]> git.saurik.com Git - apple/xnu.git/blob - tools/tests/zero-to-n/zero-to-n.c
f8cccbefa4e30ff79bd26cd8f75cc7335398ed3a
[apple/xnu.git] / tools / tests / zero-to-n / zero-to-n.c
1 /*
2 * Copyright (c) 2009 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28 #include <unistd.h>
29 #include <stdio.h>
30 #include <math.h>
31 #include <sys/kdebug.h>
32 #include <stdlib.h>
33 #include <pthread.h>
34 #include <errno.h>
35 #include <err.h>
36 #include <string.h>
37 #include <assert.h>
38 #include <sysexits.h>
39 #include <sys/sysctl.h>
40 #include <getopt.h>
41 #include <libproc.h>
42
43 #include <spawn.h>
44 #include <spawn_private.h>
45 #include <sys/spawn_internal.h>
46 #include <mach-o/dyld.h>
47
48 #include <mach/mach_time.h>
49 #include <mach/mach.h>
50 #include <mach/task.h>
51 #include <mach/semaphore.h>
52
53 #include <pthread/qos_private.h>
54
55 #include <sys/resource.h>
56
57 #include <stdatomic.h>
58
59 #include <os/tsd.h>
60 #include <os/lock.h>
61 #include <TargetConditionals.h>
62
63 typedef enum wake_type { WAKE_BROADCAST_ONESEM, WAKE_BROADCAST_PERTHREAD, WAKE_CHAIN, WAKE_HOP } wake_type_t;
64 typedef enum my_policy_type { MY_POLICY_REALTIME, MY_POLICY_TIMESHARE, MY_POLICY_TIMESHARE_NO_SMT, MY_POLICY_FIXEDPRI } my_policy_type_t;
65
66 #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)
67 #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)
68 #define assert_zero_t(tid, error) do { if ((error) != 0) { fprintf(stderr, "[FAIL] Thread %d error %d ", (tid), (error)); assert(error == 0); } } while (0)
69
70 #define CONSTRAINT_NANOS (20000000ll) /* 20 ms */
71 #define COMPUTATION_NANOS (10000000ll) /* 10 ms */
72 #define LL_CONSTRAINT_NANOS ( 2000000ll) /* 2 ms */
73 #define LL_COMPUTATION_NANOS ( 1000000ll) /* 1 ms */
74 #define RT_CHURN_COMP_NANOS ( 1000000ll) /* 1 ms */
75 #define TRACEWORTHY_NANOS (10000000ll) /* 10 ms */
76 #define TRACEWORTHY_NANOS_TEST ( 2000000ll) /* 2 ms */
77
78 #if DEBUG
79 #define debug_log(args ...) printf(args)
80 #else
81 #define debug_log(args ...) do { } while(0)
82 #endif
83
84 /* Declarations */
85 static void* worker_thread(void *arg);
86 static void usage();
87 static int thread_setup(uint32_t my_id);
88 static my_policy_type_t parse_thread_policy(const char *str);
89 static void selfexec_with_apptype(int argc, char *argv[]);
90 static void parse_args(int argc, char *argv[]);
91
92 static __attribute__((aligned(128))) _Atomic uint32_t g_done_threads;
93 static __attribute__((aligned(128))) _Atomic boolean_t g_churn_stop = FALSE;
94 static __attribute__((aligned(128))) _Atomic uint64_t g_churn_stopped_at = 0;
95
96 /* Global variables (general) */
97 static uint32_t g_numcpus;
98 static uint32_t g_nphysicalcpu;
99 static uint32_t g_nlogicalcpu;
100 static uint32_t g_numthreads;
101 static wake_type_t g_waketype;
102 static policy_t g_policy;
103 static uint32_t g_iterations;
104 static struct mach_timebase_info g_mti;
105 static semaphore_t g_main_sem;
106 static uint64_t *g_thread_endtimes_abs;
107 static boolean_t g_verbose = FALSE;
108 static boolean_t g_do_affinity = FALSE;
109 static uint64_t g_starttime_abs;
110 static uint32_t g_iteration_sleeptime_us = 0;
111 static uint32_t g_priority = 0;
112 static uint32_t g_churn_pri = 0;
113 static uint32_t g_churn_count = 0;
114 static uint32_t g_rt_churn_count = 0;
115
116 static pthread_t* g_churn_threads = NULL;
117 static pthread_t* g_rt_churn_threads = NULL;
118
119 /* should we skip test if run on non-intel */
120 static boolean_t g_run_on_intel_only = FALSE;
121
122 /* Threshold for dropping a 'bad run' tracepoint */
123 static uint64_t g_traceworthy_latency_ns = TRACEWORTHY_NANOS;
124
125 /* Have we re-execed to set apptype? */
126 static boolean_t g_seen_apptype = FALSE;
127
128 /* usleep in betweeen iterations */
129 static boolean_t g_do_sleep = TRUE;
130
131 /* Every thread spins until all threads have checked in */
132 static boolean_t g_do_all_spin = FALSE;
133
134 /* Every thread backgrounds temporarily before parking */
135 static boolean_t g_drop_priority = FALSE;
136
137 /* Use low-latency (sub 4ms deadline) realtime threads */
138 static boolean_t g_rt_ll = FALSE;
139
140 /* Test whether realtime threads are scheduled on the separate CPUs */
141 static boolean_t g_test_rt = FALSE;
142
143 static boolean_t g_rt_churn = FALSE;
144
145 /* On SMT machines, test whether realtime threads are scheduled on the correct CPUs */
146 static boolean_t g_test_rt_smt = FALSE;
147
148 /* Test whether realtime threads are successfully avoiding CPU 0 on Intel */
149 static boolean_t g_test_rt_avoid0 = FALSE;
150
151 /* Print a histgram showing how many threads ran on each CPU */
152 static boolean_t g_histogram = FALSE;
153
154 /* One randomly chosen thread holds up the train for a certain duration. */
155 static boolean_t g_do_one_long_spin = FALSE;
156 static uint32_t g_one_long_spin_id = 0;
157 static uint64_t g_one_long_spin_length_abs = 0;
158 static uint64_t g_one_long_spin_length_ns = 0;
159
160 /* Each thread spins for a certain duration after waking up before blocking again. */
161 static boolean_t g_do_each_spin = FALSE;
162 static uint64_t g_each_spin_duration_abs = 0;
163 static uint64_t g_each_spin_duration_ns = 0;
164
165 /* Global variables (broadcast) */
166 static semaphore_t g_broadcastsem;
167 static semaphore_t g_leadersem;
168 static semaphore_t g_readysem;
169 static semaphore_t g_donesem;
170 static semaphore_t g_rt_churn_sem;
171 static semaphore_t g_rt_churn_start_sem;
172
173 /* Global variables (chain) */
174 static semaphore_t *g_semarr;
175
176 typedef struct {
177 __attribute__((aligned(128))) uint32_t current;
178 uint32_t accum;
179 } histogram_t;
180
181 static histogram_t *g_cpu_histogram;
182 static _Atomic uint64_t *g_cpu_map;
183
184 static uint64_t
185 abs_to_nanos(uint64_t abstime)
186 {
187 return (uint64_t)(abstime * (((double)g_mti.numer) / ((double)g_mti.denom)));
188 }
189
