14 #include "kmp_affinity.h" 19 #include "kmp_stats.h" 21 #include "kmp_wait_release.h" 22 #include "kmp_wrapper_getpid.h" 24 #if !KMP_OS_DRAGONFLY && !KMP_OS_FREEBSD && !KMP_OS_NETBSD && !KMP_OS_OPENBSD 28 #include <sys/resource.h> 29 #include <sys/syscall.h> 31 #include <sys/times.h> 34 #if KMP_OS_LINUX && !KMP_OS_CNK 35 #include <sys/sysinfo.h> 50 #include <mach/mach.h> 51 #include <sys/sysctl.h> 52 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD 53 #include <pthread_np.h> 55 #include <sys/types.h> 56 #include <sys/sysctl.h> 63 #include "tsan_annotations.h" 65 struct kmp_sys_timer {
66 struct timespec start;
70 #define TS2NS(timespec) (((timespec).tv_sec * 1e9) + (timespec).tv_nsec) 72 static struct kmp_sys_timer __kmp_sys_timer_data;
74 #if KMP_HANDLE_SIGNALS 75 typedef void (*sig_func_t)(int);
76 STATIC_EFI2_WORKAROUND
struct sigaction __kmp_sighldrs[NSIG];
77 static sigset_t __kmp_sigset;
80 static int __kmp_init_runtime = FALSE;
82 static int __kmp_fork_count = 0;
84 static pthread_condattr_t __kmp_suspend_cond_attr;
85 static pthread_mutexattr_t __kmp_suspend_mutex_attr;
87 static kmp_cond_align_t __kmp_wait_cv;
88 static kmp_mutex_align_t __kmp_wait_mx;
90 kmp_uint64 __kmp_ticks_per_msec = 1000000;
93 static void __kmp_print_cond(
char *buffer, kmp_cond_align_t *cond) {
94 KMP_SNPRINTF(buffer, 128,
"(cond (lock (%ld, %d)), (descr (%p)))",
95 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
96 cond->c_cond.__c_waiting);
100 #if (KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED) 104 void __kmp_affinity_bind_thread(
int which) {
105 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
106 "Illegal set affinity operation when not capable");
108 kmp_affin_mask_t *mask;
109 KMP_CPU_ALLOC_ON_STACK(mask);
111 KMP_CPU_SET(which, mask);
112 __kmp_set_system_affinity(mask, TRUE);
113 KMP_CPU_FREE_FROM_STACK(mask);
119 void __kmp_affinity_determine_capable(
const char *env_var) {
122 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024) 127 buf = (
unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
132 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_SIZE_LIMIT, buf);
133 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 134 "initial getaffinity call returned %d errno = %d\n",
140 if (__kmp_affinity_verbose ||
141 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
142 (__kmp_affinity_type != affinity_default) &&
143 (__kmp_affinity_type != affinity_disabled))) {
145 kmp_msg_t err_code = KMP_ERR(error);
146 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
147 err_code, __kmp_msg_null);
148 if (__kmp_generate_warnings == kmp_warnings_off) {
149 __kmp_str_free(&err_code.str);
152 KMP_AFFINITY_DISABLE();
153 KMP_INTERNAL_FREE(buf);
161 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
162 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 163 "setaffinity for mask size %d returned %d errno = %d\n",
164 gCode, sCode, errno));
166 if (errno == ENOSYS) {
167 if (__kmp_affinity_verbose ||
168 (__kmp_affinity_warnings &&
169 (__kmp_affinity_type != affinity_none) &&
170 (__kmp_affinity_type != affinity_default) &&
171 (__kmp_affinity_type != affinity_disabled))) {
173 kmp_msg_t err_code = KMP_ERR(error);
174 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
175 err_code, __kmp_msg_null);
176 if (__kmp_generate_warnings == kmp_warnings_off) {
177 __kmp_str_free(&err_code.str);
180 KMP_AFFINITY_DISABLE();
181 KMP_INTERNAL_FREE(buf);
183 if (errno == EFAULT) {
184 KMP_AFFINITY_ENABLE(gCode);
185 KA_TRACE(10, (
"__kmp_affinity_determine_capable: " 186 "affinity supported (mask size %d)\n",
187 (
int)__kmp_affin_mask_size));
188 KMP_INTERNAL_FREE(buf);
196 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 197 "searching for proper set size\n"));
199 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
200 gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
201 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 202 "getaffinity for mask size %d returned %d errno = %d\n",
203 size, gCode, errno));
206 if (errno == ENOSYS) {
208 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 209 "inconsistent OS call behavior: errno == ENOSYS for mask " 212 if (__kmp_affinity_verbose ||
213 (__kmp_affinity_warnings &&
214 (__kmp_affinity_type != affinity_none) &&
215 (__kmp_affinity_type != affinity_default) &&
216 (__kmp_affinity_type != affinity_disabled))) {
218 kmp_msg_t err_code = KMP_ERR(error);
219 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
220 err_code, __kmp_msg_null);
221 if (__kmp_generate_warnings == kmp_warnings_off) {
222 __kmp_str_free(&err_code.str);
225 KMP_AFFINITY_DISABLE();
226 KMP_INTERNAL_FREE(buf);
232 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
233 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 234 "setaffinity for mask size %d returned %d errno = %d\n",
235 gCode, sCode, errno));
237 if (errno == ENOSYS) {
239 KA_TRACE(30, (
"__kmp_affinity_determine_capable: " 240 "inconsistent OS call behavior: errno == ENOSYS for mask " 243 if (__kmp_affinity_verbose ||
244 (__kmp_affinity_warnings &&
245 (__kmp_affinity_type != affinity_none) &&
246 (__kmp_affinity_type != affinity_default) &&
247 (__kmp_affinity_type != affinity_disabled))) {
249 kmp_msg_t err_code = KMP_ERR(error);
250 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
251 err_code, __kmp_msg_null);
252 if (__kmp_generate_warnings == kmp_warnings_off) {
253 __kmp_str_free(&err_code.str);
256 KMP_AFFINITY_DISABLE();
257 KMP_INTERNAL_FREE(buf);
260 if (errno == EFAULT) {
261 KMP_AFFINITY_ENABLE(gCode);
262 KA_TRACE(10, (
"__kmp_affinity_determine_capable: " 263 "affinity supported (mask size %d)\n",
264 (
int)__kmp_affin_mask_size));
265 KMP_INTERNAL_FREE(buf);
272 KMP_INTERNAL_FREE(buf);
277 KMP_AFFINITY_DISABLE();
278 KA_TRACE(10, (
"__kmp_affinity_determine_capable: " 279 "cannot determine mask size - affinity not supported\n"));
280 if (__kmp_affinity_verbose ||
281 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
282 (__kmp_affinity_type != affinity_default) &&
283 (__kmp_affinity_type != affinity_disabled))) {
284 KMP_WARNING(AffCantGetMaskSize, env_var);
288 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 292 int __kmp_futex_determine_capable() {
294 int rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
295 int retval = (rc == 0) || (errno != ENOSYS);
298 (
"__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
299 KA_TRACE(10, (
"__kmp_futex_determine_capable: futex syscall%s supported\n",
300 retval ?
