xref: /xnu-10002.81.5/tests/setitimer.c (revision 5e3eaea39dcf651e66cb99ba7d70e32cc4a99587)
1 #include <darwintest.h>
2 
3 #include <assert.h>
4 #include <mach/clock_types.h>
5 #include <unistd.h>
6 #include <stdlib.h>
7 #include <stdio.h>
8 #include <errno.h>
9 #include <err.h>
10 #include <sys/time.h>
11 #include <mach/mach.h>
12 #include <mach/mach_time.h>
13 #include <pthread.h>
14 #include <sys/sysctl.h>
15 #include <sys/stat.h>
16 #include <sys/mount.h>
17 #include <stdbool.h>
18 #include <signal.h>
19 #include <sys/resource.h>
20 #include <sys/resource_private.h>
21 #include <os/atomic_private.h>
22 #include <libproc.h>
23 #include <TargetConditionals.h>
24 
25 #if __has_include(<mach/mach_time_private.h>)
26 #include <mach/mach_time_private.h>
27 #else
28 kern_return_t           mach_get_times(uint64_t* absolute_time,
29     uint64_t* continuous_time,
30     struct timespec *tp);
31 #endif
32 
33 /*
34  * This test program creates up to 8 worker threads performing
35  * mixed workloads of system calls (which contribute to both
36  * user and system time), as well as spins in userspace (which
37  * only contribute to user time).
38  *
39  * setitimer(2) is used to program timers that fire signals
40  * after various thresholds. The signal handler detects
41  * which thread the signal was delivered on by matching the
42  * stack pointer to ranges for each thread.
43  *
44  * After the test scenario is complete, the distribution of
45  * threads which received interrupts is evaluated to match
46  * expected heuristics.
47  */
48 
49 T_GLOBAL_META(
50 	T_META_RUN_CONCURRENTLY(false),
51 	T_META_CHECK_LEAKS(false),
52 	T_META_ALL_VALID_ARCHS(true),
53 	T_META_RADAR_COMPONENT_NAME("xnu"),
54 	T_META_RADAR_COMPONENT_VERSION("scheduler"),
55 	T_META_OWNER("chimene"),
56 	T_META_ENABLED(TARGET_OS_OSX)
57 	);
58 
59 static void *stat_thread(void *arg);
60 static void *statfs_thread(void *arg);
61 
62 static void alrm_handler(int, struct __siginfo *, void *);
63 
64 static semaphore_t gMainWaitForWorkers;
65 static semaphore_t gWorkersStart;
66 
67 static pthread_mutex_t gShouldExitMutex = PTHREAD_MUTEX_INITIALIZER;
68 static pthread_cond_t  gShouldExitCondition = PTHREAD_COND_INITIALIZER;
69 
70 static _Atomic bool gShouldExit = false;
71 
72 static const uint32_t max_threads = 9;
73 
74 static struct threadentry {
75 	pthread_t thread;
76 	uint64_t tid;
77 	void* stack_addr;
78 	size_t stack_size;
79 	bool expect_cpu_usage;
80 	uint32_t alrm_count;
81 	uint32_t vtalrm_count;
82 	uint32_t prof_count;
83 	uint32_t xcpu_count;
84 	struct thsc_time_cpi self_stats;
85 } __attribute__((aligned(128))) gThreadList[max_threads];
86 
87 static uint32_t nworkers;
88 static uint32_t nthreads;
89 
90 static double offcore_time_percent_threshold = 75.0;
91 
92 static bool is_rosetta = false;
93 
94 static mach_timebase_info_data_t timebase_info;
95 
96 static uint64_t
abs_to_nanos(uint64_t abs)97 abs_to_nanos(uint64_t abs)
98 {
99 	return abs * timebase_info.numer / timebase_info.denom;
100 }
101 
102 /* Some statistics APIs return host abstime instead of Rosetta-translated abstime */
103 static uint64_t
abs_to_nanos_host(uint64_t abstime)104 abs_to_nanos_host(uint64_t abstime)
105 {
106 	if (is_rosetta) {
107 		return abstime * 125 / 3;
108 	} else {
109 		return abs_to_nanos(abstime);
110 	}
111 }
112 
113 static int
processIsTranslated(void)114 processIsTranslated(void)
115 {
116 	int ret = 0;
117 	size_t size = sizeof(ret);
118 	if (sysctlbyname("sysctl.proc_translated", &ret, &size, NULL, 0) == -1) {
119 		if (errno == ENOENT) {
120 			return 0;
121 		} else {
122 			return -1;
123 		}
124 	}
125 	return ret;