190 static uint64_t
191 nanos_to_abs(uint64_t ns)
192 {
193 return (uint64_t)(ns * (((double)g_mti.denom) / ((double)g_mti.numer)));
194 }
195
196 inline static void
197 yield(void)
198 {
199 #if defined(__arm__) || defined(__arm64__)
200 asm volatile ("yield");
201 #elif defined(__x86_64__) || defined(__i386__)
202 asm volatile ("pause");
203 #else
204 #error Unrecognized architecture
205 #endif
206 }
207
208 static void *
209 churn_thread(__unused void *arg)
210 {
211 uint64_t spin_count = 0;
212
213 /*
214 * As a safety measure to avoid wedging, we will bail on the spin if
215 * it's been more than 1s after the most recent run start
216 */
217
218 while (g_churn_stop == FALSE &&
219 mach_absolute_time() < (g_starttime_abs + NSEC_PER_SEC)) {
220 spin_count++;
221 yield();
222 }
223
224 /* This is totally racy, but only here to detect if anyone stops early */
225 atomic_fetch_add_explicit(&g_churn_stopped_at, spin_count, memory_order_relaxed);
226
227 return NULL;
228 }
229
230 static void
231 create_churn_threads()
232 {
233 if (g_churn_count == 0) {
234 g_churn_count = g_test_rt_smt ? g_numcpus : g_numcpus - 1;
235 }
236
237 errno_t err;
238
239 struct sched_param param = { .sched_priority = (int)g_churn_pri };
240 pthread_attr_t attr;
241
242 /* Array for churn threads */
243 g_churn_threads = (pthread_t*) valloc(sizeof(pthread_t) * g_churn_count);
244 assert(g_churn_threads);
245
246 if ((err = pthread_attr_init(&attr))) {
247 errc(EX_OSERR, err, "pthread_attr_init");
248 }
249
250 if ((err = pthread_attr_setschedparam(&attr, &param))) {
251 errc(EX_OSERR, err, "pthread_attr_setschedparam");
252 }
253
254 if ((err = pthread_attr_setschedpolicy(&attr, SCHED_RR))) {
255 errc(EX_OSERR, err, "pthread_attr_setschedpolicy");
256 }
257
258 for (uint32_t i = 0; i < g_churn_count; i++) {
259 pthread_t new_thread;
260
261 if ((err = pthread_create(&new_thread, &attr, churn_thread, NULL))) {
262 errc(EX_OSERR, err, "pthread_create");
263 }
264 g_churn_threads[i] = new_thread;
265 }
266
267 if ((err = pthread_attr_destroy(&attr))) {
268 errc(EX_OSERR, err, "pthread_attr_destroy");
269 }
270 }
271
272 static void
273 join_churn_threads(void)
274 {
275 if (atomic_load_explicit(&g_churn_stopped_at, memory_order_seq_cst) != 0) {
276 printf("Warning: Some of the churn threads may have stopped early: %lld\n",
277 g_churn_stopped_at);
278 }
279
280 atomic_store_explicit(&g_churn_stop, TRUE, memory_order_seq_cst);
281
282 /* Rejoin churn threads */
283 for (uint32_t i = 0; i < g_churn_count; i++) {
284 errno_t err = pthread_join(g_churn_threads[i], NULL);
285 if (err) {
286 errc(EX_OSERR, err, "pthread_join %d", i);
287 }
288 }
289 }
290
291 /*
292 * Set policy
293 */
294 static int
295 rt_churn_thread_setup(void)
296 {
297 kern_return_t kr;
298 thread_time_constraint_policy_data_t pol;
299
300 /* Hard-coded realtime parameters (similar to what Digi uses) */
301 pol.period = 100000;
302 pol.constraint = (uint32_t) nanos_to_abs(CONSTRAINT_NANOS * 2);
303 pol.computation = (uint32_t) nanos_to_abs(RT_CHURN_COMP_NANOS * 2);
304 pol.preemptible = 0; /* Ignored by OS */
305
306 kr = thread_policy_set(mach_thread_self(), THREAD_TIME_CONSTRAINT_POLICY,
307 (thread_policy_t) &pol, THREAD_TIME_CONSTRAINT_POLICY_COUNT);
308 mach_assert_zero_t(0, kr);
309
310 return 0;
311 }
312
313 static void *
314 rt_churn_thread(__unused void *arg)
315 {
316 rt_churn_thread_setup();
317
318 for (uint32_t i = 0; i < g_iterations; i++) {
319 kern_return_t kr = semaphore_wait_signal(g_rt_churn_start_sem, g_rt_churn_sem);
320 mach_assert_zero_t(0, kr);
321
322 volatile double x = 0.0;
323 volatile double y = 0.0;
324
325 uint64_t endspin = mach_absolute_time() + nanos_to_abs(RT_CHURN_COMP_NANOS);
326 while (mach_absolute_time() < endspin) {
327 y = y + 1.5 + x;
328 x = sqrt(y);
329 }
330 }
331
332 kern_return_t kr = semaphore_signal(g_rt_churn_sem);
333 mach_assert_zero_t(0, kr);
334
335 return NULL;
336 }
337
338 static void
339 wait_for_rt_churn_threads(void)
340 {
341 for (uint32_t i = 0; i < g_rt_churn_count; i++) {
342 kern_return_t kr = semaphore_wait(g_rt_churn_sem);
343 mach_assert_zero_t(0, kr);
344 }
345 }
346
347 static void
348 start_rt_churn_threads(void)
349 {
350 for (uint32_t i = 0; i < g_rt_churn_count; i++) {
351 kern_return_t kr = semaphore_signal(g_rt_churn_start_sem);
352 mach_assert_zero_t(0, kr);
353 }
354 }
355
356 static void
357 create_rt_churn_threads(void)
358 {
359 if (g_rt_churn_count == 0) {
360 /* Leave 1 CPU to ensure that the main thread can make progress */
361 g_rt_churn_count = g_numcpus - 1;
362 }
363
364 errno_t err;
365
366 struct sched_param param = { .sched_priority = (int)g_churn_pri };
367 pthread_attr_t attr;
368
369 /* Array for churn threads */
370 g_rt_churn_threads = (pthread_t*) valloc(sizeof(pthread_t) * g_rt_churn_count);
371 assert(g_rt_churn_threads);
372
373 if ((err = pthread_attr_init(&attr))) {
374 errc(EX_OSERR, err, "pthread_attr_init");
375 }
376
377 if ((err = pthread_attr_setschedparam(&attr, &param))) {
378 errc(EX_OSERR, err, "pthread_attr_setschedparam");
379 }
380
381 if ((err = pthread_attr_setschedpolicy(&attr, SCHED_RR))) {
382 errc(EX_OSERR, err, "pthread_attr_setschedpolicy");
383 }
384
385 for (uint32_t i = 0; i < g_rt_churn_count; i++) {
386 pthread_t new_thread;
387
388 if ((err = pthread_create(&new_thread, &attr, rt_churn_thread, NULL))) {
389 errc(EX_OSERR, err, "pthread_create");
390 }
391 g_rt_churn_threads[i] = new_thread;
392 }
393
394 if ((err = pthread_attr_destroy(&attr))) {
395 errc(EX_OSERR, err, "pthread_attr_destroy");
396 }
397
398 /* Wait until all threads have checked in */
399 wait_for_rt_churn_threads();
400 }
401
402 static void
403 join_rt_churn_threads(void)
404 {
405 /* Rejoin rt churn threads */
406 for (uint32_t i = 0; i < g_rt_churn_count; i++) {
407 errno_t err = pthread_join(g_rt_churn_threads[i], NULL);
408 if (err) {
409 errc(EX_OSERR, err, "pthread_join %d", i);
410 }
411 }