"" :
" not"));
305 #endif // KMP_USE_FUTEX 307 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS) 311 kmp_int8 __kmp_test_then_or8(
volatile kmp_int8 *p, kmp_int8 d) {
312 kmp_int8 old_value, new_value;
314 old_value = TCR_1(*p);
315 new_value = old_value | d;
317 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
319 old_value = TCR_1(*p);
320 new_value = old_value | d;
325 kmp_int8 __kmp_test_then_and8(
volatile kmp_int8 *p, kmp_int8 d) {
326 kmp_int8 old_value, new_value;
328 old_value = TCR_1(*p);
329 new_value = old_value & d;
331 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
333 old_value = TCR_1(*p);
334 new_value = old_value & d;
339 kmp_uint32 __kmp_test_then_or32(
volatile kmp_uint32 *p, kmp_uint32 d) {
340 kmp_uint32 old_value, new_value;
342 old_value = TCR_4(*p);
343 new_value = old_value | d;
345 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
347 old_value = TCR_4(*p);
348 new_value = old_value | d;
353 kmp_uint32 __kmp_test_then_and32(
volatile kmp_uint32 *p, kmp_uint32 d) {
354 kmp_uint32 old_value, new_value;
356 old_value = TCR_4(*p);
357 new_value = old_value & d;
359 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
361 old_value = TCR_4(*p);
362 new_value = old_value & d;
368 kmp_int8 __kmp_test_then_add8(
volatile kmp_int8 *p, kmp_int8 d) {
369 kmp_int8 old_value, new_value;
371 old_value = TCR_1(*p);
372 new_value = old_value + d;
374 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
376 old_value = TCR_1(*p);
377 new_value = old_value + d;
382 kmp_int64 __kmp_test_then_add64(
volatile kmp_int64 *p, kmp_int64 d) {
383 kmp_int64 old_value, new_value;
385 old_value = TCR_8(*p);
386 new_value = old_value + d;
388 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
390 old_value = TCR_8(*p);
391 new_value = old_value + d;
397 kmp_uint64 __kmp_test_then_or64(
volatile kmp_uint64 *p, kmp_uint64 d) {
398 kmp_uint64 old_value, new_value;
400 old_value = TCR_8(*p);
401 new_value = old_value | d;
402 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
404 old_value = TCR_8(*p);
405 new_value = old_value | d;
410 kmp_uint64 __kmp_test_then_and64(
volatile kmp_uint64 *p, kmp_uint64 d) {
411 kmp_uint64 old_value, new_value;
413 old_value = TCR_8(*p);
414 new_value = old_value & d;
415 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
417 old_value = TCR_8(*p);
418 new_value = old_value & d;
425 void __kmp_terminate_thread(
int gtid) {
427 kmp_info_t *th = __kmp_threads[gtid];
432 #ifdef KMP_CANCEL_THREADS 433 KA_TRACE(10, (
"__kmp_terminate_thread: kill (%d)\n", gtid));
434 status = pthread_cancel(th->th.th_info.ds.ds_thread);
435 if (status != 0 && status != ESRCH) {
436 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
447 static kmp_int32 __kmp_set_stack_info(
int gtid, kmp_info_t *th) {
449 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 450 KMP_OS_HURD || KMP_OS_KFREEBSD 459 if (!KMP_UBER_GTID(gtid)) {
462 status = pthread_attr_init(&attr);
463 KMP_CHECK_SYSFAIL(
"pthread_attr_init", status);
464 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD 465 status = pthread_attr_get_np(pthread_self(), &attr);
466 KMP_CHECK_SYSFAIL(
"pthread_attr_get_np", status);
468 status = pthread_getattr_np(pthread_self(), &attr);
469 KMP_CHECK_SYSFAIL(
"pthread_getattr_np", status);
471 status = pthread_attr_getstack(&attr, &addr, &size);
472 KMP_CHECK_SYSFAIL(
"pthread_attr_getstack", status);
474 (
"__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:" 475 " %lu, low addr: %p\n",
477 status = pthread_attr_destroy(&attr);
478 KMP_CHECK_SYSFAIL(
"pthread_attr_destroy", status);
481 if (size != 0 && addr != 0) {
483 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((
char *)addr) + size));
484 TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
485 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
491 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
492 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
493 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
497 static void *__kmp_launch_worker(
void *thr) {
498 int status, old_type, old_state;
499 #ifdef KMP_BLOCK_SIGNALS 500 sigset_t new_set, old_set;
503 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 504 KMP_OS_OPENBSD || KMP_OS_HURD || KMP_OS_KFREEBSD 505 void *
volatile padding = 0;
509 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
510 __kmp_gtid_set_specific(gtid);
511 #ifdef KMP_TDATA_GTID 514 #if KMP_STATS_ENABLED 516 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
517 __kmp_stats_thread_ptr->startLife();
518 KMP_SET_THREAD_STATE(IDLE);
523 __kmp_itt_thread_name(gtid);
526 #if KMP_AFFINITY_SUPPORTED 527 __kmp_affinity_set_init_mask(gtid, FALSE);
530 #ifdef KMP_CANCEL_THREADS 531 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
532 KMP_CHECK_SYSFAIL(
"pthread_setcanceltype", status);
534 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
535 KMP_CHECK_SYSFAIL(
"pthread_setcancelstate", status);
538 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 540 __kmp_clear_x87_fpu_status_word();
541 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
542 __kmp_load_mxcsr(&__kmp_init_mxcsr);
545 #ifdef KMP_BLOCK_SIGNALS 546 status = sigfillset(&new_set);
547 KMP_CHECK_SYSFAIL_ERRNO(
"sigfillset", status);
548 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
549 KMP_CHECK_SYSFAIL(
"pthread_sigmask", status);
552 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 554 if (__kmp_stkoffset > 0 && gtid > 0) {
555 padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
560 __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
562 __kmp_check_stack_overlap((kmp_info_t *)thr);
564 exit_val = __kmp_launch_thread((kmp_info_t *)thr);
566 #ifdef KMP_BLOCK_SIGNALS 567 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
568 KMP_CHECK_SYSFAIL(
"pthread_sigmask", status);
577 static void *__kmp_launch_monitor(
void *thr) {
578 int status, old_type, old_state;
579 #ifdef KMP_BLOCK_SIGNALS 582 struct timespec interval;
586 KA_TRACE(10, (
"__kmp_launch_monitor: #1 launched\n"));
589 __kmp_gtid_set_specific(KMP_GTID_MONITOR);
590 #ifdef KMP_TDATA_GTID 591 __kmp_gtid = KMP_GTID_MONITOR;
598 __kmp_itt_thread_ignore();
601 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
604 __kmp_check_stack_overlap((kmp_info_t *)thr);
606 #ifdef KMP_CANCEL_THREADS 607 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
