LLVM OpenMP* Runtime Library
z_Linux_util.cpp
1/*
2 * z_Linux_util.cpp -- platform specific routines.
3 */
4
5//===----------------------------------------------------------------------===//
6//
7// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8// See https://llvm.org/LICENSE.txt for license information.
9// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10//
11//===----------------------------------------------------------------------===//
12
13#include "kmp.h"
14#include "kmp_affinity.h"
15#include "kmp_i18n.h"
16#include "kmp_io.h"
17#include "kmp_itt.h"
18#include "kmp_lock.h"
19#include "kmp_stats.h"
20#include "kmp_str.h"
21#include "kmp_wait_release.h"
22#include "kmp_wrapper_getpid.h"
23
24#if !KMP_OS_DRAGONFLY && !KMP_OS_FREEBSD && !KMP_OS_NETBSD && !KMP_OS_OPENBSD
25#include <alloca.h>
26#endif
27#include <math.h> // HUGE_VAL.
28#if KMP_OS_LINUX
29#include <semaphore.h>
30#endif // KMP_OS_LINUX
31#include <sys/resource.h>
32#include <sys/syscall.h>
33#include <sys/time.h>
34#include <sys/times.h>
35#include <unistd.h>
36
37#if KMP_OS_LINUX
38#include <sys/sysinfo.h>
39#if KMP_USE_FUTEX
40// We should really include <futex.h>, but that causes compatibility problems on
41// different Linux* OS distributions that either require that you include (or
42// break when you try to include) <pci/types.h>. Since all we need is the two
43// macros below (which are part of the kernel ABI, so can't change) we just
44// define the constants here and don't include <futex.h>
45#ifndef FUTEX_WAIT
46#define FUTEX_WAIT 0
47#endif
48#ifndef FUTEX_WAKE
49#define FUTEX_WAKE 1
50#endif
51#endif
52#elif KMP_OS_DARWIN
53#include <mach/mach.h>
54#include <sys/sysctl.h>
55#elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD
56#include <sys/types.h>
57#include <sys/sysctl.h>
58#include <sys/user.h>
59#include <pthread_np.h>
60#elif KMP_OS_NETBSD || KMP_OS_OPENBSD
61#include <sys/types.h>
62#include <sys/sysctl.h>
63#endif
64
65#include <ctype.h>
66#include <dirent.h>
67#include <fcntl.h>
68
69struct kmp_sys_timer {
70 struct timespec start;
71};
72
73// Convert timespec to nanoseconds.
74#define TS2NS(timespec) \
75 (((timespec).tv_sec * (long int)1e9) + (timespec).tv_nsec)
76
77static struct kmp_sys_timer __kmp_sys_timer_data;
78
79#if KMP_HANDLE_SIGNALS
80typedef void (*sig_func_t)(int);
81STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG];
82static sigset_t __kmp_sigset;
83#endif
84
85static int __kmp_init_runtime = FALSE;
86
87static int __kmp_fork_count = 0;
88
89static pthread_condattr_t __kmp_suspend_cond_attr;
90static pthread_mutexattr_t __kmp_suspend_mutex_attr;
91
92static kmp_cond_align_t __kmp_wait_cv;
93static kmp_mutex_align_t __kmp_wait_mx;
94
95kmp_uint64 __kmp_ticks_per_msec = 1000000;
96
97#ifdef DEBUG_SUSPEND
98static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
99 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
100 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
101 cond->c_cond.__c_waiting);
102}
103#endif
104
105#if ((KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED)
106
107/* Affinity support */
108
109void __kmp_affinity_bind_thread(int which) {
110 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
111 "Illegal set affinity operation when not capable");
112
113 kmp_affin_mask_t *mask;
114 KMP_CPU_ALLOC_ON_STACK(mask);
115 KMP_CPU_ZERO(mask);
116 KMP_CPU_SET(which, mask);
117 __kmp_set_system_affinity(mask, TRUE);
118 KMP_CPU_FREE_FROM_STACK(mask);
119}
120
121/* Determine if we can access affinity functionality on this version of
122 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
123 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
124void __kmp_affinity_determine_capable(const char *env_var) {
125 // Check and see if the OS supports thread affinity.
126
127#if KMP_OS_LINUX
128#define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
129#define KMP_CPU_SET_TRY_SIZE CACHE_LINE
130#elif KMP_OS_FREEBSD
131#define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t))
132#endif
133
134 int verbose = __kmp_affinity.flags.verbose;
135 int warnings = __kmp_affinity.flags.warnings;
136 enum affinity_type type = __kmp_affinity.type;
137
138#if KMP_OS_LINUX
139 long gCode;
140 unsigned char *buf;
141 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
142
143 // If the syscall returns a suggestion for the size,
144 // then we don't have to search for an appropriate size.
145 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf);
146 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
147 "initial getaffinity call returned %ld errno = %d\n",
148 gCode, errno));
149
150 if (gCode < 0 && errno != EINVAL) {
151 // System call not supported
152 if (verbose ||
153 (warnings && (type != affinity_none) && (type != affinity_default) &&
154 (type != affinity_disabled))) {
155 int error = errno;
156 kmp_msg_t err_code = KMP_ERR(error);
157 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
158 err_code, __kmp_msg_null);
159 if (__kmp_generate_warnings == kmp_warnings_off) {
160 __kmp_str_free(&err_code.str);
161 }
162 }
163 KMP_AFFINITY_DISABLE();
164 KMP_INTERNAL_FREE(buf);
165 return;
166 } else if (gCode > 0) {
167 // The optimal situation: the OS returns the size of the buffer it expects.
168 KMP_AFFINITY_ENABLE(gCode);
169 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
170 "affinity supported (mask size %d)\n",
171 (int)__kmp_affin_mask_size));
172 KMP_INTERNAL_FREE(buf);
173 return;
174 }
175
176 // Call the getaffinity system call repeatedly with increasing set sizes
177 // until we succeed, or reach an upper bound on the search.
178 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
179 "searching for proper set size\n"));
180 int size;
181 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
182 gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
183 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
184 "getaffinity for mask size %ld returned %ld errno = %d\n",
185 size, gCode, errno));
186
187 if (gCode < 0) {
188 if (errno == ENOSYS) {
189 // We shouldn't get here
190 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
191 "inconsistent OS call behavior: errno == ENOSYS for mask "
192 "size %d\n",
193 size));
194 if (verbose ||
195 (warnings && (type != affinity_none) &&
196 (type != affinity_default) && (type != affinity_disabled))) {
197 int error = errno;
198 kmp_msg_t err_code = KMP_ERR(error);
199 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
200 err_code, __kmp_msg_null);
201 if (__kmp_generate_warnings == kmp_warnings_off) {
202 __kmp_str_free(&err_code.str);
203 }
204 }
205 KMP_AFFINITY_DISABLE();
206 KMP_INTERNAL_FREE(buf);
207 return;
208 }
209 continue;
210 }
211
212 KMP_AFFINITY_ENABLE(gCode);
213 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
214 "affinity supported (mask size %d)\n",
215 (int)__kmp_affin_mask_size));
216 KMP_INTERNAL_FREE(buf);
217 return;
218 }
219#elif KMP_OS_FREEBSD
220 long gCode;
221 unsigned char *buf;
222 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
223 gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT,
224 reinterpret_cast<cpuset_t *>(buf));
225 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
226 "initial getaffinity call returned %d errno = %d\n",
227 gCode, errno));
228 if (gCode == 0) {
229 KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT);
230 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
231 "affinity supported (mask size %d)\n",
232 (int)__kmp_affin_mask_size));
233 KMP_INTERNAL_FREE(buf);
234 return;
235 }
236#endif
237 KMP_INTERNAL_FREE(buf);
238
239 // Affinity is not supported
240 KMP_AFFINITY_DISABLE();
241 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
242 "cannot determine mask size - affinity not supported\n"));
243 if (verbose || (warnings && (type != affinity_none) &&
244 (type != affinity_default) && (type != affinity_disabled))) {
245 KMP_WARNING(AffCantGetMaskSize, env_var);
246 }
247}
248
249#endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
250
251#if KMP_USE_FUTEX
252
253int __kmp_futex_determine_capable() {
254 int loc = 0;
255 long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
256 int retval = (rc == 0) || (errno != ENOSYS);
257
258 KA_TRACE(10,
259 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
260 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
261 retval ? "" : " not"));
262
263 return retval;
264}
265
266#endif // KMP_USE_FUTEX
267
268#if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS)
269/* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
270 use compare_and_store for these routines */
271
272kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
273 kmp_int8 old_value, new_value;
274
275 old_value = TCR_1(*p);
276 new_value = old_value | d;
277
278 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
279 KMP_CPU_PAUSE();
280 old_value = TCR_1(*p);
281 new_value = old_value | d;
282 }
283 return old_value;
284}
285
286kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
287 kmp_int8 old_value, new_value;
288
289 old_value = TCR_1(*p);
290 new_value = old_value & d;
291
292 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
293 KMP_CPU_PAUSE();
294 old_value = TCR_1(*p);
295 new_value = old_value & d;
296 }
297 return old_value;
298}
299
300kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
301 kmp_uint32 old_value, new_value;
302
303 old_value = TCR_4(*p);
304 new_value = old_value | d;
305
306 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
307 KMP_CPU_PAUSE();
308 old_value = TCR_4(*p);
309 new_value = old_value | d;
310 }
311 return old_value;
312}
313
314kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
315 kmp_uint32 old_value, new_value;
316
317 old_value = TCR_4(*p);
318 new_value = old_value & d;
319
320 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
321 KMP_CPU_PAUSE();
322 old_value = TCR_4(*p);
323 new_value = old_value & d;
324 }
325 return old_value;
326}
327
328#if KMP_ARCH_X86
329kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
330 kmp_int8 old_value, new_value;
331
332 old_value = TCR_1(*p);
333 new_value = old_value + d;
334
335 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
336 KMP_CPU_PAUSE();
337 old_value = TCR_1(*p);
338 new_value = old_value + d;
339 }
340 return old_value;
341}
342
343kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
344 kmp_int64 old_value, new_value;
345
346 old_value = TCR_8(*p);
347 new_value = old_value + d;
348
349 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
350 KMP_CPU_PAUSE();
351 old_value = TCR_8(*p);
352 new_value = old_value + d;
353 }
354 return old_value;
355}
356#endif /* KMP_ARCH_X86 */
357
358kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
359 kmp_uint64 old_value, new_value;
360
361 old_value = TCR_8(*p);
362 new_value = old_value | d;
363 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
364 KMP_CPU_PAUSE();
365 old_value = TCR_8(*p);
366 new_value = old_value | d;
367 }
368 return old_value;
369}
370
371kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
372 kmp_uint64 old_value, new_value;
373
374 old_value = TCR_8(*p);
375 new_value = old_value & d;
376 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
377 KMP_CPU_PAUSE();
378 old_value = TCR_8(*p);
379 new_value = old_value & d;
380 }
381 return old_value;
382}
383
384#endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
385
386void __kmp_terminate_thread(int gtid) {
387 int status;
388 kmp_info_t *th = __kmp_threads[gtid];
389
390 if (!th)
391 return;
392
393#ifdef KMP_CANCEL_THREADS
394 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
395 status = pthread_cancel(th->th.th_info.ds.ds_thread);
396 if (status != 0 && status != ESRCH) {
397 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
398 __kmp_msg_null);
399 }
400#endif
401 KMP_YIELD(TRUE);
402} //
403
404/* Set thread stack info according to values returned by pthread_getattr_np().
