LLVM OpenMP* Runtime Library
Loading...
Searching...
No Matches
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;
96kmp_uint64 __kmp_ticks_per_usec = 1000;
97
98#ifdef DEBUG_SUSPEND
99static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
100 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
101 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
102 cond->c_cond.__c_waiting);
103}
104#endif
105
106#if ((KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED)
107
108/* Affinity support */
109
110void __kmp_affinity_bind_thread(int which) {
111 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
112 "Illegal set affinity operation when not capable");
113
114 kmp_affin_mask_t *mask;
115 KMP_CPU_ALLOC_ON_STACK(mask);
116 KMP_CPU_ZERO(mask);
117 KMP_CPU_SET(which, mask);
118 __kmp_set_system_affinity(mask, TRUE);
119 KMP_CPU_FREE_FROM_STACK(mask);
120}
121
122/* Determine if we can access affinity functionality on this version of
123 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
124 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
125void __kmp_affinity_determine_capable(const char *env_var) {
126 // Check and see if the OS supports thread affinity.
127
128#if KMP_OS_LINUX
129#define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
130#define KMP_CPU_SET_TRY_SIZE CACHE_LINE
131#elif KMP_OS_FREEBSD
132#define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t))
133#endif
134
135 int verbose = __kmp_affinity.flags.verbose;
136 int warnings = __kmp_affinity.flags.warnings;
137 enum affinity_type type = __kmp_affinity.type;
138
139#if KMP_OS_LINUX
140 long gCode;
141 unsigned char *buf;
142 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
143
144 // If the syscall returns a suggestion for the size,
145 // then we don't have to search for an appropriate size.
146 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf);
147 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
148 "initial getaffinity call returned %ld errno = %d\n",
149 gCode, errno));
150
151 if (gCode < 0 && errno != EINVAL) {
152 // System call not supported
153 if (verbose ||
154 (warnings && (type != affinity_none) && (type != affinity_default) &&
155 (type != affinity_disabled))) {
156 int error = errno;
157 kmp_msg_t err_code = KMP_ERR(error);
158 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
159 err_code, __kmp_msg_null);
160 if (__kmp_generate_warnings == kmp_warnings_off) {
161 __kmp_str_free(&err_code.str);
162 }
163 }
164 KMP_AFFINITY_DISABLE();
165 KMP_INTERNAL_FREE(buf);
166 return;
167 } else if (gCode > 0) {
168 // The optimal situation: the OS returns the size of the buffer it expects.
169 KMP_AFFINITY_ENABLE(gCode);
170 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
171 "affinity supported (mask size %d)\n",
172 (int)__kmp_affin_mask_size));
173 KMP_INTERNAL_FREE(buf);
174 return;
175 }
176
177 // Call the getaffinity system call repeatedly with increasing set sizes
178 // until we succeed, or reach an upper bound on the search.
179 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
180 "searching for proper set size\n"));
181 int size;
182 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
183 gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
184 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
185 "getaffinity for mask size %ld returned %ld errno = %d\n",
186 size, gCode, errno));
187
188 if (gCode < 0) {
189 if (errno == ENOSYS) {
190 // We shouldn't get here
191 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
192 "inconsistent OS call behavior: errno == ENOSYS for mask "
193 "size %d\n",
194 size));
195 if (verbose ||
196 (warnings && (type != affinity_none) &&
197 (type != affinity_default) && (type != affinity_disabled))) {
198 int error = errno;
199 kmp_msg_t err_code = KMP_ERR(error);
200 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
201 err_code, __kmp_msg_null);
202 if (__kmp_generate_warnings == kmp_warnings_off) {
203 __kmp_str_free(&err_code.str);
204 }
205 }
206 KMP_AFFINITY_DISABLE();
207 KMP_INTERNAL_FREE(buf);
208 return;
209 }
210 continue;
211 }
212
213 KMP_AFFINITY_ENABLE(gCode);
214 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
215 "affinity supported (mask size %d)\n",
216 (int)__kmp_affin_mask_size));
217 KMP_INTERNAL_FREE(buf);
218 return;
219 }
220#elif KMP_OS_FREEBSD
221 long gCode;
222 unsigned char *buf;
223 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
224 gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT,
225 reinterpret_cast<cpuset_t *>(buf));
226 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
227 "initial getaffinity call returned %d errno = %d\n",
228 gCode, errno));
229 if (gCode == 0) {
230 KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT);
231 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
232 "affinity supported (mask size %d)\n",
233 (int)__kmp_affin_mask_size));
234 KMP_INTERNAL_FREE(buf);
235 return;
236 }
237#endif
238 KMP_INTERNAL_FREE(buf);
239
240 // Affinity is not supported
241 KMP_AFFINITY_DISABLE();