126 }
127 
128 static void
fill_thread_stats(uint32_t i)129 fill_thread_stats(uint32_t i)
130 {
131 	struct threadentry *entry = &gThreadList[i];
132 
133 	int rv = thread_selfcounts(THSC_TIME_CPI, &entry->self_stats, sizeof(entry->self_stats));
134 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "thread_selfcounts(THSC_TIME_CPI)");
135 }
136 
137 T_DECL(setitimer,
138     "Test various setitimer delivered signals to CPU-burning threads")
139 {
140 	int rv;
141 	kern_return_t kr;
142 	uint32_t ncpu;
143 	size_t ncpu_size = sizeof(ncpu);
144 
145 	struct sched_param self_param = {.sched_priority = 47};
146 
147 	rv = pthread_setschedparam(pthread_self(), SCHED_FIFO, &self_param);
148 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_setschedparam");
149 
150 	kr = mach_timebase_info(&timebase_info);
151 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "mach_timebase_info");
152 
153 	is_rosetta = processIsTranslated();
154 
155 	rv = sysctlbyname("hw.ncpu", &ncpu, &ncpu_size, NULL, 0);
156 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "sysctlbyname(hw.ncpu)");
157 
158 	if (ncpu < 2) {
159 		T_SKIP("%d CPUs not supported for test, returning success", ncpu);
160 	}
161 
162 	nworkers = MIN(max_threads - 1, ncpu);
163 	nthreads = nworkers + 1;
164 
165 	T_LOG("rosetta = %d\n", is_rosetta);
166 	T_LOG("hw.ncpu = %d\n", ncpu);
167 	T_LOG("nworkers = %d\n", nworkers);
168 	T_LOG("nthreads = %d\n", nthreads);
169 
170 	kr = semaphore_create(mach_task_self(), &gMainWaitForWorkers, SYNC_POLICY_FIFO, 0);
171 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "semaphore_create()");
172 
173 	kr = semaphore_create(mach_task_self(), &gWorkersStart, SYNC_POLICY_FIFO, 0);
174 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "semaphore_create()");
175 
176 	pthread_attr_t attr;
177 
178 	rv = pthread_attr_init(&attr);
179 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_init");
180 
181 	struct sched_param child_param = {.sched_priority = 37};
182 
183 	rv = pthread_attr_setschedparam(&attr, &child_param);
184 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_set_qos_class_np");
185 
186 	for (uint32_t i = 0; i < nthreads; i++) {
187 		if (i == 0) {
188 			gThreadList[i].thread = pthread_self();
189 		} else {
190 			rv = pthread_create(&gThreadList[i].thread, &attr,
191 			    i % 2 ? stat_thread : statfs_thread,
192 			    (void *)(uintptr_t)i);
193 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_create");
194 			gThreadList[i].expect_cpu_usage = i % 2 == 0 ? true : false;
195 		}
196 
197 		rv = pthread_threadid_np(gThreadList[i].thread, &gThreadList[i].tid);
198 		T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_threadid_np");
199 
200 		gThreadList[i].stack_addr = pthread_get_stackaddr_np(gThreadList[i].thread);
201 		gThreadList[i].stack_size = pthread_get_stacksize_np(gThreadList[i].thread);
202 	}
203 
204 	rv = pthread_attr_destroy(&attr);
205 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_attr_destroy");
206 
207 	for (uint32_t i = 1; i < nthreads; i++) {
208 		kr = semaphore_wait(gMainWaitForWorkers);
209 		T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "semaphore_wait()");
210 	}
211 
212 	for (uint32_t i = 0; i < nthreads; i++) {
213 		T_LOG("Thread %p (0x%llx) checked in, stack %p/%p\n",
214 		    (void*)gThreadList[i].thread,
215 		    gThreadList[i].tid,
216 		    gThreadList[i].stack_addr,
217 		    (void *)gThreadList[i].stack_size);
218 	}
219 
220 	sigset_t sigmk;
221 	sigemptyset(&sigmk);
222 
223 	struct sigaction sigact = {
224 		.sa_sigaction = alrm_handler,
225 		.sa_mask = sigmk,
226 		.sa_flags = SA_SIGINFO,
227 	};