405 If values are unreasonable, assume call failed and use incremental stack
406 refinement method instead. Returns TRUE if the stack parameters could be
407 determined exactly, FALSE if incremental refinement is necessary. */
408static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
409 int stack_data;
410#if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
411 KMP_OS_HURD
412 pthread_attr_t attr;
413 int status;
414 size_t size = 0;
415 void *addr = 0;
416
417 /* Always do incremental stack refinement for ubermaster threads since the
418 initial thread stack range can be reduced by sibling thread creation so
419 pthread_attr_getstack may cause thread gtid aliasing */
420 if (!KMP_UBER_GTID(gtid)) {
421
422 /* Fetch the real thread attributes */
423 status = pthread_attr_init(&attr);
424 KMP_CHECK_SYSFAIL("pthread_attr_init", status);
425#if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
426 status = pthread_attr_get_np(pthread_self(), &attr);
427 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
428#else
429 status = pthread_getattr_np(pthread_self(), &attr);
430 KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
431#endif
432 status = pthread_attr_getstack(&attr, &addr, &size);
433 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
434 KA_TRACE(60,
435 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
436 " %lu, low addr: %p\n",
437 gtid, size, addr));
438 status = pthread_attr_destroy(&attr);
439 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
440 }
441
442 if (size != 0 && addr != 0) { // was stack parameter determination successful?
443 /* Store the correct base and size */
444 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
445 TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
446 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
447 return TRUE;
448 }
449#endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD \
450 || KMP_OS_HURD */
451 /* Use incremental refinement starting from initial conservative estimate */
452 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
453 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
454 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
455 return FALSE;
456}
457
458static void *__kmp_launch_worker(void *thr) {
459 int status, old_type, old_state;
460#ifdef KMP_BLOCK_SIGNALS
461 sigset_t new_set, old_set;
462#endif /* KMP_BLOCK_SIGNALS */
463 void *exit_val;
464#if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
465 KMP_OS_OPENBSD || KMP_OS_HURD
466 void *volatile padding = 0;
467#endif
468 int gtid;
469
470 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
471 __kmp_gtid_set_specific(gtid);
472#ifdef KMP_TDATA_GTID
473 __kmp_gtid = gtid;
474#endif
475#if KMP_STATS_ENABLED
476 // set thread local index to point to thread-specific stats
477 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
478 __kmp_stats_thread_ptr->startLife();
479 KMP_SET_THREAD_STATE(IDLE);
481#endif
482
483#if USE_ITT_BUILD
484 __kmp_itt_thread_name(gtid);
485#endif /* USE_ITT_BUILD */
486
487#if KMP_AFFINITY_SUPPORTED
488 __kmp_affinity_set_init_mask(gtid, FALSE);
489#endif
490
491#ifdef KMP_CANCEL_THREADS
492 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
493 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
494 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
495 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
496 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
497#endif
498
499#if KMP_ARCH_X86 || KMP_ARCH_X86_64
500 // Set FP control regs to be a copy of the parallel initialization thread's.
501 __kmp_clear_x87_fpu_status_word();
502 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
503 __kmp_load_mxcsr(&__kmp_init_mxcsr);
504#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
505
506#ifdef KMP_BLOCK_SIGNALS
507 status = sigfillset(&new_set);
508 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
509 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
510 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
511#endif /* KMP_BLOCK_SIGNALS */
512
513#if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
514 KMP_OS_OPENBSD
515 if (__kmp_stkoffset > 0 && gtid > 0) {
516 padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
517 (void)padding;
518 }
519#endif
520
521 KMP_MB();
522 __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
523
524 __kmp_check_stack_overlap((kmp_info_t *)thr);
525
526 exit_val = __kmp_launch_thread((kmp_info_t *)thr);
527
528#ifdef KMP_BLOCK_SIGNALS
529 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
530 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
531#endif /* KMP_BLOCK_SIGNALS */
532
533 return exit_val;
534}
535
536#if KMP_USE_MONITOR
537/* The monitor thread controls all of the threads in the complex */
538
539static void *__kmp_launch_monitor(void *thr) {
540 int status, old_type, old_state;
541#ifdef KMP_BLOCK_SIGNALS
542 sigset_t new_set;
543#endif /* KMP_BLOCK_SIGNALS */
544 struct timespec interval;
545
546 KMP_MB(); /* Flush all pending memory write invalidates. */
547
548 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
549
550 /* register us as the monitor thread */
551 __kmp_gtid_set_specific(KMP_GTID_MONITOR);
552#ifdef KMP_TDATA_GTID
553 __kmp_gtid = KMP_GTID_MONITOR;
554#endif
555
556 KMP_MB();
557
558#if USE_ITT_BUILD
559 // Instruct Intel(R) Threading Tools to ignore monitor thread.
560 __kmp_itt_thread_ignore();
561#endif /* USE_ITT_BUILD */
562
563 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
564 (kmp_info_t *)thr);
565
566 __kmp_check_stack_overlap((kmp_info_t *)thr);
567
568#ifdef KMP_CANCEL_THREADS
569 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
570 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
571 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
572 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
573 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
574#endif
575
576#if KMP_REAL_TIME_FIX
577 // This is a potential fix which allows application with real-time scheduling
578 // policy work. However, decision about the fix is not made yet, so it is
579 // disabled by default.
580 { // Are program started with real-time scheduling policy?
581 int sched = sched_getscheduler(0);
582 if (sched == SCHED_FIFO || sched == SCHED_RR) {
583 // Yes, we are a part of real-time application. Try to increase the
584 // priority of the monitor.
585 struct sched_param param;
586 int max_priority = sched_get_priority_max(sched);
587 int rc;
588 KMP_WARNING(RealTimeSchedNotSupported);
589 sched_getparam(0, &param);
590 if (param.sched_priority < max_priority) {
591 param.sched_priority += 1;
592 rc = sched_setscheduler(0, sched, &param);
593 if (rc != 0) {
594 int error = errno;
595 kmp_msg_t err_code = KMP_ERR(error);
596 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
597 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
598 if (__kmp_generate_warnings == kmp_warnings_off) {
599 __kmp_str_free(&err_code.str);
600 }
601 }
602 } else {
603 // We cannot abort here, because number of CPUs may be enough for all
604 // the threads, including the monitor thread, so application could
605 // potentially work...
606 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
607 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
608 __kmp_msg_null);
609 }
610 }
611 // AC: free thread that waits for monitor started
612 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
613 }
614#endif // KMP_REAL_TIME_FIX
615
616 KMP_MB(); /* Flush all pending memory write invalidates. */
617
618 if (__kmp_monitor_wakeups == 1) {
619 interval.tv_sec = 1;
620 interval.tv_nsec = 0;
621 } else {
622 interval.tv_sec = 0;
623 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
624 }
625
626 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
627
628 while (!TCR_4(__kmp_global.g.g_done)) {
629 struct timespec now;
630 struct timeval tval;
631
632 /* This thread monitors the state of the system */
633
634 KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
635
636 status = gettimeofday(&tval, NULL);
637 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
638 TIMEVAL_TO_TIMESPEC(&tval, &now);
639
640 now.tv_sec += interval.tv_sec;
641 now.tv_nsec += interval.tv_nsec;
642
643 if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
644 now.tv_sec += 1;
645 now.tv_nsec -= KMP_NSEC_PER_SEC;
646 }
647
648 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
649 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
650 // AC: the monitor should not fall asleep if g_done has been set
651 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
652 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
653 &__kmp_wait_mx.m_mutex, &now);
654 if (status != 0) {
655 if (status != ETIMEDOUT && status != EINTR) {
656 KMP_SYSFAIL("pthread_cond_timedwait", status);
657 }
658 }
659 }
660 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
661 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
662
663 TCW_4(__kmp_global.g.g_time.dt.t_value,
664 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
665
666 KMP_MB(); /* Flush all pending memory write invalidates. */
667 }
668
669 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
670
671#ifdef KMP_BLOCK_SIGNALS
672 status = sigfillset(&new_set);
673 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
674 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
675 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
676#endif /* KMP_BLOCK_SIGNALS */
677
678 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
679
680 if (__kmp_global.g.g_abort != 0) {
681 /* now we need to terminate the worker threads */
682 /* the value of t_abort is the signal we caught */
683
684 int gtid;
685
686 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
687 __kmp_global.g.g_abort));
688
689 /* terminate the OpenMP worker threads */
690 /* TODO this is not valid for sibling threads!!