412 }
413
414 /*
415 * Figure out what thread policy to use
416 */
417 static my_policy_type_t
418 parse_thread_policy(const char *str)
419 {
420 if (strcmp(str, "timeshare") == 0) {
421 return MY_POLICY_TIMESHARE;
422 } else if (strcmp(str, "timeshare_no_smt") == 0) {
423 return MY_POLICY_TIMESHARE_NO_SMT;
424 } else if (strcmp(str, "realtime") == 0) {
425 return MY_POLICY_REALTIME;
426 } else if (strcmp(str, "fixed") == 0) {
427 return MY_POLICY_FIXEDPRI;
428 } else {
429 errx(EX_USAGE, "Invalid thread policy \"%s\"", str);
430 }
431 }
432
433 /*
434 * Figure out what wakeup pattern to use
435 */
436 static wake_type_t
437 parse_wakeup_pattern(const char *str)
438 {
439 if (strcmp(str, "chain") == 0) {
440 return WAKE_CHAIN;
441 } else if (strcmp(str, "hop") == 0) {
442 return WAKE_HOP;
443 } else if (strcmp(str, "broadcast-single-sem") == 0) {
444 return WAKE_BROADCAST_ONESEM;
445 } else if (strcmp(str, "broadcast-per-thread") == 0) {
446 return WAKE_BROADCAST_PERTHREAD;
447 } else {
448 errx(EX_USAGE, "Invalid wakeup pattern \"%s\"", str);
449 }
450 }
451
452 /*
453 * Set policy
454 */
455 static int
456 thread_setup(uint32_t my_id)
457 {
458 kern_return_t kr;
459 errno_t ret;
460 thread_time_constraint_policy_data_t pol;
461
462 if (g_priority) {
463 int policy = SCHED_OTHER;
464 if (g_policy == MY_POLICY_FIXEDPRI) {
465 policy = SCHED_RR;
466 }
467
468 struct sched_param param = {.sched_priority = (int)g_priority};
469 if ((ret = pthread_setschedparam(pthread_self(), policy, &param))) {
470 errc(EX_OSERR, ret, "pthread_setschedparam: %d", my_id);
471 }
472 }
473
474 switch (g_policy) {
475 case MY_POLICY_TIMESHARE:
476 break;
477 case MY_POLICY_TIMESHARE_NO_SMT:
478 proc_setthread_no_smt();
479 break;
480 case MY_POLICY_REALTIME:
481 /* Hard-coded realtime parameters (similar to what Digi uses) */
482 pol.period = 100000;
483 if (g_rt_ll) {
484 pol.constraint = (uint32_t) nanos_to_abs(LL_CONSTRAINT_NANOS);
485 pol.computation = (uint32_t) nanos_to_abs(LL_COMPUTATION_NANOS);
486 } else {
487 pol.constraint = (uint32_t) nanos_to_abs(CONSTRAINT_NANOS);
488 pol.computation = (uint32_t) nanos_to_abs(COMPUTATION_NANOS);
489 }
490 pol.preemptible = 0; /* Ignored by OS */
491
492 kr = thread_policy_set(mach_thread_self(), THREAD_TIME_CONSTRAINT_POLICY,
493 (thread_policy_t) &pol, THREAD_TIME_CONSTRAINT_POLICY_COUNT);
494 mach_assert_zero_t(my_id, kr);
495 break;
496 case MY_POLICY_FIXEDPRI:
497 ret = pthread_set_fixedpriority_self();
498 if (ret) {
499 errc(EX_OSERR, ret, "pthread_set_fixedpriority_self");
500 }
501 break;
502 default:
503 errx(EX_USAGE, "invalid policy type %d", g_policy);
504 }
505
506 if (g_do_affinity) {
507 thread_affinity_policy_data_t affinity;
508
509 affinity.affinity_tag = my_id % 2;
510
511 kr = thread_policy_set(mach_thread_self(), THREAD_AFFINITY_POLICY,
512 (thread_policy_t)&affinity, THREAD_AFFINITY_POLICY_COUNT);
513 mach_assert_zero_t(my_id, kr);
514 }
515
516 return 0;
517 }
518
519 time_value_t
520 get_thread_runtime(void)
521 {
522 thread_basic_info_data_t info;
523 mach_msg_type_number_t info_count = THREAD_BASIC_INFO_COUNT;
524 thread_info(pthread_mach_thread_np(pthread_self()), THREAD_BASIC_INFO, (thread_info_t)&info, &info_count);
525
526 time_value_add(&info.user_time, &info.system_time);
527
528 return info.user_time;
529 }
530
531 time_value_t worker_threads_total_runtime = {};
532
533 /*
534 * Wait for a wakeup, potentially wake up another of the "0-N" threads,
535 * and notify the main thread when done.
536 */
537 static void*
538 worker_thread(void *arg)
539 {
540 static os_unfair_lock runtime_lock = OS_UNFAIR_LOCK_INIT;
541
542 uint32_t my_id = (uint32_t)(uintptr_t)arg;
543 kern_return_t kr;
544
545 volatile double x = 0.0;
546 volatile double y = 0.0;
547
548 /* Set policy and so forth */
549 thread_setup(my_id);
550
551 for (uint32_t i = 0; i < g_iterations; i++) {
552 if (my_id == 0) {
553 /*
554 * Leader thread either wakes everyone up or starts the chain going.
555 */
556
557 /* Give the worker threads undisturbed time to finish before waiting on them */
558 if (g_do_sleep) {
559 usleep(g_iteration_sleeptime_us);
560 }
561
562 debug_log("%d Leader thread wait for ready\n", i);
563
564 /*
565 * Wait for everyone else to declare ready
566 * Is there a better way to do this that won't interfere with the rest of the chain?
567 * TODO: Invent 'semaphore wait for N signals'
568 */
569
570 for (uint32_t j = 0; j < g_numthreads - 1; j++) {
571 kr = semaphore_wait(g_readysem);
572 mach_assert_zero_t(my_id, kr);
573 }
574
575 debug_log("%d Leader thread wait\n", i);
576
577 if (i > 0) {
578 for (int cpuid = 0; cpuid < g_numcpus; cpuid++) {
579 if (g_cpu_histogram[cpuid].current == 1) {
580 atomic_fetch_or_explicit(&g_cpu_map[i - 1], (1UL << cpuid), memory_order_relaxed);
581 g_cpu_histogram[cpuid].current = 0;
582 }
583 }
584 }
585
586 /* Signal main thread and wait for start of iteration */
587
588 kr = semaphore_wait_signal(g_leadersem, g_main_sem);
589 mach_assert_zero_t(my_id, kr);
590
591 g_thread_endtimes_abs[my_id] = mach_absolute_time();
592
593 debug_log("%d Leader thread go\n", i);
594
595 assert_zero_t(my_id, atomic_load_explicit(&g_done_threads, memory_order_relaxed));
596
597 switch (g_waketype) {
598 case WAKE_BROADCAST_ONESEM:
599 kr = semaphore_signal_all(g_broadcastsem);
600 mach_assert_zero_t(my_id, kr);
601 break;
602 case WAKE_BROADCAST_PERTHREAD:
603 for (uint32_t j = 1; j < g_numthreads; j++) {
604 kr = semaphore_signal(g_semarr[j]);
605 mach_assert_zero_t(my_id, kr);
606 }
607 break;
608 case WAKE_CHAIN:
609 kr = semaphore_signal(g_semarr[my_id + 1]);
610 mach_assert_zero_t(my_id, kr);
611 break;
612 case WAKE_HOP:
613 kr = semaphore_wait_signal(g_donesem, g_semarr[my_id + 1]);
614 mach_assert_zero_t(my_id, kr);
615 break;
616 }
617 } else {
618 /*
619 * Everyone else waits to be woken up,
620 * records when she wakes up, and possibly
621 * wakes up a friend.