228 
229 	rv = sigaction(SIGALRM, &sigact, NULL);
230 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "sigaction(SIGALRM)");
231 
232 	rv = sigaction(SIGVTALRM, &sigact, NULL);
233 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "sigaction(SIGVTALRM)");
234 
235 	rv = sigaction(SIGPROF, &sigact, NULL);
236 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "sigaction(SIGPROF)");
237 
238 	rv = sigaction(SIGXCPU, &sigact, NULL);
239 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "sigaction(SIGXCPU)");
240 
241 	struct itimerval itime = {
242 		.it_interval.tv_sec = 0,
243 		.it_interval.tv_usec = 10000,
244 		.it_value.tv_sec = 0,
245 		.it_value.tv_usec = 10,  /* immediately */
246 	};
247 
248 	rv = setitimer(ITIMER_REAL, &itime, NULL);
249 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setitimer(ITIMER_REAL)");
250 
251 	rv = setitimer(ITIMER_VIRTUAL, &itime, NULL);
252 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setitimer(ITIMER_REAL)");
253 
254 	rv = setitimer(ITIMER_PROF, &itime, NULL);
255 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setitimer(ITIMER_REAL)");
256 
257 	struct rlimit rlim = {};
258 
259 	rv = getrlimit(RLIMIT_CPU, &rlim);
260 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "getrlimit(RLIMIT_CPU)");
261 
262 	rlim.rlim_cur = 1;
263 	rv = setrlimit(RLIMIT_CPU, &rlim);
264 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setrlimit(RLIMIT_CPU)");
265 
266 	rv = pthread_mutex_lock(&gShouldExitMutex);
267 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_mutex_lock(&gShouldExitMutex)");
268 
269 	kr = semaphore_signal_all(gWorkersStart);
270 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "semaphore_signal_all()");
271 
272 	struct timespec timenow = {};
273 	uint64_t time_start;
274 
275 	kr = mach_get_times(&time_start, NULL, &timenow);
276 	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "mach_get_times()");
277 
278 	struct timespec timeout = {
279 		.tv_sec = timenow.tv_sec + 10,
280 		.tv_nsec = timenow.tv_nsec,
281 	};
282 
283 	uint64_t time_end = 0;
284 
285 	do {
286 		assert(os_atomic_load(&gShouldExit, relaxed) == false);
287 
288 		rv = pthread_cond_timedwait(&gShouldExitCondition, &gShouldExitMutex, &timeout);
289 		if (rv == ETIMEDOUT) {
290 			os_atomic_store(&gShouldExit, true, relaxed);
291 
292 			time_end = mach_absolute_time();
293 
294 			struct itimerval itime_stop = {
295 				.it_interval.tv_sec = 0,
296 				.it_interval.tv_usec = 0,
297 				.it_value.tv_sec = 0,
298 				.it_value.tv_usec = 0,  /* stop immediately */
299 			};
300 
301 			rv = setitimer(ITIMER_REAL, &itime_stop, NULL);
302 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setitimer(ITIMER_REAL)");
303 
304 			rv = setitimer(ITIMER_VIRTUAL, &itime_stop, NULL);
305 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setitimer(ITIMER_VIRTUAL)");
306 
307 			rv = setitimer(ITIMER_PROF, &itime_stop, NULL);
308 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "setitimer(ITIMER_PROF)");
309 
310 			break;
311 		} else {
312 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_cond_timedwait(&gShouldExitCondition, ...)");
313 		}
314 	} while (true);
315 
316 	rv = pthread_mutex_unlock(&gShouldExitMutex);
317 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_mutex_unlock(&gShouldExitMutex)");
318 
319 	for (uint32_t i = 1; i < nthreads; i++) {
320 		rv = pthread_join(gThreadList[i].thread, NULL);
321 		T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_join");
322 	}
323 
324 	uint64_t test_duration = time_end - time_start;
325 	uint64_t test_duration_ns = abs_to_nanos(test_duration);
326 