691 * the uber master might not be 0 anymore.. */
692 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
693 __kmp_terminate_thread(gtid);
694
695 __kmp_cleanup();
696
697 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
698 __kmp_global.g.g_abort));
699
700 if (__kmp_global.g.g_abort > 0)
701 raise(__kmp_global.g.g_abort);
702 }
703
704 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
705
706 return thr;
707}
708#endif // KMP_USE_MONITOR
709
710void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
711 pthread_t handle;
712 pthread_attr_t thread_attr;
713 int status;
714
715 th->th.th_info.ds.ds_gtid = gtid;
716
717#if KMP_STATS_ENABLED
718 // sets up worker thread stats
719 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
720
721 // th->th.th_stats is used to transfer thread-specific stats-pointer to
722 // __kmp_launch_worker. So when thread is created (goes into
723 // __kmp_launch_worker) it will set its thread local pointer to
724 // th->th.th_stats
725 if (!KMP_UBER_GTID(gtid)) {
726 th->th.th_stats = __kmp_stats_list->push_back(gtid);
727 } else {
728 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
729 // so set the th->th.th_stats field to it.
730 th->th.th_stats = __kmp_stats_thread_ptr;
731 }
732 __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
733
734#endif // KMP_STATS_ENABLED
735
736 if (KMP_UBER_GTID(gtid)) {
737 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
738 th->th.th_info.ds.ds_thread = pthread_self();
739 __kmp_set_stack_info(gtid, th);
740 __kmp_check_stack_overlap(th);
741 return;
742 }
743
744 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
745
746 KMP_MB(); /* Flush all pending memory write invalidates. */
747
748#ifdef KMP_THREAD_ATTR
749 status = pthread_attr_init(&thread_attr);
750 if (status != 0) {
751 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
752 }
753 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
754 if (status != 0) {
755 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
756 }
757
758 /* Set stack size for this thread now.
759 The multiple of 2 is there because on some machines, requesting an unusual
760 stacksize causes the thread to have an offset before the dummy alloca()
761 takes place to create the offset. Since we want the user to have a
762 sufficient stacksize AND support a stack offset, we alloca() twice the
763 offset so that the upcoming alloca() does not eliminate any premade offset,
764 and also gives the user the stack space they requested for all threads */
765 stack_size += gtid * __kmp_stkoffset * 2;
766
767 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
768 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
769 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
770
771#ifdef _POSIX_THREAD_ATTR_STACKSIZE
772 status = pthread_attr_setstacksize(&thread_attr, stack_size);
773#ifdef KMP_BACKUP_STKSIZE
774 if (status != 0) {
775 if (!__kmp_env_stksize) {
776 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
777 __kmp_stksize = KMP_BACKUP_STKSIZE;
778 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
779 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
780 "bytes\n",
781 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
782 status = pthread_attr_setstacksize(&thread_attr, stack_size);
783 }
784 }
785#endif /* KMP_BACKUP_STKSIZE */
786 if (status != 0) {
787 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
788 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
789 }
790#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
791
792#endif /* KMP_THREAD_ATTR */
793
794 status =
795 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
796 if (status != 0 || !handle) { // ??? Why do we check handle??
797#ifdef _POSIX_THREAD_ATTR_STACKSIZE
798 if (status == EINVAL) {
799 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
800 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
801 }
802 if (status == ENOMEM) {
803 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
804 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
805 }
806#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
807 if (status == EAGAIN) {
808 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
809 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
810 }
811 KMP_SYSFAIL("pthread_create", status);
812 }
813
814 th->th.th_info.ds.ds_thread = handle;
815
816#ifdef KMP_THREAD_ATTR
817 status = pthread_attr_destroy(&thread_attr);
818 if (status) {
819 kmp_msg_t err_code = KMP_ERR(status);
820 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
821 __kmp_msg_null);
822 if (__kmp_generate_warnings == kmp_warnings_off) {
823 __kmp_str_free(&err_code.str);
824 }
825 }
826#endif /* KMP_THREAD_ATTR */
827
828 KMP_MB(); /* Flush all pending memory write invalidates. */
829
830 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
831
832} // __kmp_create_worker
833
834#if KMP_USE_MONITOR
835void __kmp_create_monitor(kmp_info_t *th) {
836 pthread_t handle;
837 pthread_attr_t thread_attr;
838 size_t size;
839 int status;
840 int auto_adj_size = FALSE;
841
842 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
843 // We don't need monitor thread in case of MAX_BLOCKTIME
844 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
845 "MAX blocktime\n"));
846 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
847 th->th.th_info.ds.ds_gtid = 0;
848 return;
849 }
850 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
851
852 KMP_MB(); /* Flush all pending memory write invalidates. */
853
854 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
855 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
856#if KMP_REAL_TIME_FIX
857 TCW_4(__kmp_global.g.g_time.dt.t_value,
858 -1); // Will use it for synchronization a bit later.
859#else
860 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
861#endif // KMP_REAL_TIME_FIX
862
863#ifdef KMP_THREAD_ATTR
864 if (__kmp_monitor_stksize == 0) {
865 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
866 auto_adj_size = TRUE;
867 }
868 status = pthread_attr_init(&thread_attr);
869 if (status != 0) {
870 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
871 }
872 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
873 if (status != 0) {
874 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
875 }
876
877#ifdef _POSIX_THREAD_ATTR_STACKSIZE
878 status = pthread_attr_getstacksize(&thread_attr, &size);
879 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
880#else
881 size = __kmp_sys_min_stksize;
882#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
883#endif /* KMP_THREAD_ATTR */
884
885 if (__kmp_monitor_stksize == 0) {
886 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
887 }
888 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
889 __kmp_monitor_stksize = __kmp_sys_min_stksize;
890 }
891
892 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
893 "requested stacksize = %lu bytes\n",
894 size, __kmp_monitor_stksize));
895
896retry:
897
898/* Set stack size for this thread now. */
899#ifdef _POSIX_THREAD_ATTR_STACKSIZE
900 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
901 __kmp_monitor_stksize));
902 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
903 if (status != 0) {
904 if (auto_adj_size) {
905 __kmp_monitor_stksize *= 2;
906 goto retry;
907 }
908 kmp_msg_t err_code = KMP_ERR(status);
909 __kmp_msg(kmp_ms_warning, // should this be fatal? BB
910 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
911 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
912 if (__kmp_generate_warnings == kmp_warnings_off) {
913 __kmp_str_free(&err_code.str);
914 }
915 }
916#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
917
918 status =
919 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
920
921 if (status != 0) {
922#ifdef _POSIX_THREAD_ATTR_STACKSIZE
923 if (status == EINVAL) {
924 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
925 __kmp_monitor_stksize *= 2;
926 goto retry;
927 }
928 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
929 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
930 __kmp_msg_null);
931 }
932 if (status == ENOMEM) {
933 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
934 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
935 __kmp_msg_null);
936 }
937#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
938 if (status == EAGAIN) {
939 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
940 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
941 }
942 KMP_SYSFAIL("pthread_create", status);
943 }
944
945 th->th.th_info.ds.ds_thread = handle;
946
947#if KMP_REAL_TIME_FIX
948 // Wait for the monitor thread is really started and set its *priority*.
949 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
950 sizeof(__kmp_global.g.g_time.dt.t_value));
951 __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1,
952 &__kmp_neq_4, NULL);
953#endif // KMP_REAL_TIME_FIX
954
955#ifdef KMP_THREAD_ATTR
956 status = pthread_attr_destroy(&thread_attr);
957 if (status != 0) {
958 kmp_msg_t err_code = KMP_ERR(status);
959 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
960 __kmp_msg_null);
961 if (__kmp_generate_warnings == kmp_warnings_off) {
962 __kmp_str_free(&err_code.str);
963 }
964 }
965#endif
966
967 KMP_MB(); /* Flush all pending memory write invalidates. */
968
969 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
970 th->th.th_info.ds.ds_thread));
971
972} // __kmp_create_monitor
973#endif // KMP_USE_MONITOR
974
975void __kmp_exit_thread(int exit_status) {
976 pthread_exit((void *)(intptr_t)exit_status);
977} // __kmp_exit_thread
978
979#if KMP_USE_MONITOR
980void __kmp_resume_monitor();
981
982extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
983 int status;
984 void *exit_val;
985
986 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
987 " %#.8lx\n",
988 th->th.th_info.ds.ds_thread));
989
990 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
991 // If both tid and gtid are 0, it means the monitor did not ever start.
992 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
993 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
994 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
995 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
996 return;
997 }
998
999 KMP_MB(); /* Flush all pending memory write invalidates. */
1000
1001 /* First, check to see whether the monitor thread exists to wake it up. This
1002 is to avoid performance problem when the monitor sleeps during
1003 blocktime-size interval */
1004
1005 status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1006 if (status != ESRCH) {
1007 __kmp_resume_monitor(); // Wake up the monitor thread
1008 }
1009 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1010 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1011 if (exit_val != th) {
1012 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1013 }
1014
1015 th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1016 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1017
1018 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1019 " %#.8lx\n",
1020 th->th.th_info.ds.ds_thread));
1021
1022 KMP_MB(); /* Flush all pending memory write invalidates. */
1023}
1024#else
1025// Empty symbol to export (see exports_so.txt) when
1026// monitor thread feature is disabled
1027extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
1028 (void)th;
1029}
1030#endif // KMP_USE_MONITOR
1031
1032void __kmp_reap_worker(kmp_info_t *th) {
1033 int status;
1034 void *exit_val;
1035
1036 KMP_MB(); /* Flush all pending memory write invalidates. */
1037
1038 KA_TRACE(
1039 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1040
1041 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1042#ifdef KMP_DEBUG
1043 /* Don't expose these to the user until we understand when they trigger */
1044 if (status != 0) {
1045 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1046 }
1047 if (exit_val != th) {
1048 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1049 "exit_val = %p\n",
1050 th->th.th_info.ds.ds_gtid, exit_val));
1051 }
1052#else
1053 (void)status; // unused variable
1054#endif /* KMP_DEBUG */
1055
1056 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1057 th->th.th_info.ds.ds_gtid));
1058
1059 KMP_MB(); /* Flush all pending memory write invalidates. */
1060}
1061
1062#if KMP_HANDLE_SIGNALS
1063
1064static void __kmp_null_handler(int signo) {
1065 // Do nothing, for doing SIG_IGN-type actions.