608 KMP_CHECK_SYSFAIL(
"pthread_setcanceltype", status);
610 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
611 KMP_CHECK_SYSFAIL(
"pthread_setcancelstate", status);
614 #if KMP_REAL_TIME_FIX 619 int sched = sched_getscheduler(0);
620 if (sched == SCHED_FIFO || sched == SCHED_RR) {
623 struct sched_param param;
624 int max_priority = sched_get_priority_max(sched);
626 KMP_WARNING(RealTimeSchedNotSupported);
627 sched_getparam(0, ¶m);
628 if (param.sched_priority < max_priority) {
629 param.sched_priority += 1;
630 rc = sched_setscheduler(0, sched, ¶m);
633 kmp_msg_t err_code = KMP_ERR(error);
634 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
635 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
636 if (__kmp_generate_warnings == kmp_warnings_off) {
637 __kmp_str_free(&err_code.str);
644 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
645 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
650 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
652 #endif // KMP_REAL_TIME_FIX 656 if (__kmp_monitor_wakeups == 1) {
658 interval.tv_nsec = 0;
661 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
664 KA_TRACE(10, (
"__kmp_launch_monitor: #2 monitor\n"));
666 while (!TCR_4(__kmp_global.g.g_done)) {
672 KA_TRACE(15, (
"__kmp_launch_monitor: update\n"));
674 status = gettimeofday(&tval, NULL);
675 KMP_CHECK_SYSFAIL_ERRNO(
"gettimeofday", status);
676 TIMEVAL_TO_TIMESPEC(&tval, &now);
678 now.tv_sec += interval.tv_sec;
679 now.tv_nsec += interval.tv_nsec;
681 if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
683 now.tv_nsec -= KMP_NSEC_PER_SEC;
686 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
687 KMP_CHECK_SYSFAIL(
"pthread_mutex_lock", status);
689 if (!TCR_4(__kmp_global.g.g_done)) {
690 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
691 &__kmp_wait_mx.m_mutex, &now);
693 if (status != ETIMEDOUT && status != EINTR) {
694 KMP_SYSFAIL(
"pthread_cond_timedwait", status);
698 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
699 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
701 TCW_4(__kmp_global.g.g_time.dt.t_value,
702 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
707 KA_TRACE(10, (
"__kmp_launch_monitor: #3 cleanup\n"));
709 #ifdef KMP_BLOCK_SIGNALS 710 status = sigfillset(&new_set);
711 KMP_CHECK_SYSFAIL_ERRNO(
"sigfillset", status);
712 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
713 KMP_CHECK_SYSFAIL(
"pthread_sigmask", status);
716 KA_TRACE(10, (
"__kmp_launch_monitor: #4 finished\n"));
718 if (__kmp_global.g.g_abort != 0) {
724 KA_TRACE(10, (
"__kmp_launch_monitor: #5 terminate sig=%d\n",
725 __kmp_global.g.g_abort));
730 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
731 __kmp_terminate_thread(gtid);
735 KA_TRACE(10, (
"__kmp_launch_monitor: #6 raise sig=%d\n",
736 __kmp_global.g.g_abort));
738 if (__kmp_global.g.g_abort > 0)
739 raise(__kmp_global.g.g_abort);
742 KA_TRACE(10, (
"__kmp_launch_monitor: #7 exit\n"));
746 #endif // KMP_USE_MONITOR 748 void __kmp_create_worker(
int gtid, kmp_info_t *th,
size_t stack_size) {
750 pthread_attr_t thread_attr;
753 th->th.th_info.ds.ds_gtid = gtid;
755 #if KMP_STATS_ENABLED 757 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
763 if (!KMP_UBER_GTID(gtid)) {
764 th->th.th_stats = __kmp_stats_list->push_back(gtid);
768 th->th.th_stats = __kmp_stats_thread_ptr;
770 __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
772 #endif // KMP_STATS_ENABLED 774 if (KMP_UBER_GTID(gtid)) {
775 KA_TRACE(10, (
"__kmp_create_worker: uber thread (%d)\n", gtid));
776 th->th.th_info.ds.ds_thread = pthread_self();
777 __kmp_set_stack_info(gtid, th);
778 __kmp_check_stack_overlap(th);
782 KA_TRACE(10, (
"__kmp_create_worker: try to create thread (%d)\n", gtid));
786 #ifdef KMP_THREAD_ATTR 787 status = pthread_attr_init(&thread_attr);
789 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
791 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
793 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
803 stack_size += gtid * __kmp_stkoffset * 2;
805 KA_TRACE(10, (
"__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 806 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
807 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
809 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 810 status = pthread_attr_setstacksize(&thread_attr, stack_size);
811 #ifdef KMP_BACKUP_STKSIZE 813 if (!__kmp_env_stksize) {
814 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
815 __kmp_stksize = KMP_BACKUP_STKSIZE;
816 KA_TRACE(10, (
"__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 817 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu " 819 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
820 status = pthread_attr_setstacksize(&thread_attr, stack_size);
825 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
826 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
833 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (
void *)th);
834 if (status != 0 || !handle) {
835 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 836 if (status == EINVAL) {
837 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
838 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
840 if (status == ENOMEM) {
841 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
842 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
845 if (status == EAGAIN) {
846 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
847 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
849 KMP_SYSFAIL(
"pthread_create", status);
852 th->th.th_info.ds.ds_thread = handle;
854 #ifdef KMP_THREAD_ATTR 855 status = pthread_attr_destroy(&thread_attr);
857 kmp_msg_t err_code = KMP_ERR(status);
858 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
860 if (__kmp_generate_warnings == kmp_warnings_off) {
861 __kmp_str_free(&err_code.str);
868 KA_TRACE(10, (
"__kmp_create_worker: done creating thread (%d)\n", gtid));
873 void __kmp_create_monitor(kmp_info_t *th) {
875 pthread_attr_t thread_attr;
878 int auto_adj_size = FALSE;
880 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
882 KA_TRACE(10, (
"__kmp_create_monitor: skipping monitor thread because of " 884 th->th.th_info.ds.ds_tid = 0;
885 th->th.th_info.ds.ds_gtid = 0;
888 KA_TRACE(10, (
"__kmp_create_monitor: try to create monitor\n"));