622 */
623 switch (g_waketype) {
624 case WAKE_BROADCAST_ONESEM:
625 kr = semaphore_wait_signal(g_broadcastsem, g_readysem);
626 mach_assert_zero_t(my_id, kr);
627
628 g_thread_endtimes_abs[my_id] = mach_absolute_time();
629 break;
630
631 case WAKE_BROADCAST_PERTHREAD:
632 kr = semaphore_wait_signal(g_semarr[my_id], g_readysem);
633 mach_assert_zero_t(my_id, kr);
634
635 g_thread_endtimes_abs[my_id] = mach_absolute_time();
636 break;
637
638 case WAKE_CHAIN:
639 kr = semaphore_wait_signal(g_semarr[my_id], g_readysem);
640 mach_assert_zero_t(my_id, kr);
641
642 /* Signal the next thread *after* recording wake time */
643
644 g_thread_endtimes_abs[my_id] = mach_absolute_time();
645
646 if (my_id < (g_numthreads - 1)) {
647 kr = semaphore_signal(g_semarr[my_id + 1]);
648 mach_assert_zero_t(my_id, kr);
649 }
650
651 break;
652
653 case WAKE_HOP:
654 kr = semaphore_wait_signal(g_semarr[my_id], g_readysem);
655 mach_assert_zero_t(my_id, kr);
656
657 /* Signal the next thread *after* recording wake time */
658
659 g_thread_endtimes_abs[my_id] = mach_absolute_time();
660
661 if (my_id < (g_numthreads - 1)) {
662 kr = semaphore_wait_signal(g_donesem, g_semarr[my_id + 1]);
663 mach_assert_zero_t(my_id, kr);
664 } else {
665 kr = semaphore_signal_all(g_donesem);
666 mach_assert_zero_t(my_id, kr);
667 }
668
669 break;
670 }
671 }
672
673 unsigned int cpuid = _os_cpu_number();
674 assert(cpuid < g_numcpus);
675 debug_log("Thread %p woke up on CPU %d for iteration %d.\n", pthread_self(), cpuid, i);
676 g_cpu_histogram[cpuid].current = 1;
677 g_cpu_histogram[cpuid].accum++;
678
679 if (g_do_one_long_spin && g_one_long_spin_id == my_id) {
680 /* One randomly chosen thread holds up the train for a while. */
681
682 uint64_t endspin = g_starttime_abs + g_one_long_spin_length_abs;
683 while (mach_absolute_time() < endspin) {
684 y = y + 1.5 + x;
685 x = sqrt(y);
686 }
687 }
688
689 if (g_do_each_spin) {
690 /* Each thread spins for a certain duration after waking up before blocking again. */
691
692 uint64_t endspin = mach_absolute_time() + g_each_spin_duration_abs;
693 while (mach_absolute_time() < endspin) {
694 y = y + 1.5 + x;
695 x = sqrt(y);
696 }
697 }
698
699 uint32_t done_threads;
700 done_threads = atomic_fetch_add_explicit(&g_done_threads, 1, memory_order_relaxed) + 1;
701
702 debug_log("Thread %p new value is %d, iteration %d\n", pthread_self(), done_threads, i);
703
704 if (g_drop_priority) {
705 /* Drop priority to BG momentarily */
706 errno_t ret = setpriority(PRIO_DARWIN_THREAD, 0, PRIO_DARWIN_BG);
707 if (ret) {
708 errc(EX_OSERR, ret, "setpriority PRIO_DARWIN_BG");
709 }
710 }
711
712 if (g_do_all_spin) {
713 /* Everyone spins until the last thread checks in. */
714
715 while (atomic_load_explicit(&g_done_threads, memory_order_relaxed) < g_numthreads) {
716 y = y + 1.5 + x;
717 x = sqrt(y);
718 }
719 }
720
721 if (g_drop_priority) {
722 /* Restore normal priority */
723 errno_t ret = setpriority(PRIO_DARWIN_THREAD, 0, 0);
724 if (ret) {
725 errc(EX_OSERR, ret, "setpriority 0");
726 }
727 }
728
729 debug_log("Thread %p done spinning, iteration %d\n", pthread_self(), i);
730 }
731
732 if (my_id == 0) {
733 /* Give the worker threads undisturbed time to finish before waiting on them */
734 if (g_do_sleep) {
735 usleep(g_iteration_sleeptime_us);
736 }
737
738 /* Wait for the worker threads to finish */
739 for (uint32_t i = 0; i < g_numthreads - 1; i++) {
740 kr = semaphore_wait(g_readysem);
741 mach_assert_zero_t(my_id, kr);
742 }
743
744 /* Tell everyone and the main thread that the last iteration is done */
745 debug_log("%d Leader thread done\n", g_iterations - 1);
746
747 for (int cpuid = 0; cpuid < g_numcpus; cpuid++) {
748 if (g_cpu_histogram[cpuid].current == 1) {
749 atomic_fetch_or_explicit(&g_cpu_map[g_iterations - 1], (1UL << cpuid), memory_order_relaxed);
750 g_cpu_histogram[cpuid].current = 0;
751 }
752 }
753
754 kr = semaphore_signal_all(g_main_sem);
755 mach_assert_zero_t(my_id, kr);
756 } else {
757 /* Hold up thread teardown so it doesn't affect the last iteration */
758 kr = semaphore_wait_signal(g_main_sem, g_readysem);
759 mach_assert_zero_t(my_id, kr);
760 }
761
762 time_value_t runtime = get_thread_runtime();
763 os_unfair_lock_lock(&runtime_lock);
764 time_value_add(&worker_threads_total_runtime, &runtime);
765 os_unfair_lock_unlock(&runtime_lock);
766
767 return 0;
768 }
769
770 /*
771 * Given an array of uint64_t values, compute average, max, min, and standard deviation
772 */
773 static void
774 compute_stats(uint64_t *values, uint64_t count, float *averagep, uint64_t *maxp, uint64_t *minp, float *stddevp)
775 {
776 uint32_t i;
777 uint64_t _sum = 0;
778 uint64_t _max = 0;
779 uint64_t _min = UINT64_MAX;
780 float _avg = 0;
781 float _dev = 0;
782
783 for (i = 0; i < count; i++) {
784 _sum += values[i];
785 _max = values[i] > _max ? values[i] : _max;
786 _min = values[i] < _min ? values[i] : _min;
787 }
788
789 _avg = ((float)_sum) / ((float)count);
790
791 _dev = 0;
792 for (i = 0; i < count; i++) {
793 _dev += powf((((float)values[i]) - _avg), 2);
794 }
795
796 _dev /= count;
797 _dev = sqrtf(_dev);
798
799 *averagep = _avg;
800 *maxp = _max;
801 *minp = _min;
802 *stddevp = _dev;
803 }