1066} // __kmp_null_handler
1067
1068static void __kmp_team_handler(int signo) {
1069 if (__kmp_global.g.g_abort == 0) {
1070/* Stage 1 signal handler, let's shut down all of the threads */
1071#ifdef KMP_DEBUG
1072 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1073#endif
1074 switch (signo) {
1075 case SIGHUP:
1076 case SIGINT:
1077 case SIGQUIT:
1078 case SIGILL:
1079 case SIGABRT:
1080 case SIGFPE:
1081 case SIGBUS:
1082 case SIGSEGV:
1083#ifdef SIGSYS
1084 case SIGSYS:
1085#endif
1086 case SIGTERM:
1087 if (__kmp_debug_buf) {
1088 __kmp_dump_debug_buffer();
1089 }
1090 __kmp_unregister_library(); // cleanup shared memory
1091 KMP_MB(); // Flush all pending memory write invalidates.
1092 TCW_4(__kmp_global.g.g_abort, signo);
1093 KMP_MB(); // Flush all pending memory write invalidates.
1094 TCW_4(__kmp_global.g.g_done, TRUE);
1095 KMP_MB(); // Flush all pending memory write invalidates.
1096 break;
1097 default:
1098#ifdef KMP_DEBUG
1099 __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1100#endif
1101 break;
1102 }
1103 }
1104} // __kmp_team_handler
1105
1106static void __kmp_sigaction(int signum, const struct sigaction *act,
1107 struct sigaction *oldact) {
1108 int rc = sigaction(signum, act, oldact);
1109 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1110}
1111
1112static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1113 int parallel_init) {
1114 KMP_MB(); // Flush all pending memory write invalidates.
1115 KB_TRACE(60,
1116 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1117 if (parallel_init) {
1118 struct sigaction new_action;
1119 struct sigaction old_action;
1120 new_action.sa_handler = handler_func;
1121 new_action.sa_flags = 0;
1122 sigfillset(&new_action.sa_mask);
1123 __kmp_sigaction(sig, &new_action, &old_action);
1124 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1125 sigaddset(&__kmp_sigset, sig);
1126 } else {
1127 // Restore/keep user's handler if one previously installed.
1128 __kmp_sigaction(sig, &old_action, NULL);
1129 }
1130 } else {
1131 // Save initial/system signal handlers to see if user handlers installed.
1132 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1133 }
1134 KMP_MB(); // Flush all pending memory write invalidates.
1135} // __kmp_install_one_handler
1136
1137static void __kmp_remove_one_handler(int sig) {
1138 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1139 if (sigismember(&__kmp_sigset, sig)) {
1140 struct sigaction old;
1141 KMP_MB(); // Flush all pending memory write invalidates.
1142 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1143 if ((old.sa_handler != __kmp_team_handler) &&
1144 (old.sa_handler != __kmp_null_handler)) {
1145 // Restore the users signal handler.
1146 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1147 "restoring: sig=%d\n",
1148 sig));
1149 __kmp_sigaction(sig, &old, NULL);
1150 }
1151 sigdelset(&__kmp_sigset, sig);
1152 KMP_MB(); // Flush all pending memory write invalidates.
1153 }
1154} // __kmp_remove_one_handler
1155
1156void __kmp_install_signals(int parallel_init) {
1157 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1158 if (__kmp_handle_signals || !parallel_init) {
1159 // If ! parallel_init, we do not install handlers, just save original
1160 // handlers. Let us do it even __handle_signals is 0.
1161 sigemptyset(&__kmp_sigset);
1162 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1163 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1164 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1165 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1166 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1167 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1168 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1169 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1170#ifdef SIGSYS
1171 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1172#endif // SIGSYS
1173 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1174#ifdef SIGPIPE
1175 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1176#endif // SIGPIPE
1177 }
1178} // __kmp_install_signals
1179
1180void __kmp_remove_signals(void) {
1181 int sig;
1182 KB_TRACE(10, ("__kmp_remove_signals()\n"));
1183 for (sig = 1; sig < NSIG; ++sig) {
1184 __kmp_remove_one_handler(sig);
1185 }
1186} // __kmp_remove_signals
1187
1188#endif // KMP_HANDLE_SIGNALS
1189
1190void __kmp_enable(int new_state) {
1191#ifdef KMP_CANCEL_THREADS
1192 int status, old_state;
1193 status = pthread_setcancelstate(new_state, &old_state);
1194 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1195 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1196#endif
1197}
1198
1199void __kmp_disable(int *old_state) {
1200#ifdef KMP_CANCEL_THREADS
1201 int status;
1202 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1203 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1204#endif
1205}
1206
1207static void __kmp_atfork_prepare(void) {
1208 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1209 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1210}
1211
1212static void __kmp_atfork_parent(void) {
1213 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1214 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1215}
1216
1217/* Reset the library so execution in the child starts "all over again" with
1218 clean data structures in initial states. Don't worry about freeing memory
1219 allocated by parent, just abandon it to be safe. */
1220static void __kmp_atfork_child(void) {
1221 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1222 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1223 /* TODO make sure this is done right for nested/sibling */
1224 // ATT: Memory leaks are here? TODO: Check it and fix.
1225 /* KMP_ASSERT( 0 ); */
1226
1227 ++__kmp_fork_count;
1228
1229#if KMP_AFFINITY_SUPPORTED
1230#if KMP_OS_LINUX || KMP_OS_FREEBSD
1231 // reset the affinity in the child to the initial thread
1232 // affinity in the parent
1233 kmp_set_thread_affinity_mask_initial();
1234#endif
1235 // Set default not to bind threads tightly in the child (we're expecting
1236 // over-subscription after the fork and this can improve things for
1237 // scripting languages that use OpenMP inside process-parallel code).
1238 if (__kmp_nested_proc_bind.bind_types != NULL) {
1239 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1240 }
1241 for (kmp_affinity_t *affinity : __kmp_affinities)
1242 *affinity = KMP_AFFINITY_INIT(affinity->env_var);
1243 __kmp_affin_fullMask = nullptr;
1244 __kmp_affin_origMask = nullptr;
1245 __kmp_topology = nullptr;
1246#endif // KMP_AFFINITY_SUPPORTED
1247
1248#if KMP_USE_MONITOR
1249 __kmp_init_monitor = 0;
1250#endif
1251 __kmp_init_parallel = FALSE;
1252 __kmp_init_middle = FALSE;
1253 __kmp_init_serial = FALSE;
1254 TCW_4(__kmp_init_gtid, FALSE);
1255 __kmp_init_common = FALSE;
1256
1257 TCW_4(__kmp_init_user_locks, FALSE);
1258#if !KMP_USE_DYNAMIC_LOCK
1259 __kmp_user_lock_table.used = 1;
1260 __kmp_user_lock_table.allocated = 0;
1261 __kmp_user_lock_table.table = NULL;
1262 __kmp_lock_blocks = NULL;
1263#endif
1264
1265 __kmp_all_nth = 0;
1266 TCW_4(__kmp_nth, 0);
1267
1268 __kmp_thread_pool = NULL;
1269 __kmp_thread_pool_insert_pt = NULL;
1270 __kmp_team_pool = NULL;
1271
1272 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1273 here so threadprivate doesn't use stale data */
1274 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1275 __kmp_threadpriv_cache_list));
1276
1277 while (__kmp_threadpriv_cache_list != NULL) {
1278
1279 if (*__kmp_threadpriv_cache_list->addr != NULL) {
1280 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1281 &(*__kmp_threadpriv_cache_list->addr)));
1282
1283 *__kmp_threadpriv_cache_list->addr = NULL;
1284 }
1285 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1286 }
1287
1288 __kmp_init_runtime = FALSE;
1289
1290 /* reset statically initialized locks */
1291 __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1292 __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1293 __kmp_init_bootstrap_lock(&__kmp_console_lock);
1294 __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1295
1296#if USE_ITT_BUILD
1297 __kmp_itt_reset(); // reset ITT's global state
1298#endif /* USE_ITT_BUILD */
1299
1300 {
1301 // Child process often get terminated without any use of OpenMP. That might
1302 // cause mapped shared memory file to be left unattended. Thus we postpone
1303 // library registration till middle initialization in the child process.
1304 __kmp_need_register_serial = FALSE;
1305 __kmp_serial_initialize();
1306 }
1307
1308 /* This is necessary to make sure no stale data is left around */
1309 /* AC: customers complain that we use unsafe routines in the atfork
1310 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1311 in dynamic_link when check the presence of shared tbbmalloc library.
1312 Suggestion is to make the library initialization lazier, similar
1313 to what done for __kmpc_begin(). */
1314 // TODO: synchronize all static initializations with regular library
1315 // startup; look at kmp_global.cpp and etc.