892 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
893 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
894 #if KMP_REAL_TIME_FIX 895 TCW_4(__kmp_global.g.g_time.dt.t_value,
898 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
899 #endif // KMP_REAL_TIME_FIX 901 #ifdef KMP_THREAD_ATTR 902 if (__kmp_monitor_stksize == 0) {
903 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
904 auto_adj_size = TRUE;
906 status = pthread_attr_init(&thread_attr);
908 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
910 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
912 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
915 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 916 status = pthread_attr_getstacksize(&thread_attr, &size);
917 KMP_CHECK_SYSFAIL(
"pthread_attr_getstacksize", status);
919 size = __kmp_sys_min_stksize;
923 if (__kmp_monitor_stksize == 0) {
924 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
926 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
927 __kmp_monitor_stksize = __kmp_sys_min_stksize;
930 KA_TRACE(10, (
"__kmp_create_monitor: default stacksize = %lu bytes," 931 "requested stacksize = %lu bytes\n",
932 size, __kmp_monitor_stksize));
937 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 938 KA_TRACE(10, (
"__kmp_create_monitor: setting stacksize = %lu bytes,",
939 __kmp_monitor_stksize));
940 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
943 __kmp_monitor_stksize *= 2;
946 kmp_msg_t err_code = KMP_ERR(status);
947 __kmp_msg(kmp_ms_warning,
948 KMP_MSG(CantSetMonitorStackSize, (
long int)__kmp_monitor_stksize),
949 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
950 if (__kmp_generate_warnings == kmp_warnings_off) {
951 __kmp_str_free(&err_code.str);
957 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (
void *)th);
960 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 961 if (status == EINVAL) {
962 if (auto_adj_size && (__kmp_monitor_stksize < (
size_t)0x40000000)) {
963 __kmp_monitor_stksize *= 2;
966 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
967 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
970 if (status == ENOMEM) {
971 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
972 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
976 if (status == EAGAIN) {
977 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
978 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
980 KMP_SYSFAIL(
"pthread_create", status);
983 th->th.th_info.ds.ds_thread = handle;
985 #if KMP_REAL_TIME_FIX 987 KMP_DEBUG_ASSERT(
sizeof(kmp_uint32) ==
988 sizeof(__kmp_global.g.g_time.dt.t_value));
989 __kmp_wait_4((kmp_uint32
volatile *)&__kmp_global.g.g_time.dt.t_value, -1,
991 #endif // KMP_REAL_TIME_FIX 993 #ifdef KMP_THREAD_ATTR 994 status = pthread_attr_destroy(&thread_attr);
996 kmp_msg_t err_code = KMP_ERR(status);
997 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
999 if (__kmp_generate_warnings == kmp_warnings_off) {
1000 __kmp_str_free(&err_code.str);
1007 KA_TRACE(10, (
"__kmp_create_monitor: monitor created %#.8lx\n",
1008 th->th.th_info.ds.ds_thread));
1011 #endif // KMP_USE_MONITOR 1013 void __kmp_exit_thread(
int exit_status) {
1014 pthread_exit((
void *)(intptr_t)exit_status);
1018 void __kmp_resume_monitor();
1020 void __kmp_reap_monitor(kmp_info_t *th) {
1024 KA_TRACE(10, (
"__kmp_reap_monitor: try to reap monitor thread with handle" 1026 th->th.th_info.ds.ds_thread));
1031 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1032 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1033 KA_TRACE(10, (
"__kmp_reap_monitor: monitor did not start, returning\n"));
1043 status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1044 if (status != ESRCH) {
1045 __kmp_resume_monitor();
1047 KA_TRACE(10, (
"__kmp_reap_monitor: try to join with monitor\n"));
1048 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1049 if (exit_val != th) {
1050 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1053 th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1054 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1056 KA_TRACE(10, (
"__kmp_reap_monitor: done reaping monitor thread with handle" 1058 th->th.th_info.ds.ds_thread));
1062 #endif // KMP_USE_MONITOR 1064 void __kmp_reap_worker(kmp_info_t *th) {
1071 10, (
"__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1073 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1077 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1079 if (exit_val != th) {
1080 KA_TRACE(10, (
"__kmp_reap_worker: worker T#%d did not reap properly, " 1082 th->th.th_info.ds.ds_gtid, exit_val));
1086 KA_TRACE(10, (
"__kmp_reap_worker: done reaping T#%d\n",
1087 th->th.th_info.ds.ds_gtid));
1092 #if KMP_HANDLE_SIGNALS 1094 static void __kmp_null_handler(
int signo) {
1098 static void __kmp_team_handler(
int signo) {
1099 if (__kmp_global.g.g_abort == 0) {
1102 __kmp_debug_printf(
"__kmp_team_handler: caught signal = %d\n", signo);
1117 if (__kmp_debug_buf) {
1118 __kmp_dump_debug_buffer();
1121 TCW_4(__kmp_global.g.g_abort, signo);
1123 TCW_4(__kmp_global.g.g_done, TRUE);
1128 __kmp_debug_printf(
"__kmp_team_handler: unknown signal type");
1135 static void __kmp_sigaction(
int signum,
const struct sigaction *act,
1136 struct sigaction *oldact) {
1137 int rc = sigaction(signum, act, oldact);
1138 KMP_CHECK_SYSFAIL_ERRNO(
"sigaction", rc);
1141 static void __kmp_install_one_handler(
int sig, sig_func_t handler_func,
1142 int parallel_init) {
1145 (
"__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1146 if (parallel_init) {
1147 struct sigaction new_action;
1148 struct sigaction old_action;
1149 new_action.sa_handler = handler_func;
1150 new_action.sa_flags = 0;
1151 sigfillset(&new_action.sa_mask);
1152 __kmp_sigaction(sig, &new_action, &old_action);
1153 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1154 sigaddset(&__kmp_sigset, sig);
1157 __kmp_sigaction(sig, &old_action, NULL);
1161 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1166 static void __kmp_remove_one_handler(
int sig) {
1167 KB_TRACE(60, (
"__kmp_remove_one_handler( %d )\n", sig));
1168 if (sigismember(&__kmp_sigset, sig)) {
1169 struct sigaction old;
1171 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1172 if ((old.sa_handler != __kmp_team_handler) &&
1173 (old.sa_handler != __kmp_null_handler)) {
1175 KB_TRACE(10, (