804
805 typedef struct {
806 natural_t sys;
807 natural_t user;
808 natural_t idle;
809 } cpu_time_t;
810
811 void
812 record_cpu_time(cpu_time_t *cpu_time)
813 {
814 host_cpu_load_info_data_t load;
815 mach_msg_type_number_t count = HOST_CPU_LOAD_INFO_COUNT;
816 kern_return_t kr = host_statistics(mach_host_self(), HOST_CPU_LOAD_INFO, (int *)&load, &count);
817 mach_assert_zero_t(0, kr);
818
819 natural_t total_system_time = load.cpu_ticks[CPU_STATE_SYSTEM];
820 natural_t total_user_time = load.cpu_ticks[CPU_STATE_USER] + load.cpu_ticks[CPU_STATE_NICE];
821 natural_t total_idle_time = load.cpu_ticks[CPU_STATE_IDLE];
822
823 cpu_time->sys = total_system_time;
824 cpu_time->user = total_user_time;
825 cpu_time->idle = total_idle_time;
826 }
827
828 int
829 main(int argc, char **argv)
830 {
831 errno_t ret;
832 kern_return_t kr;
833
834 pthread_t *threads;
835 uint64_t *worst_latencies_ns;
836 uint64_t *worst_latencies_from_first_ns;
837 uint64_t max, min;
838 float avg, stddev;
839
840 bool test_fail = false;
841 bool test_warn = false;
842
843 for (int i = 0; i < argc; i++) {
844 if (strcmp(argv[i], "--switched_apptype") == 0) {
845 g_seen_apptype = TRUE;
846 }
847 }
848
849 if (!g_seen_apptype) {
850 selfexec_with_apptype(argc, argv);
851 }
852
853 parse_args(argc, argv);
854
855 srand((unsigned int)time(NULL));
856
857 mach_timebase_info(&g_mti);
858
859 #if TARGET_OS_OSX
860 /* SKIP test if running on arm platform */
861 if (g_run_on_intel_only) {
862 int is_arm = 0;
863 size_t is_arm_size = sizeof(is_arm);
864 ret = sysctlbyname("hw.optional.arm64", &is_arm, &is_arm_size, NULL, 0);
865 if (ret == 0 && is_arm) {
866 printf("Unsupported platform. Skipping test.\n");
867 exit(0);
868 }
869 }
870 #endif /* TARGET_OS_OSX */
871
872 size_t ncpu_size = sizeof(g_numcpus);
873 ret = sysctlbyname("hw.ncpu", &g_numcpus, &ncpu_size, NULL, 0);
874 if (ret) {
875 err(EX_OSERR, "Failed sysctlbyname(hw.ncpu)");
876 }
877 assert(g_numcpus <= 64); /* g_cpu_map needs to be extended for > 64 cpus */
878
879 size_t physicalcpu_size = sizeof(g_nphysicalcpu);
880 ret = sysctlbyname("hw.physicalcpu", &g_nphysicalcpu, &physicalcpu_size, NULL, 0);
881 if (ret) {
882 err(EX_OSERR, "Failed sysctlbyname(hw.physicalcpu)");
883 }
884
885 size_t logicalcpu_size = sizeof(g_nlogicalcpu);
886 ret = sysctlbyname("hw.logicalcpu", &g_nlogicalcpu, &logicalcpu_size, NULL, 0);
887 if (ret) {
888 err(EX_OSERR, "Failed sysctlbyname(hw.logicalcpu)");
889 }
890
891 if (g_test_rt) {
892 if (g_numthreads == 0) {
893 g_numthreads = g_numcpus;
894 }
895 g_policy = MY_POLICY_REALTIME;
896 g_do_all_spin = TRUE;
897 g_histogram = true;
898 /* Don't change g_traceworthy_latency_ns if it's explicity been set to something other than the default */
899 if (g_traceworthy_latency_ns == TRACEWORTHY_NANOS) {
900 g_traceworthy_latency_ns = TRACEWORTHY_NANOS_TEST;
901 }
902 } else if (g_test_rt_smt) {
903 if (g_nlogicalcpu != 2 * g_nphysicalcpu) {
904 /* Not SMT */
905 printf("Attempt to run --test-rt-smt on a non-SMT device\n");
906 exit(0);
907 }
908
909 if (g_numthreads == 0) {
910 g_numthreads = g_nphysicalcpu;
911 }
912 g_policy = MY_POLICY_REALTIME;
913 g_do_all_spin = TRUE;
914 g_histogram = true;
915 } else if (g_test_rt_avoid0) {
916 #if defined(__x86_64__) || defined(__i386__)
917 if (g_numthreads == 0) {
918 g_numthreads = g_nphysicalcpu - 1;
919 }
920 if (g_numthreads == 0) {
921 printf("Attempt to run --test-rt-avoid0 on a uniprocessor\n");
922 exit(0);
923 }
924 g_policy = MY_POLICY_REALTIME;
925 g_do_all_spin = TRUE;
926 g_histogram = true;
927 #else
928 printf("Attempt to run --test-rt-avoid0 on a non-Intel device\n");
929 exit(0);
930 #endif
931 } else if (g_numthreads == 0) {
932 g_numthreads = g_numcpus;
933 }
934
935 if (g_do_each_spin) {
936 g_each_spin_duration_abs = nanos_to_abs(g_each_spin_duration_ns);
937 }
938
939 /* Configure the long-spin thread to take up half of its computation */
940 if (g_do_one_long_spin) {
941 g_one_long_spin_length_ns = COMPUTATION_NANOS / 2;
942 g_one_long_spin_length_abs = nanos_to_abs(g_one_long_spin_length_ns);
943 }
944
945 /* Estimate the amount of time the cleanup phase needs to back off */
946 g_iteration_sleeptime_us = g_numthreads * 20;
947
948 uint32_t threads_per_core = (g_numthreads / g_numcpus) + 1;
949 if (g_do_each_spin) {
950 g_iteration_sleeptime_us += threads_per_core * (g_each_spin_duration_ns / NSEC_PER_USEC);
951 }
952 if (g_do_one_long_spin) {
953 g_iteration_sleeptime_us += g_one_long_spin_length_ns / NSEC_PER_USEC;
954 }
955
956 /* Arrays for threads and their wakeup times */
957 threads = (pthread_t*) valloc(sizeof(pthread_t) * g_numthreads);
958 assert(threads);
959
960 size_t endtimes_size = sizeof(uint64_t) * g_numthreads;
961
962 g_thread_endtimes_abs = (uint64_t*) valloc(endtimes_size);
963 assert(g_thread_endtimes_abs);
964
965 /* Ensure the allocation is pre-faulted */
966 ret = memset_s(g_thread_endtimes_abs, endtimes_size, 0, endtimes_size);
967 if (ret) {
968 errc(EX_OSERR, ret, "memset_s endtimes");
969 }
970
971 size_t latencies_size = sizeof(uint64_t) * g_iterations;
972
973 worst_latencies_ns = (uint64_t*) valloc(latencies_size);
974 assert(worst_latencies_ns);
975