1316 //__kmp_internal_begin ();
1317}
1318
1319void __kmp_register_atfork(void) {
1320 if (__kmp_need_register_atfork) {
1321 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1322 __kmp_atfork_child);
1323 KMP_CHECK_SYSFAIL("pthread_atfork", status);
1324 __kmp_need_register_atfork = FALSE;
1325 }
1326}
1327
1328void __kmp_suspend_initialize(void) {
1329 int status;
1330 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1331 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1332 status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1333 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1334}
1335
1336void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1337 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count);
1338 int new_value = __kmp_fork_count + 1;
1339 // Return if already initialized
1340 if (old_value == new_value)
1341 return;
1342 // Wait, then return if being initialized
1343 if (old_value == -1 || !__kmp_atomic_compare_store(
1344 &th->th.th_suspend_init_count, old_value, -1)) {
1345 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) {
1346 KMP_CPU_PAUSE();
1347 }
1348 } else {
1349 // Claim to be the initializer and do initializations
1350 int status;
1351 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1352 &__kmp_suspend_cond_attr);
1353 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1354 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1355 &__kmp_suspend_mutex_attr);
1356 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1357 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value);
1358 }
1359}
1360
1361void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1362 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) {
1363 /* this means we have initialize the suspension pthread objects for this
1364 thread in this instance of the process */
1365 int status;
1366
1367 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1368 if (status != 0 && status != EBUSY) {
1369 KMP_SYSFAIL("pthread_cond_destroy", status);
1370 }
1371 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1372 if (status != 0 && status != EBUSY) {
1373 KMP_SYSFAIL("pthread_mutex_destroy", status);
1374 }
1375 --th->th.th_suspend_init_count;
1376 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) ==
1377 __kmp_fork_count);
1378 }
1379}
1380
1381// return true if lock obtained, false otherwise
1382int __kmp_try_suspend_mx(kmp_info_t *th) {
1383 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1384}
1385
1386void __kmp_lock_suspend_mx(kmp_info_t *th) {
1387 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1388 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1389}
1390
1391void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1392 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1393 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1394}
1395
1396/* This routine puts the calling thread to sleep after setting the
1397 sleep bit for the indicated flag variable to true. */
1398template <class C>
1399static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1400 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1401 kmp_info_t *th = __kmp_threads[th_gtid];
1402 int status;
1403 typename C::flag_t old_spin;
1404
1405 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1406 flag->get()));
1407
1408 __kmp_suspend_initialize_thread(th);
1409
1410 __kmp_lock_suspend_mx(th);
1411
1412 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1413 th_gtid, flag->get()));
1414
1415 /* TODO: shouldn't this use release semantics to ensure that
1416 __kmp_suspend_initialize_thread gets called first? */
1417 old_spin = flag->set_sleeping();
1418 TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1419 th->th.th_sleep_loc_type = flag->get_type();
1420 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1421 __kmp_pause_status != kmp_soft_paused) {
1422 flag->unset_sleeping();
1423 TCW_PTR(th->th.th_sleep_loc, NULL);
1424 th->th.th_sleep_loc_type = flag_unset;
1425 __kmp_unlock_suspend_mx(th);
1426 return;
1427 }
1428 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1429 " was %x\n",
1430 th_gtid, flag->get(), flag->load(), old_spin));
1431
1432 if (flag->done_check_val(old_spin) || flag->done_check()) {
1433 flag->unset_sleeping();
1434 TCW_PTR(th->th.th_sleep_loc, NULL);
1435 th->th.th_sleep_loc_type = flag_unset;
1436 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1437 "for spin(%p)\n",
1438 th_gtid, flag->get()));
1439 } else {
1440 /* Encapsulate in a loop as the documentation states that this may
1441 "with low probability" return when the condition variable has
1442 not been signaled or broadcast */
1443 int deactivated = FALSE;
1444
1445 while (flag->is_sleeping()) {
1446#ifdef DEBUG_SUSPEND
1447 char buffer[128];
1448 __kmp_suspend_count++;
1449 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1450 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1451 buffer);
1452#endif
1453 // Mark the thread as no longer active (only in the first iteration of the
1454 // loop).
1455 if (!deactivated) {
1456 th->th.th_active = FALSE;
1457 if (th->th.th_active_in_pool) {
1458 th->th.th_active_in_pool = FALSE;
1459 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1460 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1461 }
1462 deactivated = TRUE;
1463 }
1464
1465 KMP_DEBUG_ASSERT(th->th.th_sleep_loc);
1466 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type);
1467
1468#if USE_SUSPEND_TIMEOUT
1469 struct timespec now;
1470 struct timeval tval;
1471 int msecs;
1472
1473 status = gettimeofday(&tval, NULL);
1474 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1475 TIMEVAL_TO_TIMESPEC(&tval, &now);
1476
1477 msecs = (4 * __kmp_dflt_blocktime) + 200;
1478 now.tv_sec += msecs / 1000;
1479 now.tv_nsec += (msecs % 1000) * 1000;
1480
1481 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1482 "pthread_cond_timedwait\n",
1483 th_gtid));
1484 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1485 &th->th.th_suspend_mx.m_mutex, &now);
1486#else
1487 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1488 " pthread_cond_wait\n",
1489 th_gtid));
1490 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1491 &th->th.th_suspend_mx.m_mutex);
1492#endif // USE_SUSPEND_TIMEOUT
1493
1494 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1495 KMP_SYSFAIL("pthread_cond_wait", status);
1496 }
1497
1498 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type());
1499
1500 if (!flag->is_sleeping() &&
1501 ((status == EINTR) || (status == ETIMEDOUT))) {
1502 // if interrupt or timeout, and thread is no longer sleeping, we need to
1503 // make sure sleep_loc gets reset; however, this shouldn't be needed if
1504 // we woke up with resume
1505 flag->unset_sleeping();
1506 TCW_PTR(th->th.th_sleep_loc, NULL);
1507 th->th.th_sleep_loc_type = flag_unset;
1508 }
1509#ifdef KMP_DEBUG
1510 if (status == ETIMEDOUT) {
1511 if (flag->is_sleeping()) {
1512 KF_TRACE(100,
1513 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1514 } else {
1515 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1516 "not set!\n",
1517 th_gtid));
1518 TCW_PTR(th->th.th_sleep_loc, NULL);
1519 th->th.th_sleep_loc_type = flag_unset;
1520 }
1521 } else if (flag->is_sleeping()) {
1522 KF_TRACE(100,
1523 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1524 }
1525#endif
1526 } // while
1527
1528 // Mark the thread as active again (if it was previous marked as inactive)
1529 if (deactivated) {
1530 th->th.th_active = TRUE;
1531 if (TCR_4(th->th.th_in_pool)) {
1532 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1533 th->th.th_active_in_pool = TRUE;
1534 }
1535 }
1536 }
1537 // We may have had the loop variable set before entering the loop body;
1538 // so we need to reset sleep_loc.
1539 TCW_PTR(th->th.th_sleep_loc, NULL);
1540 th->th.th_sleep_loc_type = flag_unset;
1541
1542 KMP_DEBUG_ASSERT(!flag->is_sleeping());
1543 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc);
1544#ifdef DEBUG_SUSPEND
1545 {
1546 char buffer[128];
1547 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1548 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1549 buffer);
1550 }
1551#endif
1552
1553 __kmp_unlock_suspend_mx(th);
1554 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1555}
1556
1557template <bool C, bool S>
1558void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) {
1559 __kmp_suspend_template(th_gtid, flag);
1560}
1561template <bool C, bool S>
1562void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) {
1563 __kmp_suspend_template(th_gtid, flag);
1564}
1565template <bool C, bool S>
1566void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) {
1567 __kmp_suspend_template(th_gtid, flag);
1568}
1569void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1570 __kmp_suspend_template(th_gtid, flag);
1571}
1572
1573template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *);
1574template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *);
1575template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *);
1576template void
1577__kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1578template void
1579__kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *);
1580
1581/* This routine signals the thread specified by target_gtid to wake up
1582 after setting the sleep bit indicated by the flag argument to FALSE.
1583 The target thread must already have called __kmp_suspend_template() */
1584template <class C>
1585static inline void __kmp_resume_template(int target_gtid, C *flag) {
1586 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1587 kmp_info_t *th = __kmp_threads[target_gtid];
1588 int status;
1589
1590#ifdef KMP_DEBUG
1591 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1592#endif
1593
1594 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1595 gtid, target_gtid));
1596 KMP_DEBUG_ASSERT(gtid != target_gtid);
1597
1598 __kmp_suspend_initialize_thread(th);
1599
1600 __kmp_lock_suspend_mx(th);
1601
1602 if (!flag || flag != th->th.th_sleep_loc) {
1603 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a
1604 // different location; wake up at new location
1605 flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1606 }
1607
1608 // First, check if the flag is null or its type has changed. If so, someone
1609 // else woke it up.
1610 if (!flag) { // Thread doesn't appear to be sleeping on anything
1611 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1612 "awake: flag(%p)\n",
1613 gtid, target_gtid, (void *)NULL));
1614 __kmp_unlock_suspend_mx(th);
1615 return;
1616 } else if (flag->get_type() != th->th.th_sleep_loc_type) {
1617 // Flag type does not appear to match this function template; possibly the
1618 // thread is sleeping on something else. Try null resume again.