"__kmp_remove_one_handler: oops, not our handler, " 1176 "restoring: sig=%d\n",
1178 __kmp_sigaction(sig, &old, NULL);
1180 sigdelset(&__kmp_sigset, sig);
1185 void __kmp_install_signals(
int parallel_init) {
1186 KB_TRACE(10, (
"__kmp_install_signals( %d )\n", parallel_init));
1187 if (__kmp_handle_signals || !parallel_init) {
1190 sigemptyset(&__kmp_sigset);
1191 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1192 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1193 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1194 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1195 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1196 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1197 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1198 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1200 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1202 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1204 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1209 void __kmp_remove_signals(
void) {
1211 KB_TRACE(10, (
"__kmp_remove_signals()\n"));
1212 for (sig = 1; sig < NSIG; ++sig) {
1213 __kmp_remove_one_handler(sig);
1217 #endif // KMP_HANDLE_SIGNALS 1219 void __kmp_enable(
int new_state) {
1220 #ifdef KMP_CANCEL_THREADS 1221 int status, old_state;
1222 status = pthread_setcancelstate(new_state, &old_state);
1223 KMP_CHECK_SYSFAIL(
"pthread_setcancelstate", status);
1224 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1228 void __kmp_disable(
int *old_state) {
1229 #ifdef KMP_CANCEL_THREADS 1231 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1232 KMP_CHECK_SYSFAIL(
"pthread_setcancelstate", status);
1236 static void __kmp_atfork_prepare(
void) {
1237 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1238 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1241 static void __kmp_atfork_parent(
void) {
1242 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1243 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1249 static void __kmp_atfork_child(
void) {
1250 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1257 #if KMP_AFFINITY_SUPPORTED 1261 kmp_set_thread_affinity_mask_initial();
1266 __kmp_affinity_type = affinity_none;
1268 if (__kmp_nested_proc_bind.bind_types != NULL) {
1269 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1271 #endif // OMP_40_ENABLED 1272 #endif // KMP_AFFINITY_SUPPORTED 1274 __kmp_init_runtime = FALSE;
1276 __kmp_init_monitor = 0;
1278 __kmp_init_parallel = FALSE;
1279 __kmp_init_middle = FALSE;
1280 __kmp_init_serial = FALSE;
1281 TCW_4(__kmp_init_gtid, FALSE);
1282 __kmp_init_common = FALSE;
1284 TCW_4(__kmp_init_user_locks, FALSE);
1285 #if !KMP_USE_DYNAMIC_LOCK 1286 __kmp_user_lock_table.used = 1;
1287 __kmp_user_lock_table.allocated = 0;
1288 __kmp_user_lock_table.table = NULL;
1289 __kmp_lock_blocks = NULL;
1293 TCW_4(__kmp_nth, 0);
1295 __kmp_thread_pool = NULL;
1296 __kmp_thread_pool_insert_pt = NULL;
1297 __kmp_team_pool = NULL;
1301 KA_TRACE(10, (
"__kmp_atfork_child: checking cache address list %p\n",
1302 __kmp_threadpriv_cache_list));
1304 while (__kmp_threadpriv_cache_list != NULL) {
1306 if (*__kmp_threadpriv_cache_list->addr != NULL) {
1307 KC_TRACE(50, (
"__kmp_atfork_child: zeroing cache at address %p\n",
1308 &(*__kmp_threadpriv_cache_list->addr)));
1310 *__kmp_threadpriv_cache_list->addr = NULL;
1312 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1315 __kmp_init_runtime = FALSE;
1318 __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1319 __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1320 __kmp_init_bootstrap_lock(&__kmp_console_lock);
1321 __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1338 void __kmp_register_atfork(
void) {
1339 if (__kmp_need_register_atfork) {
1340 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1341 __kmp_atfork_child);
1342 KMP_CHECK_SYSFAIL(
"pthread_atfork", status);
1343 __kmp_need_register_atfork = FALSE;
1347 void __kmp_suspend_initialize(
void) {
1349 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1350 KMP_CHECK_SYSFAIL(
"pthread_mutexattr_init", status);
1351 status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1352 KMP_CHECK_SYSFAIL(
"pthread_condattr_init", status);
1355 void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1356 ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count);
1357 if (th->th.th_suspend_init_count <= __kmp_fork_count) {
1361 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1362 &__kmp_suspend_cond_attr);
1363 KMP_CHECK_SYSFAIL(
"pthread_cond_init", status);
1364 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1365 &__kmp_suspend_mutex_attr);
1366 KMP_CHECK_SYSFAIL(
"pthread_mutex_init", status);
1367 *(
volatile int *)&th->th.th_suspend_init_count = __kmp_fork_count + 1;
1368 ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count);
1372 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1373 if (th->th.th_suspend_init_count > __kmp_fork_count) {
1378 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1379 if (status != 0 && status != EBUSY) {
1380 KMP_SYSFAIL(
"pthread_cond_destroy", status);
1382 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1383 if (status != 0 && status != EBUSY) {
1384 KMP_SYSFAIL(
"pthread_mutex_destroy", status);
1386 --th->th.th_suspend_init_count;
1387 KMP_DEBUG_ASSERT(th->th.th_suspend_init_count == __kmp_fork_count);
1392 int __kmp_try_suspend_mx(kmp_info_t *th) {
1393 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1396 void __kmp_lock_suspend_mx(kmp_info_t *th) {
1397 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1398 KMP_CHECK_SYSFAIL(
"pthread_mutex_lock", status);
1401 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1402 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1403 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1409 static inline void __kmp_suspend_template(
int th_gtid, C *flag) {
1410 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1411 kmp_info_t *th = __kmp_threads[th_gtid];
1413 typename C::flag_t old_spin;
1415 KF_TRACE(30, (
"__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1418 __kmp_suspend_initialize_thread(th);
1420 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1421 KMP_CHECK_SYSFAIL(
"pthread_mutex_lock", status);
1423 KF_TRACE(10, (
"__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1424 th_gtid, flag->get()));