327 	double elapsed_secs = (double) test_duration_ns / (uint64_t)NSEC_PER_SEC;
328 
329 	T_LOG("test duration %3.3f seconds\n", elapsed_secs);
330 
331 	fill_thread_stats(0);
332 
333 	struct rusage_info_v6 ru = {};
334 	rv = proc_pid_rusage(getpid(), RUSAGE_INFO_V6, (rusage_info_t *)&ru);
335 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "proc_pid_rusage");
336 
337 	uint64_t total_user_time_ns = abs_to_nanos_host(ru.ri_user_time);
338 	double total_user_time_s = (double)total_user_time_ns / (uint64_t)NSEC_PER_SEC;
339 
340 	uint64_t total_system_time_ns = abs_to_nanos_host(ru.ri_system_time);
341 	double total_system_time_s = (double)total_system_time_ns / (uint64_t)NSEC_PER_SEC;
342 
343 	uint64_t total_time_ns = (total_user_time_ns + total_system_time_ns);
344 	double total_time_s = (double)total_time_ns / (uint64_t)NSEC_PER_SEC;
345 
346 	uint64_t total_runnable_time_ns = abs_to_nanos_host(ru.ri_runnable_time);
347 	double total_runnable_time_s = (double)total_runnable_time_ns / (uint64_t)NSEC_PER_SEC;
348 
349 	uint64_t total_pending_time_ns = total_runnable_time_ns - (total_time_ns);
350 	double total_pending_time_s = (double)total_pending_time_ns / (uint64_t)NSEC_PER_SEC;
351 
352 	uint64_t total_p_time_ns = abs_to_nanos_host(ru.ri_user_ptime + ru.ri_system_ptime);
353 	double total_p_time_s = (double)total_p_time_ns / (uint64_t)NSEC_PER_SEC;
354 
355 	T_LOG("total usage: time: %3.3f user: %3.3f kernel: %3.3f runnable: %3.3f pending: %3.3f pcore: %3.3f\n",
356 	    total_time_s, total_user_time_s, total_system_time_s,
357 	    total_runnable_time_s, total_pending_time_s,
358 	    total_p_time_s);
359 
360 	/*
361 	 * "Good" data looks like:
362 	 *
363 	 * total usage: time: 77.696 user: 16.570 kernel: 61.126 runnable: 79.951 pending: 2.255 pcore: 72.719
364 	 * Thread        ALRM VTALRM   PROF   XCPU      inst        cycle               user                  kernel          offcore  type
365 	 * 0x16f78f000      0    251    811      0  27680301973  28913501188   3706622958 (  38.14%)   6012631083 (  61.86%)    2.81%  statfs
366 	 * 0x16f81b000      0      2    889      0  27962710058  28780576123    439297291 (   4.53%)   9259942583 (  95.47%)    3.01%  stat
367 	 * 0x16f8a7000      0    251    836      0  27558331077  28889228535   3699010000 (  38.08%)   6016015083 (  61.92%)    2.85%  statfs
368 	 * 0x16f933000      0      0    939      0  28078084696  28880195679    443067500 (   4.56%)   9269807666 (  95.44%)    2.87%  stat
369 	 * 0x16f9bf000      0    283    874      0  27691851016  28969873070   3710916750 (  38.16%)   6012783541 (  61.84%)    2.76%  statfs
370 	 * 0x16fa4b000      0      2    908      1  27945063330  28769971396    438583000 (   4.53%)   9252694291 (  95.47%)    3.09%  stat
371 	 * 0x16fad7000      0    262    889      0  27328496429  28772748055   3689245375 (  38.03%)   6011061458 (  61.97%)    3.00%  statfs
372 	 * 0x16fb63000      0      0    914      0  27942195343  28757254100    439690166 (   4.53%)   9256659500 (  95.47%)    3.04%  stat
373 	 * 0x1fe2bb400   1001      0      3      0     72144372    102339334      3532125 (   9.35%)     34249208 (  90.65%)   99.62%  main
374 	 */
375 	uint32_t total_alrm = 0;
376 	uint32_t total_vtalrm = 0;
377 	uint32_t total_prof = 0;
378 	uint32_t total_xcpu = 0;
379 	uint32_t total_vtalrm_in_cpubound = 0;
380 
381 	uint32_t total_threads_not_finding_cpus = 0;
382 
383 	T_LOG("Thread         ALRM VTALRM   PROF   XCPU      "
384 	    "inst        cycle               user                  kernel          "
385 	    "offcore type\n");
386 
387 	for (uint32_t i = 0; i < nthreads; i++) {
388 		uint64_t user_time = abs_to_nanos_host(gThreadList[i].self_stats.ttci_user_time_mach);