1619 KF_TRACE(
1620 5,
1621 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), "
1622 "spin(%p) type=%d ptr_type=%d\n",
1623 gtid, target_gtid, flag, flag->get(), flag->get_type(),
1624 th->th.th_sleep_loc_type));
1625 __kmp_unlock_suspend_mx(th);
1626 __kmp_null_resume_wrapper(th);
1627 return;
1628 } else { // if multiple threads are sleeping, flag should be internally
1629 // referring to a specific thread here
1630 if (!flag->is_sleeping()) {
1631 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1632 "awake: flag(%p): %u\n",
1633 gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1634 __kmp_unlock_suspend_mx(th);
1635 return;
1636 }
1637 }
1638 KMP_DEBUG_ASSERT(flag);
1639 flag->unset_sleeping();
1640 TCW_PTR(th->th.th_sleep_loc, NULL);
1641 th->th.th_sleep_loc_type = flag_unset;
1642
1643 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1644 "sleep bit for flag's loc(%p): %u\n",
1645 gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1646
1647#ifdef DEBUG_SUSPEND
1648 {
1649 char buffer[128];
1650 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1651 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1652 target_gtid, buffer);
1653 }
1654#endif
1655 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1656 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1657 __kmp_unlock_suspend_mx(th);
1658 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1659 " for T#%d\n",
1660 gtid, target_gtid));
1661}
1662
1663template <bool C, bool S>
1664void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) {
1665 __kmp_resume_template(target_gtid, flag);
1666}
1667template <bool C, bool S>
1668void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) {
1669 __kmp_resume_template(target_gtid, flag);
1670}
1671template <bool C, bool S>
1672void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) {
1673 __kmp_resume_template(target_gtid, flag);
1674}
1675void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1676 __kmp_resume_template(target_gtid, flag);
1677}
1678
1679template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *);
1680template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *);
1681template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *);
1682template void
1683__kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1684
1685#if KMP_USE_MONITOR
1686void __kmp_resume_monitor() {
1687 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1688 int status;
1689#ifdef KMP_DEBUG
1690 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1691 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1692 KMP_GTID_MONITOR));
1693 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1694#endif
1695 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1696 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1697#ifdef DEBUG_SUSPEND
1698 {
1699 char buffer[128];
1700 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1701 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1702 KMP_GTID_MONITOR, buffer);
1703 }
1704#endif
1705 status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1706 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1707 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1708 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1709 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1710 " for T#%d\n",
1711 gtid, KMP_GTID_MONITOR));
1712}
1713#endif // KMP_USE_MONITOR
1714
1715void __kmp_yield() { sched_yield(); }
1716
1717void __kmp_gtid_set_specific(int gtid) {
1718 if (__kmp_init_gtid) {
1719 int status;
1720 status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1721 (void *)(intptr_t)(gtid + 1));
1722 KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1723 } else {
1724 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1725 }
1726}
1727
1728int __kmp_gtid_get_specific() {
1729 int gtid;
1730 if (!__kmp_init_gtid) {
1731 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1732 "KMP_GTID_SHUTDOWN\n"));
1733 return KMP_GTID_SHUTDOWN;
1734 }
1735 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1736 if (gtid == 0) {
1737 gtid = KMP_GTID_DNE;
1738 } else {
1739 gtid--;
1740 }
1741 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1742 __kmp_gtid_threadprivate_key, gtid));
1743 return gtid;
1744}
1745
1746double __kmp_read_cpu_time(void) {
1747 /*clock_t t;*/
1748 struct tms buffer;
1749
1750 /*t =*/times(&buffer);
1751
1752 return (double)(buffer.tms_utime + buffer.tms_cutime) /
1753 (double)CLOCKS_PER_SEC;
1754}
1755
1756int __kmp_read_system_info(struct kmp_sys_info *info) {
1757 int status;
1758 struct rusage r_usage;
1759
1760 memset(info, 0, sizeof(*info));
1761
1762 status = getrusage(RUSAGE_SELF, &r_usage);
1763 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1764
1765 // The maximum resident set size utilized (in kilobytes)
1766 info->maxrss = r_usage.ru_maxrss;
1767 // The number of page faults serviced without any I/O
1768 info->minflt = r_usage.ru_minflt;
1769 // The number of page faults serviced that required I/O
1770 info->majflt = r_usage.ru_majflt;
1771 // The number of times a process was "swapped" out of memory
1772 info->nswap = r_usage.ru_nswap;
1773 // The number of times the file system had to perform input
1774 info->inblock = r_usage.ru_inblock;
1775 // The number of times the file system had to perform output
1776 info->oublock = r_usage.ru_oublock;
1777 // The number of times a context switch was voluntarily
1778 info->nvcsw = r_usage.ru_nvcsw;
1779 // The number of times a context switch was forced
1780 info->nivcsw = r_usage.ru_nivcsw;
1781
1782 return (status != 0);
1783}
1784
1785void __kmp_read_system_time(double *delta) {
1786 double t_ns;
1787 struct timeval tval;
1788 struct timespec stop;
1789 int status;
1790
1791 status = gettimeofday(&tval, NULL);
1792 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1793 TIMEVAL_TO_TIMESPEC(&tval, &stop);
1794 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start));
1795 *delta = (t_ns * 1e-9);
1796}
1797
1798void __kmp_clear_system_time(void) {
1799 struct timeval tval;
1800 int status;
1801 status = gettimeofday(&tval, NULL);
1802 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1803 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1804}
1805
1806static int __kmp_get_xproc(void) {
1807
1808 int r = 0;
1809
1810#if KMP_OS_LINUX
1811
1812 __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r));
1813
1814#elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \
1815 KMP_OS_HURD
1816
1817 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r));
1818
1819#elif KMP_OS_DARWIN
1820
1821 // Bug C77011 High "OpenMP Threads and number of active cores".
1822
1823 // Find the number of available CPUs.
1824 kern_return_t rc;
1825 host_basic_info_data_t info;
1826 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1827 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1828 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1829 // Cannot use KA_TRACE() here because this code works before trace support
1830 // is initialized.
1831 r = info.avail_cpus;
1832 } else {
1833 KMP_WARNING(CantGetNumAvailCPU);
1834 KMP_INFORM(AssumedNumCPU);
1835 }
1836
1837#else
1838
1839#error "Unknown or unsupported OS."
1840
1841#endif
1842
1843 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1844
1845} // __kmp_get_xproc
1846
1847int __kmp_read_from_file(char const *path, char const *format, ...) {
1848 int result;
1849 va_list args;
1850
1851 va_start(args, format);
1852 FILE *f = fopen(path, "rb");
1853 if (f == NULL) {
1854 va_end(args);
1855 return 0;
1856 }
1857 result = vfscanf(f, format, args);
1858 fclose(f);
1859 va_end(args);
1860
1861 return result;
1862}
1863
1864void __kmp_runtime_initialize(void) {
1865 int status;
1866 pthread_mutexattr_t mutex_attr;
1867 pthread_condattr_t cond_attr;
1868
1869 if (__kmp_init_runtime) {
1870 return;
1871 }
1872
1873#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1874 if (!__kmp_cpuinfo.initialized) {
1875 __kmp_query_cpuid(&__kmp_cpuinfo);
1876 }
1877#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1878
1879 __kmp_xproc = __kmp_get_xproc();
1880
1881#if !KMP_32_BIT_ARCH
1882 struct rlimit rlim;
1883 // read stack size of calling thread, save it as default for worker threads;
1884 // this should be done before reading environment variables
1885 status = getrlimit(RLIMIT_STACK, &rlim);
1886 if (status == 0) { // success?
1887 __kmp_stksize = rlim.rlim_cur;
1888 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed
1889 }
1890#endif /* KMP_32_BIT_ARCH */
1891
1892 if (sysconf(_SC_THREADS)) {
1893
1894 /* Query the maximum number of threads */
1895 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth));
1896 if (__kmp_sys_max_nth == -1) {
1897 /* Unlimited threads for NPTL */
1898 __kmp_sys_max_nth = INT_MAX;
1899 } else if (__kmp_sys_max_nth <= 1) {
1900 /* Can't tell, just use PTHREAD_THREADS_MAX */
1901 __kmp_sys_max_nth = KMP_MAX_NTH;
1902 }
1903
1904 /* Query the minimum stack size */
1905 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1906 if (__kmp_sys_min_stksize <= 1) {
1907 __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1908 }
1909 }
1910
1911 /* Set up minimum number of threads to switch to TLS gtid */
1912 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1913
1914 status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1915 __kmp_internal_end_dest);
1916 KMP_CHECK_SYSFAIL("pthread_key_create", status);
1917 status = pthread_mutexattr_init(&mutex_attr);
1918 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1919 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1920 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1921 status = pthread_mutexattr_destroy(&mutex_attr);
1922 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status);
1923 status = pthread_condattr_init(&cond_attr);
1924 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1925 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1926 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1927 status = pthread_condattr_destroy(&cond_attr);
1928 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status);
1929#if USE_ITT_BUILD
1930 __kmp_itt_initialize();
1931#endif /* USE_ITT_BUILD */
1932
1933 __kmp_init_runtime = TRUE;
1934}
1935
1936void __kmp_runtime_destroy(void) {
1937 int status;
1938
1939 if (!__kmp_init_runtime) {
1940 return; // Nothing to do.
1941 }
1942
1943#if USE_ITT_BUILD
1944 __kmp_itt_destroy();
1945#endif /* USE_ITT_BUILD */
1946
1947 status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1948 KMP_CHECK_SYSFAIL("pthread_key_delete", status);
1949
1950 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1951 if (status != 0 && status != EBUSY) {
1952 KMP_SYSFAIL("pthread_mutex_destroy", status);
1953 }
1954 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1955 if (status != 0 && status != EBUSY) {
1956 KMP_SYSFAIL("pthread_cond_destroy", status);
1957 }
1958#if KMP_AFFINITY_SUPPORTED
1959 __kmp_affinity_uninitialize();
1960#endif
1961
1962 __kmp_init_runtime = FALSE;
1963}
1964
1965/* Put the thread to sleep for a time period */
1966/* NOTE: not currently used anywhere */
1967void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
1968
1969/* Calculate the elapsed wall clock time for the user */
1970void __kmp_elapsed(double *t) {
1971 int status;
1972#ifdef FIX_SGI_CLOCK
1973 struct timespec ts;
1974
1975 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1976 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
1977 *t =
1978 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
1979#else
1980 struct timeval tv;
1981
1982 status = gettimeofday(&tv, NULL);
1983 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1984 *t =
1985 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
1986#endif
1987}
1988
1989/* Calculate the elapsed wall clock tick for the user */
1990void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1991
1992/* Return the current time stamp in nsec */
1993kmp_uint64 __kmp_now_nsec() {
1994 struct timeval t;
1995 gettimeofday(&t, NULL);
1996 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
1997 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
1998 return nsec;
1999}
2000
2001#if KMP_ARCH_X86 || KMP_ARCH_X86_64
2002/* Measure clock ticks per millisecond */
2003void __kmp_initialize_system_tick() {
2004 kmp_uint64 now, nsec2, diff;
2005 kmp_uint64 delay = 100000; // 50~100 usec on most machines.
2006 kmp_uint64 nsec = __kmp_now_nsec();
2007 kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
2008 while ((now = __kmp_hardware_timestamp()) < goal)
2009 ;
2010 nsec2 = __kmp_now_nsec();
2011 diff = nsec2 - nsec;
2012 if (diff > 0) {
2013 kmp_uint64 tpms = ((kmp_uint64)1e6 * (delay + (now - goal)) / diff);
2014 if (tpms > 0)
2015 __kmp_ticks_per_msec = tpms;
2016 }
2017}
2018#endif
2019
2020/* Determine whether the given address is mapped into the current address
2021 space. */
2022
2023int __kmp_is_address_mapped(void *addr) {
2024
2025 int found = 0;
2026 int rc;
2027
2028#if KMP_OS_LINUX || KMP_OS_HURD
2029
2030 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the
2031 address ranges mapped into the address space. */
2032
2033 char *name = __kmp_str_format("/proc/%d/maps", getpid());
2034 FILE *file = NULL;
2035
2036 file = fopen(name, "r");
2037 KMP_ASSERT(file != NULL);
2038
2039 for (;;) {
2040
2041 void *beginning = NULL;
2042 void *ending = NULL;
2043 char perms[5];
2044
2045 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
2046 if (rc == EOF) {
2047 break;
2048 }
2049 KMP_ASSERT(rc == 3 &&
2050 KMP_STRLEN(perms) == 4); // Make sure all fields are read.
2051
2052 // Ending address is not included in the region, but beginning is.
2053 if ((addr >= beginning) && (addr < ending)) {
2054 perms[2] = 0; // 3th and 4th character does not matter.
2055 if (strcmp(perms, "rw") == 0) {
2056 // Memory we are looking for should be readable and writable.
2057 found = 1;
2058 }
2059 break;
2060 }
2061 }
2062
2063 // Free resources.
