 248     int enable_callout_stats = 1;
 249     size_t enable_size = sizeof(enable_callout_stats);
 250
 251     if (mt_supported) {
 252         /* Enable callout stat collection */
 253         T_ASSERT_POSIX_ZERO(sysctlbyname("kern.perfcontrol_callout.stats_enabled",
 254                 NULL, 0, &enable_callout_stats, enable_size),
 255                 "sysctlbyname kern.perfcontrol_callout.stats_enabled");
 256     }
 257
 258     T_ASSERT_MACH_SUCCESS(semaphore_signal_all(semaphore), "semaphore_signal");
 259
 260
 261     for (uint32_t thread_id = 0; thread_id < n_spinners + 2; thread_id++) {
 262         T_ASSERT_POSIX_ZERO(pthread_join(threads[thread_id].thread, NULL),
 263                 "pthread_join %d", thread_id);
 264     }
 265
 266     if (mt_supported) {
 267         record_perfcontrol_stats("kern.perfcontrol_callout.oncore_instr",
 268                 "instructions", "oncore.instructions");
 269         record_perfcontrol_stats("kern.perfcontrol_callout.offcore_instr",
 270                 "instructions", "offcore.instructions");
 271         record_perfcontrol_stats("kern.perfcontrol_callout.oncore_cycles",
 272                 "cycles", "oncore.cycles");
 273         record_perfcontrol_stats("kern.perfcontrol_callout.offcore_cycles",
 274                 "cycles", "offcore.cycles");
 275
 276         /* Disable callout stat collection */
 277         enable_callout_stats = 0;
 278         T_ASSERT_POSIX_ZERO(sysctlbyname("kern.perfcontrol_callout.stats_enabled",
 279                 NULL, 0, &enable_callout_stats, enable_size),
 280                 "sysctlbyname kern.perfcontrol_callout.stats_enabled");
 281     }
 282
 283     check_device_temperature();
 284     dt_stat_finalize(s);
 285 }