1428 old_spin = flag->set_sleeping();
1430 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1431 __kmp_pause_status != kmp_soft_paused) {
1432 flag->unset_sleeping();
1433 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1434 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1438 KF_TRACE(5, (
"__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1440 th_gtid, flag->get(), flag->load(), old_spin));
1442 if (flag->done_check_val(old_spin)) {
1443 old_spin = flag->unset_sleeping();
1444 KF_TRACE(5, (
"__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1446 th_gtid, flag->get()));
1451 int deactivated = FALSE;
1452 TCW_PTR(th->th.th_sleep_loc, (
void *)flag);
1454 while (flag->is_sleeping()) {
1455 #ifdef DEBUG_SUSPEND 1457 __kmp_suspend_count++;
1458 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1459 __kmp_printf(
"__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1465 th->th.th_active = FALSE;
1466 if (th->th.th_active_in_pool) {
1467 th->th.th_active_in_pool = FALSE;
1468 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1469 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1474 #if USE_SUSPEND_TIMEOUT 1475 struct timespec now;
1476 struct timeval tval;
1479 status = gettimeofday(&tval, NULL);
1480 KMP_CHECK_SYSFAIL_ERRNO(
"gettimeofday", status);
1481 TIMEVAL_TO_TIMESPEC(&tval, &now);
1483 msecs = (4 * __kmp_dflt_blocktime) + 200;
1484 now.tv_sec += msecs / 1000;
1485 now.tv_nsec += (msecs % 1000) * 1000;
1487 KF_TRACE(15, (
"__kmp_suspend_template: T#%d about to perform " 1488 "pthread_cond_timedwait\n",
1490 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1491 &th->th.th_suspend_mx.m_mutex, &now);
1493 KF_TRACE(15, (
"__kmp_suspend_template: T#%d about to perform" 1494 " pthread_cond_wait\n",
1496 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1497 &th->th.th_suspend_mx.m_mutex);
1500 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1501 KMP_SYSFAIL(
"pthread_cond_wait", status);
1504 if (status == ETIMEDOUT) {
1505 if (flag->is_sleeping()) {
1507 (
"__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1509 KF_TRACE(2, (
"__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1513 }
else if (flag->is_sleeping()) {
1515 (
"__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1522 th->th.th_active = TRUE;
1523 if (TCR_4(th->th.th_in_pool)) {
1524 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1525 th->th.th_active_in_pool = TRUE;
1529 #ifdef DEBUG_SUSPEND 1532 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1533 __kmp_printf(
"__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1538 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1539 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1540 KF_TRACE(30, (
"__kmp_suspend_template: T#%d exit\n", th_gtid));
1543 void __kmp_suspend_32(
int th_gtid, kmp_flag_32 *flag) {
1544 __kmp_suspend_template(th_gtid, flag);
1546 void __kmp_suspend_64(
int th_gtid, kmp_flag_64 *flag) {
1547 __kmp_suspend_template(th_gtid, flag);
1549 void __kmp_suspend_oncore(
int th_gtid, kmp_flag_oncore *flag) {
1550 __kmp_suspend_template(th_gtid, flag);
1557 static inline void __kmp_resume_template(
int target_gtid, C *flag) {
1558 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1559 kmp_info_t *th = __kmp_threads[target_gtid];
1563 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1566 KF_TRACE(30, (
"__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1567 gtid, target_gtid));
1568 KMP_DEBUG_ASSERT(gtid != target_gtid);
1570 __kmp_suspend_initialize_thread(th);
1572 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1573 KMP_CHECK_SYSFAIL(
"pthread_mutex_lock", status);
1576 flag = (C *)CCAST(
void *, th->th.th_sleep_loc);
1581 if (!flag || flag->get_type() != flag->get_ptr_type()) {
1584 KF_TRACE(5, (
"__kmp_resume_template: T#%d exiting, thread T#%d already " 1585 "awake: flag(%p)\n",
1586 gtid, target_gtid, NULL));
1587 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1588 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1592 typename C::flag_t old_spin = flag->unset_sleeping();
1593 if (!flag->is_sleeping_val(old_spin)) {
1594 KF_TRACE(5, (
"__kmp_resume_template: T#%d exiting, thread T#%d already " 1597 gtid, target_gtid, flag->get(), old_spin, flag->load()));
1598 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1599 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1602 KF_TRACE(5, (
"__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1603 "sleep bit for flag's loc(%p): " 1605 gtid, target_gtid, flag->get(), old_spin, flag->load()));
1607 TCW_PTR(th->th.th_sleep_loc, NULL);
1609 #ifdef DEBUG_SUSPEND 1612 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1613 __kmp_printf(
"__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1614 target_gtid, buffer);
1617 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1618 KMP_CHECK_SYSFAIL(
"pthread_cond_signal", status);
1619 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1620 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1621 KF_TRACE(30, (
"__kmp_resume_template: T#%d exiting after signaling wake up" 1623 gtid, target_gtid));
1626 void __kmp_resume_32(
int target_gtid, kmp_flag_32 *flag) {
1627 __kmp_resume_template(target_gtid, flag);
1629 void __kmp_resume_64(
int target_gtid, kmp_flag_64 *flag) {
1630 __kmp_resume_template(target_gtid, flag);
1632 void __kmp_resume_oncore(
int target_gtid, kmp_flag_oncore *flag) {
1633 __kmp_resume_template(target_gtid, flag);
1637 void __kmp_resume_monitor() {
1638 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1641 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1642 KF_TRACE(30, (
"__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1644 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1646 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1647 KMP_CHECK_SYSFAIL(
"pthread_mutex_lock", status);
1648 #ifdef DEBUG_SUSPEND 1651 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1652 __kmp_printf(
"__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1653 KMP_GTID_MONITOR, buffer);
1656 status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1657 KMP_CHECK_SYSFAIL(
"pthread_cond_signal", status);
1658 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1659 KMP_CHECK_SYSFAIL(
"pthread_mutex_unlock", status);
1660 KF_TRACE(30, (