976 /* Ensure the allocation is pre-faulted */
977 ret = memset_s(worst_latencies_ns, latencies_size, 0, latencies_size);
978 if (ret) {
979 errc(EX_OSERR, ret, "memset_s latencies");
980 }
981
982 worst_latencies_from_first_ns = (uint64_t*) valloc(latencies_size);
983 assert(worst_latencies_from_first_ns);
984
985 /* Ensure the allocation is pre-faulted */
986 ret = memset_s(worst_latencies_from_first_ns, latencies_size, 0, latencies_size);
987 if (ret) {
988 errc(EX_OSERR, ret, "memset_s latencies_from_first");
989 }
990
991 size_t histogram_size = sizeof(histogram_t) * g_numcpus;
992 g_cpu_histogram = (histogram_t *)valloc(histogram_size);
993 assert(g_cpu_histogram);
994 /* Ensure the allocation is pre-faulted */
995 ret = memset_s(g_cpu_histogram, histogram_size, 0, histogram_size);
996 if (ret) {
997 errc(EX_OSERR, ret, "memset_s g_cpu_histogram");
998 }
999
1000 size_t map_size = sizeof(uint64_t) * g_iterations;
1001 g_cpu_map = (_Atomic uint64_t *)valloc(map_size);
1002 assert(g_cpu_map);
1003 /* Ensure the allocation is pre-faulted */
1004 ret = memset_s(g_cpu_map, map_size, 0, map_size);
1005 if (ret) {
1006 errc(EX_OSERR, ret, "memset_s g_cpu_map");
1007 }
1008
1009 kr = semaphore_create(mach_task_self(), &g_main_sem, SYNC_POLICY_FIFO, 0);
1010 mach_assert_zero(kr);
1011
1012 /* Either one big semaphore or one per thread */
1013 if (g_waketype == WAKE_CHAIN ||
1014 g_waketype == WAKE_BROADCAST_PERTHREAD ||
1015 g_waketype == WAKE_HOP) {
1016 g_semarr = valloc(sizeof(semaphore_t) * g_numthreads);
1017 assert(g_semarr);
1018
1019 for (uint32_t i = 0; i < g_numthreads; i++) {
1020 kr = semaphore_create(mach_task_self(), &g_semarr[i], SYNC_POLICY_FIFO, 0);
1021 mach_assert_zero(kr);
1022 }
1023
1024 g_leadersem = g_semarr[0];
1025 } else {
1026 kr = semaphore_create(mach_task_self(), &g_broadcastsem, SYNC_POLICY_FIFO, 0);
1027 mach_assert_zero(kr);
1028 kr = semaphore_create(mach_task_self(), &g_leadersem, SYNC_POLICY_FIFO, 0);
1029 mach_assert_zero(kr);
1030 }
1031
1032 if (g_waketype == WAKE_HOP) {
1033 kr = semaphore_create(mach_task_self(), &g_donesem, SYNC_POLICY_FIFO, 0);
1034 mach_assert_zero(kr);
1035 }
1036
1037 kr = semaphore_create(mach_task_self(), &g_readysem, SYNC_POLICY_FIFO, 0);
1038 mach_assert_zero(kr);
1039
1040 kr = semaphore_create(mach_task_self(), &g_rt_churn_sem, SYNC_POLICY_FIFO, 0);
1041 mach_assert_zero(kr);
1042
1043 kr = semaphore_create(mach_task_self(), &g_rt_churn_start_sem, SYNC_POLICY_FIFO, 0);
1044 mach_assert_zero(kr);
1045
1046 atomic_store_explicit(&g_done_threads, 0, memory_order_relaxed);
1047
1048 /* Create the threads */
1049 for (uint32_t i = 0; i < g_numthreads; i++) {
1050 ret = pthread_create(&threads[i], NULL, worker_thread, (void*)(uintptr_t)i);
1051 if (ret) {
1052 errc(EX_OSERR, ret, "pthread_create %d", i);
1053 }
1054 }
1055
1056 ret = setpriority(PRIO_DARWIN_ROLE, 0, PRIO_DARWIN_ROLE_UI_FOCAL);
1057 if (ret) {
1058 errc(EX_OSERR, ret, "setpriority");
1059 }
1060
1061 thread_setup(0);
1062
1063 g_starttime_abs = mach_absolute_time();
1064
1065 if (g_churn_pri) {
1066 create_churn_threads();
1067 }
1068 if (g_rt_churn) {
1069 create_rt_churn_threads();
1070 }
1071
1072 /* Let everyone get settled */
1073 kr = semaphore_wait(g_main_sem);
1074 mach_assert_zero(kr);
1075
1076 /* Give the system a bit more time to settle */
1077 if (g_do_sleep) {
1078 usleep(g_iteration_sleeptime_us);
1079 }
1080
1081 cpu_time_t start_time;
1082 cpu_time_t finish_time;
1083
1084 record_cpu_time(&start_time);
1085
1086 /* Go! */
1087 for (uint32_t i = 0; i < g_iterations; i++) {
1088 uint32_t j;
1089 uint64_t worst_abs = 0, best_abs = UINT64_MAX;
1090
1091 if (g_do_one_long_spin) {
1092 g_one_long_spin_id = (uint32_t)rand() % g_numthreads;
1093 }
1094
1095 if (g_rt_churn) {
1096 start_rt_churn_threads();
1097 usleep(100);
1098 }
1099
1100 debug_log("%d Main thread reset\n", i);
1101
1102 atomic_store_explicit(&g_done_threads, 0, memory_order_seq_cst);
1103
1104 g_starttime_abs = mach_absolute_time();
1105
1106 /* Fire them off and wait for worker threads to finish */
1107 kr = semaphore_wait_signal(g_main_sem, g_leadersem);
1108 mach_assert_zero(kr);
1109
1110 debug_log("%d Main thread return\n", i);
1111
1112 assert(atomic_load_explicit(&g_done_threads, memory_order_relaxed) == g_numthreads);
1113
1114 if (g_rt_churn) {
1115 wait_for_rt_churn_threads();
1116 }
1117
1118 /*
1119 * We report the worst latencies relative to start time
1120 * and relative to the lead worker thread.
1121 */
1122 for (j = 0; j < g_numthreads; j++) {
1123 uint64_t latency_abs;
1124
1125 latency_abs = g_thread_endtimes_abs[j] - g_starttime_abs;
1126 worst_abs = worst_abs < latency_abs ? latency_abs : worst_abs;
1127 }
1128
1129 worst_latencies_ns[i] = abs_to_nanos(worst_abs);
1130
1131 worst_abs = 0;
1132 for (j = 1; j < g_numthreads; j++) {
1133 uint64_t latency_abs;
1134
1135 latency_abs = g_thread_endtimes_abs[j] - g_thread_endtimes_abs[0];
1136 worst_abs = worst_abs < latency_abs ? latency_abs : worst_abs;
1137 best_abs = best_abs > latency_abs ? latency_abs : best_abs;
1138 }
1139
1140 worst_latencies_from_first_ns[i] = abs_to_nanos(worst_abs);
1141
1142 /*
1143 * In the event of a bad run, cut a trace point.