2064 fclose(file);
2065 KMP_INTERNAL_FREE(name);
2066#elif KMP_OS_FREEBSD
2067 char *buf;
2068 size_t lstsz;
2069 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()};
2070 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0);
2071 if (rc < 0)
2072 return 0;
2073 // We pass from number of vm entry's semantic
2074 // to size of whole entry map list.
2075 lstsz = lstsz * 4 / 3;
2076 buf = reinterpret_cast<char *>(kmpc_malloc(lstsz));
2077 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0);
2078 if (rc < 0) {
2079 kmpc_free(buf);
2080 return 0;
2081 }
2082
2083 char *lw = buf;
2084 char *up = buf + lstsz;
2085
2086 while (lw < up) {
2087 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw);
2088 size_t cursz = cur->kve_structsize;
2089 if (cursz == 0)
2090 break;
2091 void *start = reinterpret_cast<void *>(cur->kve_start);
2092 void *end = reinterpret_cast<void *>(cur->kve_end);
2093 // Readable/Writable addresses within current map entry
2094 if ((addr >= start) && (addr < end)) {
2095 if ((cur->kve_protection & KVME_PROT_READ) != 0 &&
2096 (cur->kve_protection & KVME_PROT_WRITE) != 0) {
2097 found = 1;
2098 break;
2099 }
2100 }
2101 lw += cursz;
2102 }
2103 kmpc_free(buf);
2104
2105#elif KMP_OS_DARWIN
2106
2107 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2108 using vm interface. */
2109
2110 int buffer;
2111 vm_size_t count;
2112 rc = vm_read_overwrite(
2113 mach_task_self(), // Task to read memory of.
2114 (vm_address_t)(addr), // Address to read from.
2115 1, // Number of bytes to be read.
2116 (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2117 &count // Address of var to save number of read bytes in.
2118 );
2119 if (rc == 0) {
2120 // Memory successfully read.
2121 found = 1;
2122 }
2123
2124#elif KMP_OS_NETBSD
2125
2126 int mib[5];
2127 mib[0] = CTL_VM;
2128 mib[1] = VM_PROC;
2129 mib[2] = VM_PROC_MAP;
2130 mib[3] = getpid();
2131 mib[4] = sizeof(struct kinfo_vmentry);
2132
2133 size_t size;
2134 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2135 KMP_ASSERT(!rc);
2136 KMP_ASSERT(size);
2137
2138 size = size * 4 / 3;
2139 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2140 KMP_ASSERT(kiv);
2141
2142 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2143 KMP_ASSERT(!rc);
2144 KMP_ASSERT(size);
2145
2146 for (size_t i = 0; i < size; i++) {
2147 if (kiv[i].kve_start >= (uint64_t)addr &&
2148 kiv[i].kve_end <= (uint64_t)addr) {
2149 found = 1;
2150 break;
2151 }
2152 }
2153 KMP_INTERNAL_FREE(kiv);
2154#elif KMP_OS_OPENBSD
2155
2156 int mib[3];
2157 mib[0] = CTL_KERN;
2158 mib[1] = KERN_PROC_VMMAP;
2159 mib[2] = getpid();
2160
2161 size_t size;
2162 uint64_t end;
2163 rc = sysctl(mib, 3, NULL, &size, NULL, 0);
2164 KMP_ASSERT(!rc);
2165 KMP_ASSERT(size);
2166 end = size;
2167
2168 struct kinfo_vmentry kiv = {.kve_start = 0};
2169
2170 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) {
2171 KMP_ASSERT(size);
2172 if (kiv.kve_end == end)
2173 break;
2174
2175 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) {
2176 found = 1;
2177 break;
2178 }
2179 kiv.kve_start += 1;
2180 }
2181#elif KMP_OS_DRAGONFLY
2182
2183 // FIXME(DragonFly): Implement this
2184 found = 1;
2185
2186#else
2187
2188#error "Unknown or unsupported OS"
2189
2190#endif
2191
2192 return found;
2193
2194} // __kmp_is_address_mapped
2195
2196#ifdef USE_LOAD_BALANCE
2197
2198#if KMP_OS_DARWIN || KMP_OS_NETBSD
2199
2200// The function returns the rounded value of the system load average
2201// during given time interval which depends on the value of
2202// __kmp_load_balance_interval variable (default is 60 sec, other values
2203// may be 300 sec or 900 sec).
2204// It returns -1 in case of error.
2205int __kmp_get_load_balance(int max) {
2206 double averages[3];
2207 int ret_avg = 0;
2208
2209 int res = getloadavg(averages, 3);
2210
2211 // Check __kmp_load_balance_interval to determine which of averages to use.
2212 // getloadavg() may return the number of samples less than requested that is
2213 // less than 3.
2214 if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2215 ret_avg = (int)averages[0]; // 1 min
2216 } else if ((__kmp_load_balance_interval >= 180 &&
2217 __kmp_load_balance_interval < 600) &&
2218 (res >= 2)) {
2219 ret_avg = (int)averages[1]; // 5 min
2220 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2221 ret_avg = (int)averages[2]; // 15 min
2222 } else { // Error occurred
2223 return -1;
2224 }
2225
2226 return ret_avg;
2227}
2228
2229#else // Linux* OS
2230
2231// The function returns number of running (not sleeping) threads, or -1 in case
2232// of error. Error could be reported if Linux* OS kernel too old (without
2233// "/proc" support). Counting running threads stops if max running threads
2234// encountered.
2235int __kmp_get_load_balance(int max) {
2236 static int permanent_error = 0;
2237 static int glb_running_threads = 0; // Saved count of the running threads for
2238 // the thread balance algorithm
2239 static double glb_call_time = 0; /* Thread balance algorithm call time */
2240
2241 int running_threads = 0; // Number of running threads in the system.
2242
2243 DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2244 struct dirent *proc_entry = NULL;
2245
2246 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2247 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2248 struct dirent *task_entry = NULL;
2249 int task_path_fixed_len;
2250
2251 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2252 int stat_file = -1;
2253 int stat_path_fixed_len;
2254
2255#ifdef KMP_DEBUG
2256 int total_processes = 0; // Total number of processes in system.
2257#endif
2258
2259 double call_time = 0.0;
2260
2261 __kmp_str_buf_init(&task_path);
2262 __kmp_str_buf_init(&stat_path);
2263
2264 __kmp_elapsed(&call_time);
2265
2266 if (glb_call_time &&
2267 (call_time - glb_call_time < __kmp_load_balance_interval)) {
2268 running_threads = glb_running_threads;
2269 goto finish;
2270 }
2271
2272 glb_call_time = call_time;
2273
2274 // Do not spend time on scanning "/proc/" if we have a permanent error.
2275 if (permanent_error) {
2276 running_threads = -1;
2277 goto finish;
2278 }
2279
2280 if (max <= 0) {
2281 max = INT_MAX;
2282 }
2283
2284 // Open "/proc/" directory.
2285 proc_dir = opendir("/proc");
2286 if (proc_dir == NULL) {
2287 // Cannot open "/prroc/". Probably the kernel does not support it. Return an
2288 // error now and in subsequent calls.
2289 running_threads = -1;
2290 permanent_error = 1;
2291 goto finish;
2292 }
2293
2294 // Initialize fixed part of task_path. This part will not change.
2295 __kmp_str_buf_cat(&task_path, "/proc/", 6);
2296 task_path_fixed_len = task_path.used; // Remember number of used characters.
2297
2298 proc_entry = readdir(proc_dir);
2299 while (proc_entry != NULL) {
2300 // Proc entry is a directory and name starts with a digit. Assume it is a
2301 // process' directory.
2302 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2303
2304#ifdef KMP_DEBUG
2305 ++total_processes;
2306#endif
2307 // Make sure init process is the very first in "/proc", so we can replace
2308 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2309 // 1. We are going to check that total_processes == 1 => d_name == "1" is
2310 // true (where "=>" is implication). Since C++ does not have => operator,
2311 // let us replace it with its equivalent: a => b == ! a || b.
2312 KMP_DEBUG_ASSERT(total_processes != 1 ||
2313 strcmp(proc_entry->d_name, "1") == 0);
2314
2315 // Construct task_path.
2316 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2317 __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2318 KMP_STRLEN(proc_entry->d_name));
2319 __kmp_str_buf_cat(&task_path, "/task", 5);
2320
2321 task_dir = opendir(task_path.str);
2322 if (task_dir == NULL) {
2323 // Process can finish between reading "/proc/" directory entry and
2324 // opening process' "task/" directory. So, in general case we should not
2325 // complain, but have to skip this process and read the next one. But on
2326 // systems with no "task/" support we will spend lot of time to scan
2327 // "/proc/" tree again and again without any benefit. "init" process
2328 // (its pid is 1) should exist always, so, if we cannot open
2329 // "/proc/1/task/" directory, it means "task/" is not supported by
2330 // kernel. Report an error now and in the future.
2331 if (strcmp(proc_entry->d_name, "1") == 0) {
2332 running_threads = -1;
2333 permanent_error = 1;
2334 goto finish;
2335 }
2336 } else {
2337 // Construct fixed part of stat file path.
2338 __kmp_str_buf_clear(&stat_path);
2339 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2340 __kmp_str_buf_cat(&stat_path, "/", 1);
2341 stat_path_fixed_len = stat_path.used;
2342
2343 task_entry = readdir(task_dir);
2344 while (task_entry != NULL) {
2345 // It is a directory and name starts with a digit.
2346 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2347
2348 // Construct complete stat file path. Easiest way would be:
2349 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2350 // task_entry->d_name );
2351 // but seriae of __kmp_str_buf_cat works a bit faster.
2352 stat_path.used =
2353 stat_path_fixed_len; // Reset stat path to its fixed part.
2354 __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2355 KMP_STRLEN(task_entry->d_name));
2356 __kmp_str_buf_cat(&stat_path, "/stat", 5);
2357
2358 // Note: Low-level API (open/read/close) is used. High-level API
2359 // (fopen/fclose) works ~ 30 % slower.
2360 stat_file = open(stat_path.str, O_RDONLY);
2361 if (stat_file == -1) {
2362 // We cannot report an error because task (thread) can terminate
2363 // just before reading this file.