389 		uint64_t system_time = abs_to_nanos_host(gThreadList[i].self_stats.ttci_system_time_mach);
390 
391 
392 		uint64_t total_time = user_time + system_time;
393 
394 		double percentage_user = (double)user_time / (double) total_time * 100;
395 		double percentage_system = (double)system_time / (double) total_time * 100;
396 		double percentage_not_running = (double)(test_duration_ns - total_time) / (double) test_duration_ns * 100;
397 
398 		char* thread_type_str = "";
399 		char* warning_str = "";
400 
401 		if (i == 0) {
402 			thread_type_str = "main ";
403 		} else {
404 			thread_type_str = i % 2 ? "stat   " : "statfs ";
405 
406 			if (percentage_not_running > offcore_time_percent_threshold) {
407 				total_threads_not_finding_cpus++;
408 				warning_str = "** too much offcore time **";
409 			}
410 		}
411 
412 		T_LOG("0x%010llx %6d %6d %6d %6d %12lld %12lld %12lld (%7.2f%%) %12lld (%7.2f%%) %7.2f%% %s%s\n",
413 		    gThreadList[i].tid,
414 		    gThreadList[i].alrm_count,
415 		    gThreadList[i].vtalrm_count,
416 		    gThreadList[i].prof_count,
417 		    gThreadList[i].xcpu_count,
418 		    gThreadList[i].self_stats.ttci_instructions,
419 		    gThreadList[i].self_stats.ttci_cycles,
420 		    user_time, percentage_user,
421 		    system_time, percentage_system,
422 		    percentage_not_running,
423 		    thread_type_str, warning_str);
424 
425 		total_alrm += gThreadList[i].alrm_count;
426 		total_vtalrm += gThreadList[i].vtalrm_count;
427 		total_prof += gThreadList[i].prof_count;
428 		total_xcpu += gThreadList[i].xcpu_count;
429 
430 		if (gThreadList[i].expect_cpu_usage) {
431 			total_vtalrm_in_cpubound += gThreadList[i].vtalrm_count;
432 		}
433 	}
434 
435 	/*
436 	 * We expect all SIGALRM to go to the main thread, because it is the
437 	 * first thread in the process with the signal unmasked, and we
438 	 * never expect the signal handler itself to take >10ms
439 	 *
440 	 * This can happen if the main thread is preempted for the entire 10ms duration, though.
441 	 * Being high priority, it shouldn't be delayed for more than 10ms too often.
442 	 * Allow up to 10% to deliver to other threads.
443 	 */
444 	if ((double)gThreadList[0].alrm_count * 100 / total_alrm < 90.0) {
445 		T_FAIL("SIGALRM delivered to non-main thread more than 10%% of the time (%d of %d)",
446 		    gThreadList[0].alrm_count,
447 		    total_alrm);
448 	}
449 
450 	/* We expect all worker threads to find CPUs of their own for most of the test */
451 	if (total_threads_not_finding_cpus != 0) {
452 		T_FAIL("%d worker threads spent more than %2.0f%% of time off-core",
453 		    total_threads_not_finding_cpus, offcore_time_percent_threshold);
454 	}
455 
456 	/*
457 	 * SIGVTALRM is delivered based on user time, and we expect the busy
458 	 * threads to have an advantage and account for 80% (non-scientific) of events,
459 	 * since the other threads will spend more time in kernel mode.
460 	 */
461 	if (total_vtalrm_in_cpubound * 100 / total_vtalrm < 80) {
462 		T_FAIL("SIGVTALRM delivered to threads without extra userspace spin (only %d of %d)",
463 		    total_vtalrm_in_cpubound, total_vtalrm);
464 	}
465 
466 	/*
467 	 * SIGPROF is delivered based on user+system time, and we expect it to be distributed
468 	 * among non-blocked threads (so not the main thread, which only handles SIGALRM).
469 	 */
470 	if (gThreadList[0].prof_count * 100 / total_prof > 1) {
471 		T_FAIL("SIGPROF delivered to main thread more than 1%% (%d of %d)",
472 		    gThreadList[0].prof_count,
473 		    total_prof);
474 	}
475 
476 	/*
477 	 * SIGXCPU should be delivered exactly once.