1144 */
1145 if (worst_latencies_from_first_ns[i] > g_traceworthy_latency_ns) {
1146 /* Ariadne's ad-hoc test signpost */
1147 kdebug_trace(ARIADNEDBG_CODE(0, 0), worst_latencies_from_first_ns[i], g_traceworthy_latency_ns, 0, 0);
1148
1149 if (g_verbose) {
1150 printf("Worst on this round was %.2f us.\n", ((float)worst_latencies_from_first_ns[i]) / 1000.0);
1151 }
1152 }
1153
1154 /* Give the system a bit more time to settle */
1155 if (g_do_sleep) {
1156 usleep(g_iteration_sleeptime_us);
1157 }
1158 }
1159
1160 record_cpu_time(&finish_time);
1161
1162 /* Rejoin threads */
1163 for (uint32_t i = 0; i < g_numthreads; i++) {
1164 ret = pthread_join(threads[i], NULL);
1165 if (ret) {
1166 errc(EX_OSERR, ret, "pthread_join %d", i);
1167 }
1168 }
1169
1170 if (g_rt_churn) {
1171 join_rt_churn_threads();
1172 }
1173
1174 if (g_churn_pri) {
1175 join_churn_threads();
1176 }
1177
1178 uint32_t cpu_idle_time = (finish_time.idle - start_time.idle) * 10;
1179 uint32_t worker_threads_runtime = worker_threads_total_runtime.seconds * 1000 + worker_threads_total_runtime.microseconds / 1000;
1180
1181 compute_stats(worst_latencies_ns, g_iterations, &avg, &max, &min, &stddev);
1182 printf("Results (from a stop):\n");
1183 printf("Max:\t\t%.2f us\n", ((float)max) / 1000.0);
1184 printf("Min:\t\t%.2f us\n", ((float)min) / 1000.0);
1185 printf("Avg:\t\t%.2f us\n", avg / 1000.0);
1186 printf("Stddev:\t\t%.2f us\n", stddev / 1000.0);
1187
1188 putchar('\n');
1189
1190 compute_stats(worst_latencies_from_first_ns, g_iterations, &avg, &max, &min, &stddev);
1191 printf("Results (relative to first thread):\n");
1192 printf("Max:\t\t%.2f us\n", ((float)max) / 1000.0);
1193 printf("Min:\t\t%.2f us\n", ((float)min) / 1000.0);
1194 printf("Avg:\t\t%.2f us\n", avg / 1000.0);
1195 printf("Stddev:\t\t%.2f us\n", stddev / 1000.0);
1196
1197 #if 0
1198 for (uint32_t i = 0; i < g_iterations; i++) {
1199 printf("Iteration %d: %f us\n", i, worst_latencies_ns[i] / 1000.0);
1200 }
1201 #endif
1202
1203 if (g_histogram) {
1204 putchar('\n');
1205
1206 for (uint32_t i = 0; i < g_numcpus; i++) {
1207 printf("%d\t%d\n", i, g_cpu_histogram[i].accum);
1208 }
1209 }
1210
1211 if (g_test_rt || g_test_rt_smt || g_test_rt_avoid0) {
1212 #define PRIMARY 0x5555555555555555ULL
1213 #define SECONDARY 0xaaaaaaaaaaaaaaaaULL
1214
1215 int fail_count = 0;
1216
1217 for (uint32_t i = 0; i < g_iterations; i++) {
1218 bool secondary = false;
1219 bool fail = false;
1220 uint64_t map = g_cpu_map[i];
1221 if (g_test_rt_smt) {
1222 /* Test for one or more threads running on secondary cores unexpectedly (WARNING) */
1223 secondary = (map & SECONDARY);
1224 /* Test for threads running on both primary and secondary cpus of the same core (FAIL) */
1225 fail = ((map & PRIMARY) & ((map & SECONDARY) >> 1));
1226 } else if (g_test_rt) {
1227 fail = (__builtin_popcountll(map) != g_numthreads) && (worst_latencies_ns[i] > g_traceworthy_latency_ns);
1228 } else if (g_test_rt_avoid0) {
1229 fail = ((map & 0x1) == 0x1);
1230 }
1231 if (secondary || fail) {
1232 printf("Iteration %d: 0x%llx%s%s\n", i, map,
1233 secondary ? " SECONDARY" : "",
1234 fail ? " FAIL" : "");
1235 }
1236 test_warn |= (secondary || fail);
1237 test_fail |= fail;
1238 fail_count += fail;
1239 }
1240
1241 if (test_fail && (g_iterations >= 100) && (fail_count <= g_iterations / 100)) {
1242 printf("99%% or better success rate\n");
1243 test_fail = 0;
1244 }
1245 }
1246
1247 if (g_test_rt_smt && (g_each_spin_duration_ns >= 200000) && !test_warn) {
1248 printf("cpu_idle_time=%dms worker_threads_runtime=%dms\n", cpu_idle_time, worker_threads_runtime);
1249 if (cpu_idle_time < worker_threads_runtime / 4) {
1250 printf("FAIL cpu_idle_time unexpectedly small\n");
1251 test_fail = 1;
1252 } else if (cpu_idle_time > worker_threads_runtime * 2) {
1253 printf("FAIL cpu_idle_time unexpectedly large\n");
1254 test_fail = 1;
1255 }
1256 }
1257
1258 free(threads);
1259 free(g_thread_endtimes_abs);
1260 free(worst_latencies_ns);
1261 free(worst_latencies_from_first_ns);
1262 free(g_cpu_histogram);
1263 free(g_cpu_map);
1264
1265 return test_fail;
1266 }
1267
1268 /*
1269 * WARNING: This is SPI specifically intended for use by launchd to start UI
1270 * apps. We use it here for a test tool only to opt into QoS using the same
1271 * policies. Do not use this outside xnu or libxpc/launchd.
1272 */
1273 static void
1274 selfexec_with_apptype(int argc, char *argv[])
1275 {
1276 int ret;
1277 posix_spawnattr_t attr;
1278 extern char **environ;
1279 char *new_argv[argc + 1 + 1 /* NULL */];
1280 int i;
1281 char prog[PATH_MAX];
1282 uint32_t prog_size = PATH_MAX;
1283
1284 ret = _NSGetExecutablePath(prog, &prog_size);
1285 if (ret) {
1286 err(EX_OSERR, "_NSGetExecutablePath");
1287 }
1288
1289 for (i = 0; i < argc; i++) {
1290 new_argv[i] = argv[i];
1291 }
1292
1293 new_argv[i] = "--switched_apptype";
1294 new_argv[i + 1] = NULL;
1295
1296 ret = posix_spawnattr_init(&attr);
1297 if (ret) {
1298 errc(EX_OSERR, ret, "posix_spawnattr_init");
1299 }
1300
1301 ret = posix_spawnattr_setflags(&attr, POSIX_SPAWN_SETEXEC);
1302 if (ret) {
1303 errc(EX_OSERR, ret, "posix_spawnattr_setflags");
1304 }
1305
1306 ret = posix_spawnattr_setprocesstype_np(&attr, POSIX_SPAWN_PROC_TYPE_APP_DEFAULT);
1307 if (ret) {
1308 errc(EX_OSERR, ret, "posix_spawnattr_setprocesstype_np");
1309 }
1310
1311 ret = posix_spawn(NULL, prog, NULL, &attr, new_argv, environ);
1312 if (ret) {
1313 errc(EX_OSERR, ret, "posix_spawn");
1314 }
1315 }
1316
1317 /*
1318 * Admittedly not very attractive.