2364 } else {
2365 /* Content of "stat" file looks like:
2366 24285 (program) S ...
2367
2368 It is a single line (if program name does not include funny
2369 symbols). First number is a thread id, then name of executable
2370 file name in paretheses, then state of the thread. We need just
2371 thread state.
2372
2373 Good news: Length of program name is 15 characters max. Longer
2374 names are truncated.
2375
2376 Thus, we need rather short buffer: 15 chars for program name +
2377 2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2378
2379 Bad news: Program name may contain special symbols like space,
2380 closing parenthesis, or even new line. This makes parsing
2381 "stat" file not 100 % reliable. In case of fanny program names
2382 parsing may fail (report incorrect thread state).
2383
2384 Parsing "status" file looks more promissing (due to different
2385 file structure and escaping special symbols) but reading and
2386 parsing of "status" file works slower.
2387 -- ln
2388 */
2389 char buffer[65];
2390 ssize_t len;
2391 len = read(stat_file, buffer, sizeof(buffer) - 1);
2392 if (len >= 0) {
2393 buffer[len] = 0;
2394 // Using scanf:
2395 // sscanf( buffer, "%*d (%*s) %c ", & state );
2396 // looks very nice, but searching for a closing parenthesis
2397 // works a bit faster.
2398 char *close_parent = strstr(buffer, ") ");
2399 if (close_parent != NULL) {
2400 char state = *(close_parent + 2);
2401 if (state == 'R') {
2402 ++running_threads;
2403 if (running_threads >= max) {
2404 goto finish;
2405 }
2406 }
2407 }
2408 }
2409 close(stat_file);
2410 stat_file = -1;
2411 }
2412 }
2413 task_entry = readdir(task_dir);
2414 }
2415 closedir(task_dir);
2416 task_dir = NULL;
2417 }
2418 }
2419 proc_entry = readdir(proc_dir);
2420 }
2421
2422 // There _might_ be a timing hole where the thread executing this
2423 // code get skipped in the load balance, and running_threads is 0.
2424 // Assert in the debug builds only!!!
2425 KMP_DEBUG_ASSERT(running_threads > 0);
2426 if (running_threads <= 0) {
2427 running_threads = 1;
2428 }
2429
2430finish: // Clean up and exit.
2431 if (proc_dir != NULL) {
2432 closedir(proc_dir);
2433 }
2434 __kmp_str_buf_free(&task_path);
2435 if (task_dir != NULL) {
2436 closedir(task_dir);
2437 }
2438 __kmp_str_buf_free(&stat_path);
2439 if (stat_file != -1) {
2440 close(stat_file);
2441 }
2442
2443 glb_running_threads = running_threads;
2444
2445 return running_threads;
2446
2447} // __kmp_get_load_balance
2448
2449#endif // KMP_OS_DARWIN
2450
2451#endif // USE_LOAD_BALANCE
2452
2453#if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \
2454 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \
2455 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \
2456 KMP_ARCH_ARM)
2457
2458// we really only need the case with 1 argument, because CLANG always build
2459// a struct of pointers to shared variables referenced in the outlined function
2460int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2461 void *p_argv[]
2462#if OMPT_SUPPORT
2463 ,
2464 void **exit_frame_ptr
2465#endif
2466) {
2467#if OMPT_SUPPORT
2468 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2469#endif
2470
2471 switch (argc) {
2472 default:
2473 fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2474 fflush(stderr);
2475 exit(-1);
2476 case 0:
2477 (*pkfn)(&gtid, &tid);
2478 break;
2479 case 1:
2480 (*pkfn)(&gtid, &tid, p_argv[0]);
2481 break;
2482 case 2:
2483 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1]);
2484 break;
2485 case 3:
2486 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2]);
2487 break;
2488 case 4:
2489 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2490 break;
2491 case 5:
2492 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2493 break;
2494 case 6:
2495 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2496 p_argv[5]);
2497 break;
2498 case 7:
2499 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2500 p_argv[5], p_argv[6]);
2501 break;
2502 case 8:
2503 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2504 p_argv[5], p_argv[6], p_argv[7]);
2505 break;
2506 case 9:
2507 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2508 p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2509 break;
2510 case 10:
2511 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2512 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2513 break;
2514 case 11:
2515 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2516 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2517 break;
2518 case 12:
2519 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2520 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2521 p_argv[11]);
2522 break;
2523 case 13:
2524 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2525 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2526 p_argv[11], p_argv[12]);
2527 break;
2528 case 14:
2529 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2530 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2531 p_argv[11], p_argv[12], p_argv[13]);
2532 break;
2533 case 15:
2534 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2535 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2536 p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2537 break;
2538 }
2539
2540 return 1;
2541}
2542
2543#endif
2544
2545#if KMP_OS_LINUX
2546// Functions for hidden helper task
2547namespace {
2548// Condition variable for initializing hidden helper team
2549pthread_cond_t hidden_helper_threads_initz_cond_var;
2550pthread_mutex_t hidden_helper_threads_initz_lock;
2551volatile int hidden_helper_initz_signaled = FALSE;
2552
2553// Condition variable for deinitializing hidden helper team
2554pthread_cond_t hidden_helper_threads_deinitz_cond_var;
2555pthread_mutex_t hidden_helper_threads_deinitz_lock;
2556volatile int hidden_helper_deinitz_signaled = FALSE;
2557
2558// Condition variable for the wrapper function of main thread
2559pthread_cond_t hidden_helper_main_thread_cond_var;
2560pthread_mutex_t hidden_helper_main_thread_lock;
2561volatile int hidden_helper_main_thread_signaled = FALSE;
2562
2563// Semaphore for worker threads. We don't use condition variable here in case
2564// that when multiple signals are sent at the same time, only one thread might
2565// be waken.
2566sem_t hidden_helper_task_sem;
2567} // namespace
2568
2569void __kmp_hidden_helper_worker_thread_wait() {
2570 int status = sem_wait(&hidden_helper_task_sem);
2571 KMP_CHECK_SYSFAIL("sem_wait", status);
2572}
2573
2574void __kmp_do_initialize_hidden_helper_threads() {
2575 // Initialize condition variable
2576 int status =
2577 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr);
2578 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2579
2580 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr);
2581 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2582
2583 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr);
2584 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2585
2586 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr);
2587 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2588
2589 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr);
2590 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2591
2592 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr);
2593 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2594
2595 // Initialize the semaphore
2596 status = sem_init(&hidden_helper_task_sem, 0, 0);
2597 KMP_CHECK_SYSFAIL("sem_init", status);
2598
2599 // Create a new thread to finish initialization
2600 pthread_t handle;
2601 status = pthread_create(
2602 &handle, nullptr,
2603 [](void *) -> void * {
2604 __kmp_hidden_helper_threads_initz_routine();
2605 return nullptr;
2606 },
2607 nullptr);
2608 KMP_CHECK_SYSFAIL("pthread_create", status);
2609}
2610
2611void __kmp_hidden_helper_threads_initz_wait() {
2612 // Initial thread waits here for the completion of the initialization. The
2613 // condition variable will be notified by main thread of hidden helper teams.
2614 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2615 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2616
2617 if (!TCR_4(hidden_helper_initz_signaled)) {
2618 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var,
2619 &hidden_helper_threads_initz_lock);
2620 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2621 }
2622
2623 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2624 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2625}
2626
2627void __kmp_hidden_helper_initz_release() {
2628 // After all initialization, reset __kmp_init_hidden_helper_threads to false.
2629 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2630 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2631
2632 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var);
2633 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2634
2635 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE);
2636
2637 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2638 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2639}
2640
2641void __kmp_hidden_helper_main_thread_wait() {
2642 // The main thread of hidden helper team will be blocked here. The
2643 // condition variable can only be signal in the destructor of RTL.
2644 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2645 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2646
2647 if (!TCR_4(hidden_helper_main_thread_signaled)) {
2648 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var,
2649 &hidden_helper_main_thread_lock);
2650 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2651 }
2652
2653 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2654 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2655}
2656
2657void __kmp_hidden_helper_main_thread_release() {
2658 // The initial thread of OpenMP RTL should call this function to wake up the
2659 // main thread of hidden helper team.
2660 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2661 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2662
2663 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var);
2664 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
2665
2666 // The hidden helper team is done here
2667 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE);
2668
2669 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2670 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2671}
2672
2673void __kmp_hidden_helper_worker_thread_signal() {
2674 int status = sem_post(&hidden_helper_task_sem);
2675 KMP_CHECK_SYSFAIL("sem_post", status);
2676}
2677
2678void __kmp_hidden_helper_threads_deinitz_wait() {
2679 // Initial thread waits here for the completion of the deinitialization. The
2680 // condition variable will be notified by main thread of hidden helper teams.
2681 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
2682 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2683
2684 if (!TCR_4(hidden_helper_deinitz_signaled)) {
2685 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var,
2686 &hidden_helper_threads_deinitz_lock);
2687 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2688 }
2689
2690 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
2691 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2692}
2693
2694void __kmp_hidden_helper_threads_deinitz_release() {
2695 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
2696 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2697
2698 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var);
2699 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2700
2701 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE);
2702
2703 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
2704 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2705}
2706#else // KMP_OS_LINUX
2707void __kmp_hidden_helper_worker_thread_wait() {
2708 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2709}
2710
2711void __kmp_do_initialize_hidden_helper_threads() {
2712 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2713}
2714
2715void __kmp_hidden_helper_threads_initz_wait() {
2716 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2717}
2718
2719void __kmp_hidden_helper_initz_release() {
2720 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2721}
2722
2723void __kmp_hidden_helper_main_thread_wait() {
2724 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2725}
2726
2727void __kmp_hidden_helper_main_thread_release() {
2728 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2729}
2730
2731void __kmp_hidden_helper_worker_thread_signal() {
2732 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2733}
2734
2735void __kmp_hidden_helper_threads_deinitz_wait() {
2736 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2737}
2738
2739void __kmp_hidden_helper_threads_deinitz_release() {
2740 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2741}
2742#endif // KMP_OS_LINUX
2743
2744// end of file //
#define KMP_INIT_PARTITIONED_TIMERS(name)
Initializes the partitioned timers to begin with name.
Definition: kmp_stats.h:940