"__kmp_resume_monitor: T#%d exiting after signaling wake up" 1662 gtid, KMP_GTID_MONITOR));
1664 #endif // KMP_USE_MONITOR 1666 void __kmp_yield() { sched_yield(); }
1668 void __kmp_gtid_set_specific(
int gtid) {
1669 if (__kmp_init_gtid) {
1671 status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1672 (
void *)(intptr_t)(gtid + 1));
1673 KMP_CHECK_SYSFAIL(
"pthread_setspecific", status);
1675 KA_TRACE(50, (
"__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1679 int __kmp_gtid_get_specific() {
1681 if (!__kmp_init_gtid) {
1682 KA_TRACE(50, (
"__kmp_gtid_get_specific: runtime shutdown, returning " 1683 "KMP_GTID_SHUTDOWN\n"));
1684 return KMP_GTID_SHUTDOWN;
1686 gtid = (int)(
size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1688 gtid = KMP_GTID_DNE;
1692 KA_TRACE(50, (
"__kmp_gtid_get_specific: key:%d gtid:%d\n",
1693 __kmp_gtid_threadprivate_key, gtid));
1697 double __kmp_read_cpu_time(
void) {
1703 return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC;
1706 int __kmp_read_system_info(
struct kmp_sys_info *info) {
1708 struct rusage r_usage;
1710 memset(info, 0,
sizeof(*info));
1712 status = getrusage(RUSAGE_SELF, &r_usage);
1713 KMP_CHECK_SYSFAIL_ERRNO(
"getrusage", status);
1716 info->maxrss = r_usage.ru_maxrss;
1718 info->minflt = r_usage.ru_minflt;
1720 info->majflt = r_usage.ru_majflt;
1722 info->nswap = r_usage.ru_nswap;
1724 info->inblock = r_usage.ru_inblock;
1726 info->oublock = r_usage.ru_oublock;
1728 info->nvcsw = r_usage.ru_nvcsw;
1730 info->nivcsw = r_usage.ru_nivcsw;
1732 return (status != 0);
1735 void __kmp_read_system_time(
double *delta) {
1737 struct timeval tval;
1738 struct timespec stop;
1741 status = gettimeofday(&tval, NULL);
1742 KMP_CHECK_SYSFAIL_ERRNO(
"gettimeofday", status);
1743 TIMEVAL_TO_TIMESPEC(&tval, &stop);
1744 t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start);
1745 *delta = (t_ns * 1e-9);
1748 void __kmp_clear_system_time(
void) {
1749 struct timeval tval;
1751 status = gettimeofday(&tval, NULL);
1752 KMP_CHECK_SYSFAIL_ERRNO(
"gettimeofday", status);
1753 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1756 static int __kmp_get_xproc(
void) {
1760 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 1761 KMP_OS_OPENBSD || KMP_OS_HURD || KMP_OS_KFREEBSD 1763 r = sysconf(_SC_NPROCESSORS_ONLN);
1771 host_basic_info_data_t info;
1772 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1773 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1774 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1777 r = info.avail_cpus;
1779 KMP_WARNING(CantGetNumAvailCPU);
1780 KMP_INFORM(AssumedNumCPU);
1785 #error "Unknown or unsupported OS." 1789 return r > 0 ? r : 2;
1793 int __kmp_read_from_file(
char const *path,
char const *format, ...) {
1797 va_start(args, format);
1798 FILE *f = fopen(path,
"rb");
1801 result = vfscanf(f, format, args);
1807 void __kmp_runtime_initialize(
void) {
1809 pthread_mutexattr_t mutex_attr;
1810 pthread_condattr_t cond_attr;
1812 if (__kmp_init_runtime) {
1816 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1817 if (!__kmp_cpuinfo.initialized) {
1818 __kmp_query_cpuid(&__kmp_cpuinfo);
1822 __kmp_xproc = __kmp_get_xproc();
1824 if (sysconf(_SC_THREADS)) {
1827 __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX);
1828 if (__kmp_sys_max_nth == -1) {
1830 __kmp_sys_max_nth = INT_MAX;
1831 }
else if (__kmp_sys_max_nth <= 1) {
1833 __kmp_sys_max_nth = KMP_MAX_NTH;
1837 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1838 if (__kmp_sys_min_stksize <= 1) {
1839 __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1844 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1846 status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1847 __kmp_internal_end_dest);
1848 KMP_CHECK_SYSFAIL(
"pthread_key_create", status);
1849 status = pthread_mutexattr_init(&mutex_attr);
1850 KMP_CHECK_SYSFAIL(
"pthread_mutexattr_init", status);
1851 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1852 KMP_CHECK_SYSFAIL(
"pthread_mutex_init", status);
1853 status = pthread_condattr_init(&cond_attr);
1854 KMP_CHECK_SYSFAIL(
"pthread_condattr_init", status);
1855 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1856 KMP_CHECK_SYSFAIL(
"pthread_cond_init", status);
1858 __kmp_itt_initialize();
1861 __kmp_init_runtime = TRUE;
1864 void __kmp_runtime_destroy(
void) {
1867 if (!__kmp_init_runtime) {
1872 __kmp_itt_destroy();
1875 status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1876 KMP_CHECK_SYSFAIL(
"pthread_key_delete", status);
1878 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1879 if (status != 0 && status != EBUSY) {
1880 KMP_SYSFAIL(
"pthread_mutex_destroy", status);
1882 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1883 if (status != 0 && status != EBUSY) {
1884 KMP_SYSFAIL(
"pthread_cond_destroy", status);
1886 #if KMP_AFFINITY_SUPPORTED 1887 __kmp_affinity_uninitialize();
1890 __kmp_init_runtime = FALSE;
1895 void __kmp_thread_sleep(
int millis) { sleep((millis + 500) / 1000); }
1898 void __kmp_elapsed(
double *t) {
1900 #ifdef FIX_SGI_CLOCK 1903 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1904 KMP_CHECK_SYSFAIL_ERRNO(
"clock_gettime", status);
1906 (double)ts.tv_nsec * (1.0 / (
double)KMP_NSEC_PER_SEC) + (
double)ts.tv_sec;
1910 status = gettimeofday(&tv, NULL);
1911 KMP_CHECK_SYSFAIL_ERRNO(
"gettimeofday", status);
1913 (double)tv.tv_usec * (1.0 / (
double)KMP_USEC_PER_SEC) + (
double)tv.tv_sec;
1918 void __kmp_elapsed_tick(
double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1921 kmp_uint64 __kmp_now_nsec() {
1923 gettimeofday(&t, NULL);
1924 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
1925 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
1929 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1931 void __kmp_initialize_system_tick() {
1932 kmp_uint64 now, nsec2, diff;
1933 kmp_uint64 delay = 100000;
1934 kmp_uint64 nsec = __kmp_now_nsec();
1935 kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
1936 while ((now = __kmp_hardware_timestamp()) < goal)
1938 nsec2 = __kmp_now_nsec();
1939 diff = nsec2 - nsec;
1941 kmp_uint64 tpms = (kmp_uint64)(1e6 * (delay + (now - goal)) / diff);
1943 __kmp_ticks_per_msec = tpms;
1951 int __kmp_is_address_mapped(
void *addr) {
1956 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_HURD || KMP_OS_KFREEBSD 1961 char *name = __kmp_str_format(