478 	 */
479 	if (total_xcpu == 0) {
480 		T_FAIL("SIGXCPU delivered %d times (expected at least once)", total_xcpu);
481 	}
482 }
483 
484 static void *
stat_thread(void * arg)485 stat_thread(void *arg)
486 {
487 	kern_return_t kr;
488 	int rv;
489 
490 	/* This wait can be aborted by one of the signals, so we make sure to wait for the first iteration of main */
491 	kr = semaphore_wait_signal(gWorkersStart, gMainWaitForWorkers);
492 	if (kr != KERN_ABORTED) {
493 		T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "semaphore_wait_signal()");
494 	}
495 
496 	rv = pthread_mutex_lock(&gShouldExitMutex);
497 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_mutex_lock(&gShouldExitMutex)");
498 	rv = pthread_mutex_unlock(&gShouldExitMutex);
499 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_mutex_unlock(&gShouldExitMutex)");
500 
501 	do {
502 		struct stat sb;
503 
504 		rv = stat("/", &sb);
505 		if (rv != 0) {
506 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "stat");
507 		}
508 	} while (os_atomic_load(&gShouldExit, relaxed) == false);
509 
510 	fill_thread_stats((uint32_t)(uintptr_t)arg);
511 
512 	return NULL;
513 }
514 
515 static void *
statfs_thread(void * arg)516 statfs_thread(void *arg)
517 {
518 	kern_return_t kr;
519 	uint64_t previous_spin_timestamp;
520 	int iteration = 0;
521 	int rv;
522 
523 	/* This wait can be aborted by one of the signals, so we make sure to wait for the first iteration of main */
524 	kr = semaphore_wait_signal(gWorkersStart, gMainWaitForWorkers);
525 	if (kr != KERN_ABORTED) {
526 		T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "semaphore_wait_signal()");
527 	}
528 
529 	rv = pthread_mutex_lock(&gShouldExitMutex);
530 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_mutex_lock(&gShouldExitMutex)");
531 	rv = pthread_mutex_unlock(&gShouldExitMutex);
532 	T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "pthread_mutex_unlock(&gShouldExitMutex)");
533 
534 	previous_spin_timestamp = mach_absolute_time();
535 
536 	do {
537 		struct statfs sf;
538 
539 		/*
540 		 * Every so many system calls, inject a spin in userspace
541 		 * proportional to how much time was spent performing the
542 		 * system calls.
543 		 */
544 #define SYSCALL_ITERATIONS_BETWEEN_SPINS (10000)
545 		if (++iteration % SYSCALL_ITERATIONS_BETWEEN_SPINS == 0) {
546 			uint64_t now = mach_absolute_time();
547 			uint64_t spin_deadline = now + (now - previous_spin_timestamp) / 2;
548 
549 			while (mach_absolute_time() < spin_deadline) {
550 				;
551 			}
552 
553 			previous_spin_timestamp = mach_absolute_time();
554 		}
555 
556 		rv = statfs("/", &sf);
557 		if (rv != 0) {
558 			T_QUIET; T_ASSERT_POSIX_SUCCESS(rv, "statfs");
559 		}
560 	} while (os_atomic_load(&gShouldExit, relaxed) == false);
561 
562 	fill_thread_stats((uint32_t)(uintptr_t)arg);
563 
564 	return NULL;
565 }
566 
567 static void
alrm_handler(int signum,struct __siginfo * info __unused,void * uap)568 alrm_handler(int signum, struct __siginfo *info __unused, void *uap)
569 {
570 	ucontext_t *context = (ucontext_t *)uap;
571 	struct threadentry *entry = NULL;
572 	void *sp;
573 
574 #if defined(__arm64__)
575 	sp = (void *)__darwin_arm_thread_state64_get_sp((context->uc_mcontext)->__ss);
576 #elif defined(__i386__)
577 	sp = (void *)(context->uc_mcontext)->__ss.__esp;
578 #elif defined(__x86_64__)
579 	sp = (void *)(context->uc_mcontext)->__ss.__rsp;
580 #else
581 #error Unrecognized architecture
582 #endif
583 
584 	for (uint32_t i = 0; i < nworkers + 1; i++) {
585 		struct threadentry *t = &gThreadList[i];
586 		if (((uintptr_t)sp >= ((uintptr_t)t->stack_addr - t->stack_size) &&
587 		    ((uintptr_t)sp < (uintptr_t)t->stack_addr))) {
588 			entry = t;
589 			break;
590 		}
591 	}
592 
593 	if (entry == NULL) {
594 		T_ASSERT_FAIL("Signal %d delivered to unknown thread, SP=%p", signum, sp);
595 	}
596 
597 	switch (signum) {
598 	case SIGALRM:
599 		os_atomic_inc(&entry->alrm_count, relaxed);
600 		break;
601 	case SIGVTALRM:
602 		os_atomic_inc(&entry->vtalrm_count, relaxed);
603 		break;
604 	case SIGPROF:
605 		os_atomic_inc(&entry->prof_count, relaxed);
606 		break;
607 	case SIGXCPU:
608 		os_atomic_inc(&entry->xcpu_count, relaxed);
609 		break;
610 	}
611 }
612