1319 */
1320 static void __attribute__((noreturn))
1321 usage()
1322 {
1323 errx(EX_USAGE, "Usage: %s <threads> <chain | hop | broadcast-single-sem | broadcast-per-thread> "
1324 "<realtime | timeshare | timeshare_no_smt | fixed> <iterations>\n\t\t"
1325 "[--trace <traceworthy latency in ns>] "
1326 "[--verbose] [--spin-one] [--spin-all] [--spin-time <nanos>] [--affinity]\n\t\t"
1327 "[--no-sleep] [--drop-priority] [--churn-pri <pri>] [--churn-count <n>]\n\t\t"
1328 "[--rt-churn] [--rt-churn-count <n>] [--rt-ll] [--test-rt] [--test-rt-smt] [--test-rt-avoid0]",
1329 getprogname());
1330 }
1331
1332 static struct option* g_longopts;
1333 static int option_index;
1334
1335 static uint32_t
1336 read_dec_arg()
1337 {
1338 char *cp;
1339 /* char* optarg is a magic global */
1340
1341 uint32_t arg_val = (uint32_t)strtoull(optarg, &cp, 10);
1342
1343 if (cp == optarg || *cp) {
1344 errx(EX_USAGE, "arg --%s requires a decimal number, found \"%s\"",
1345 g_longopts[option_index].name, optarg);
1346 }
1347
1348 return arg_val;
1349 }
1350
1351 static void
1352 parse_args(int argc, char *argv[])
1353 {
1354 enum {
1355 OPT_GETOPT = 0,
1356 OPT_SPIN_TIME,
1357 OPT_TRACE,
1358 OPT_PRIORITY,
1359 OPT_CHURN_PRI,
1360 OPT_CHURN_COUNT,
1361 OPT_RT_CHURN_COUNT,
1362 };
1363
1364 static struct option longopts[] = {
1365 /* BEGIN IGNORE CODESTYLE */
1366 { "spin-time", required_argument, NULL, OPT_SPIN_TIME },
1367 { "trace", required_argument, NULL, OPT_TRACE },
1368 { "priority", required_argument, NULL, OPT_PRIORITY },
1369 { "churn-pri", required_argument, NULL, OPT_CHURN_PRI },
1370 { "churn-count", required_argument, NULL, OPT_CHURN_COUNT },
1371 { "rt-churn-count", required_argument, NULL, OPT_RT_CHURN_COUNT },
1372 { "switched_apptype", no_argument, (int*)&g_seen_apptype, TRUE },
1373 { "spin-one", no_argument, (int*)&g_do_one_long_spin, TRUE },
1374 { "intel-only", no_argument, (int*)&g_run_on_intel_only, TRUE },
1375 { "spin-all", no_argument, (int*)&g_do_all_spin, TRUE },
1376 { "affinity", no_argument, (int*)&g_do_affinity, TRUE },
1377 { "no-sleep", no_argument, (int*)&g_do_sleep, FALSE },
1378 { "drop-priority", no_argument, (int*)&g_drop_priority, TRUE },
1379 { "test-rt", no_argument, (int*)&g_test_rt, TRUE },
1380 { "test-rt-smt", no_argument, (int*)&g_test_rt_smt, TRUE },
1381 { "test-rt-avoid0", no_argument, (int*)&g_test_rt_avoid0, TRUE },
1382 { "rt-churn", no_argument, (int*)&g_rt_churn, TRUE },
1383 { "rt-ll", no_argument, (int*)&g_rt_ll, TRUE },
1384 { "histogram", no_argument, (int*)&g_histogram, TRUE },
1385 { "verbose", no_argument, (int*)&g_verbose, TRUE },
1386 { "help", no_argument, NULL, 'h' },
1387 { NULL, 0, NULL, 0 }
1388 /* END IGNORE CODESTYLE */
1389 };
1390
1391 g_longopts = longopts;
1392 int ch = 0;
1393
1394 while ((ch = getopt_long(argc, argv, "h", longopts, &option_index)) != -1) {
1395 switch (ch) {
1396 case OPT_GETOPT:
1397 /* getopt_long set a variable */
1398 break;
1399 case OPT_SPIN_TIME:
1400 g_do_each_spin = TRUE;
1401 g_each_spin_duration_ns = read_dec_arg();
1402 break;
1403 case OPT_TRACE:
1404 g_traceworthy_latency_ns = read_dec_arg();
1405 break;
1406 case OPT_PRIORITY:
1407 g_priority = read_dec_arg();
1408 break;
1409 case OPT_CHURN_PRI:
1410 g_churn_pri = read_dec_arg();
1411 break;
1412 case OPT_CHURN_COUNT:
1413 g_churn_count = read_dec_arg();
1414 break;
1415 case OPT_RT_CHURN_COUNT:
1416 g_rt_churn_count = read_dec_arg();
1417 break;
1418 case '?':
1419 case 'h':
1420 default:
1421 usage();
1422 /* NORETURN */
1423 }
1424 }
1425
1426 /*
1427 * getopt_long reorders all the options to the beginning of the argv array.
1428 * Jump past them to the non-option arguments.
1429 */
1430
1431 argc -= optind;
1432 argv += optind;
1433
1434 if (argc > 4) {
1435 warnx("Too many non-option arguments passed");
1436 usage();
1437 }
1438
1439 if (argc != 4) {
1440 warnx("Missing required <threads> <waketype> <policy> <iterations> arguments");
1441 usage();
1442 }
1443
1444 char *cp;
1445
1446 /* How many threads? */
1447 g_numthreads = (uint32_t)strtoull(argv[0], &cp, 10);
1448
1449 if (cp == argv[0] || *cp) {
1450 errx(EX_USAGE, "numthreads requires a decimal number, found \"%s\"", argv[0]);
1451 }
1452
1453 /* What wakeup pattern? */
1454 g_waketype = parse_wakeup_pattern(argv[1]);
1455
1456 /* Policy */
1457 g_policy = parse_thread_policy(argv[2]);
1458
1459 /* Iterations */
1460 g_iterations = (uint32_t)strtoull(argv[3], &cp, 10);
1461
1462 if (cp == argv[3] || *cp) {
1463 errx(EX_USAGE, "numthreads requires a decimal number, found \"%s\"", argv[3]);
1464 }
1465
1466 if (g_iterations < 1) {
1467 errx(EX_USAGE, "Must have at least one iteration");
1468 }
1469
1470 if (g_numthreads == 1 && g_waketype == WAKE_CHAIN) {
1471 errx(EX_USAGE, "chain mode requires more than one thread");
1472 }
1473
1474 if (g_numthreads == 1 && g_waketype == WAKE_HOP) {
1475 errx(EX_USAGE, "hop mode requires more than one thread");
1476 }
1477 }