"/proc/%d/maps", getpid());
1964 file = fopen(name,
"r");
1965 KMP_ASSERT(file != NULL);
1969 void *beginning = NULL;
1970 void *ending = NULL;
1973 rc = fscanf(file,
"%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
1977 KMP_ASSERT(rc == 3 &&
1978 KMP_STRLEN(perms) == 4);
1981 if ((addr >= beginning) && (addr < ending)) {
1983 if (strcmp(perms,
"rw") == 0) {
1993 KMP_INTERNAL_FREE(name);
2002 rc = vm_read_overwrite(
2004 (vm_address_t)(addr),
2006 (vm_address_t)(&buffer),
2019 mib[2] = VM_PROC_MAP;
2021 mib[4] =
sizeof(
struct kinfo_vmentry);
2024 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2028 size = size * 4 / 3;
2029 struct kinfo_vmentry *kiv = (
struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2032 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2036 for (
size_t i = 0; i < size; i++) {
2037 if (kiv[i].kve_start >= (uint64_t)addr &&
2038 kiv[i].kve_end <= (uint64_t)addr) {
2043 KMP_INTERNAL_FREE(kiv);
2044 #elif KMP_OS_DRAGONFLY || KMP_OS_OPENBSD 2051 #error "Unknown or unsupported OS" 2059 #ifdef USE_LOAD_BALANCE 2061 #if KMP_OS_DARWIN || KMP_OS_NETBSD 2068 int __kmp_get_load_balance(
int max) {
2072 int res = getloadavg(averages, 3);
2077 if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2078 ret_avg = averages[0];
2079 }
else if ((__kmp_load_balance_interval >= 180 &&
2080 __kmp_load_balance_interval < 600) &&
2082 ret_avg = averages[1];
2083 }
else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2084 ret_avg = averages[2];
2098 int __kmp_get_load_balance(
int max) {
2099 static int permanent_error = 0;
2100 static int glb_running_threads = 0;
2102 static double glb_call_time = 0;
2104 int running_threads = 0;
2106 DIR *proc_dir = NULL;
2107 struct dirent *proc_entry = NULL;
2109 kmp_str_buf_t task_path;
2110 DIR *task_dir = NULL;
2111 struct dirent *task_entry = NULL;
2112 int task_path_fixed_len;
2114 kmp_str_buf_t stat_path;
2116 int stat_path_fixed_len;
2118 int total_processes = 0;
2119 int total_threads = 0;
2121 double call_time = 0.0;
2123 __kmp_str_buf_init(&task_path);
2124 __kmp_str_buf_init(&stat_path);
2126 __kmp_elapsed(&call_time);
2128 if (glb_call_time &&
2129 (call_time - glb_call_time < __kmp_load_balance_interval)) {
2130 running_threads = glb_running_threads;
2134 glb_call_time = call_time;
2137 if (permanent_error) {
2138 running_threads = -1;
2147 proc_dir = opendir(
"/proc");
2148 if (proc_dir == NULL) {
2151 running_threads = -1;
2152 permanent_error = 1;
2157 __kmp_str_buf_cat(&task_path,
"/proc/", 6);
2158 task_path_fixed_len = task_path.used;
2160 proc_entry = readdir(proc_dir);
2161 while (proc_entry != NULL) {
2164 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2172 KMP_DEBUG_ASSERT(total_processes != 1 ||
2173 strcmp(proc_entry->d_name,
"1") == 0);
2176 task_path.used = task_path_fixed_len;
2177 __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2178 KMP_STRLEN(proc_entry->d_name));
2179 __kmp_str_buf_cat(&task_path,
"/task", 5);
2181 task_dir = opendir(task_path.str);
2182 if (task_dir == NULL) {
2191 if (strcmp(proc_entry->d_name,
"1") == 0) {
2192 running_threads = -1;
2193 permanent_error = 1;
2198 __kmp_str_buf_clear(&stat_path);
2199 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2200 __kmp_str_buf_cat(&stat_path,
"/", 1);
2201 stat_path_fixed_len = stat_path.used;
2203 task_entry = readdir(task_dir);
2204 while (task_entry != NULL) {
2206 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2214 stat_path_fixed_len;
2215 __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2216 KMP_STRLEN(task_entry->d_name));
2217 __kmp_str_buf_cat(&stat_path,
"/stat", 5);
2221 stat_file = open(stat_path.str, O_RDONLY);
2222 if (stat_file == -1) {
2252 len = read(stat_file, buffer,
sizeof(buffer) - 1);
2259 char *close_parent = strstr(buffer,
") ");
2260 if (close_parent != NULL) {
2261 char state = *(close_parent + 2);
2264 if (running_threads >= max) {
2274 task_entry = readdir(task_dir);
2280 proc_entry = readdir(proc_dir);
2286 KMP_DEBUG_ASSERT(running_threads > 0);
2287 if (running_threads <= 0) {
2288 running_threads = 1;
2292 if (proc_dir != NULL) {
2295 __kmp_str_buf_free(&task_path);
2296 if (task_dir != NULL) {
2299 __kmp_str_buf_free(&stat_path);
2300 if (stat_file != -1) {
2304 glb_running_threads = running_threads;
2306 return running_threads;
2310 #endif // KMP_OS_DARWIN 2312 #endif // USE_LOAD_BALANCE 2314 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2315 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || KMP_ARCH_PPC64) 2319 int __kmp_invoke_microtask(microtask_t pkfn,
int gtid,
int tid,
int argc,
2323 void **exit_frame_ptr
2327 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2332 fprintf(stderr,
"Too many args to microtask: %d!\n", argc);
2336 (*pkfn)(>id, &tid);
2339 (*pkfn)(>id, &tid, p_argv[0]);
2342 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]);
2345 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]);
2348 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2351 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2354 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2358 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2359 p_argv[5], p_argv[6]);
2362 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2363 p_argv[5], p_argv[6], p_argv[7]);
2366 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2367 p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2370 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2371 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2374 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2375 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2378 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2379 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2383 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2384 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2385 p_argv[11], p_argv[12]);
2388 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2389 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2390 p_argv[11], p_argv[12], p_argv[13]);
2393 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2394 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2395 p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2400 *exit_frame_ptr = 0;
#define KMP_INIT_PARTITIONED_TIMERS(name)
Initializes the paritioned timers to begin with name.