242 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
243 "cannot determine mask size - affinity not supported\n"));
244 if (verbose || (warnings && (type != affinity_none) &&
245 (type != affinity_default) && (type != affinity_disabled))) {
246 KMP_WARNING(AffCantGetMaskSize, env_var);
247 }
248}
249
250#endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
251
252#if KMP_USE_FUTEX
253
254int __kmp_futex_determine_capable() {
255 int loc = 0;
256 long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
257 int retval = (rc == 0) || (errno != ENOSYS);
258
259 KA_TRACE(10,
260 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
261 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
262 retval ? "" : " not"));
263
264 return retval;
265}
266
267#endif // KMP_USE_FUTEX
268
269#if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS)
270/* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
271 use compare_and_store for these routines */
272
273kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
274 kmp_int8 old_value, new_value;
275
276 old_value = TCR_1(*p);
277 new_value = old_value | d;
278
279 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
280 KMP_CPU_PAUSE();
281 old_value = TCR_1(*p);
282 new_value = old_value | d;
283 }
284 return old_value;
285}
286
287kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
288 kmp_int8 old_value, new_value;
289
290 old_value = TCR_1(*p);
291 new_value = old_value & d;
292
293 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
294 KMP_CPU_PAUSE();
295 old_value = TCR_1(*p);
296 new_value = old_value & d;
297 }
298 return old_value;
299}
300
301kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
302 kmp_uint32 old_value, new_value;
303
304 old_value = TCR_4(*p);
305 new_value = old_value | d;
306
307 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
308 KMP_CPU_PAUSE();
309 old_value = TCR_4(*p);
310 new_value = old_value | d;
311 }
312 return old_value;
313}
314
315kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
316 kmp_uint32 old_value, new_value;
317
318 old_value = TCR_4(*p);
319 new_value = old_value & d;
320
321 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
322 KMP_CPU_PAUSE();
323 old_value = TCR_4(*p);
324 new_value = old_value & d;
325 }
326 return old_value;
327}
328
329#if KMP_ARCH_X86
330kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
331 kmp_int8 old_value, new_value;
332
333 old_value = TCR_1(*p);
334 new_value = old_value + d;
335
336 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
337 KMP_CPU_PAUSE();
338 old_value = TCR_1(*p);
339 new_value = old_value + d;
340 }
341 return old_value;
342}
343
344kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
345 kmp_int64 old_value, new_value;
346
347 old_value = TCR_8(*p);
348 new_value = old_value + d;
349
350 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
351 KMP_CPU_PAUSE();
352 old_value = TCR_8(*p);
353 new_value = old_value + d;
354 }
355 return old_value;
356}
357#endif /* KMP_ARCH_X86 */
358
359kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
360 kmp_uint64 old_value, new_value;
361
362 old_value = TCR_8(*p);
363 new_value = old_value | d;
364 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
365 KMP_CPU_PAUSE();
366 old_value = TCR_8(*p);
367 new_value = old_value | d;
368 }
369 return old_value;
370}
371
372kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
373 kmp_uint64 old_value, new_value;
374
375 old_value = TCR_8(*p);
376 new_value = old_value & d;
377 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
378 KMP_CPU_PAUSE();
379 old_value = TCR_8(*p);
380 new_value = old_value & d;
381 }
382 return old_value;
383}
384
385#endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
386
387void __kmp_terminate_thread(int gtid) {
388 int status;
389 kmp_info_t *th = __kmp_threads[gtid];
390
391 if (!th)
392 return;
393
394#ifdef KMP_CANCEL_THREADS
395 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
396 status = pthread_cancel(th->th.th_info.ds.ds_thread);
397 if (status != 0 && status != ESRCH) {
398 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
399 __kmp_msg_null);
400 }
401#endif
402 KMP_YIELD(TRUE);
403} //
404
405/* Set thread stack info according to values returned by pthread_getattr_np().
406 If values are unreasonable, assume call failed and use incremental stack
407 refinement method instead. Returns TRUE if the stack parameters could be
408 determined exactly, FALSE if incremental refinement is necessary. */
409static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
410 int stack_data;
411#if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
412 KMP_OS_HURD
413 pthread_attr_t attr;
414 int status;
415 size_t size = 0;
416 void *addr = 0;
417
418 /* Always do incremental stack refinement for ubermaster threads since the
419 initial thread stack range can be reduced by sibling thread creation so
420 pthread_attr_getstack may cause thread gtid aliasing */
421 if (!KMP_UBER_GTID(gtid)) {
422
423 /* Fetch the real thread attributes */
424 status = pthread_attr_init(&attr);
425 KMP_CHECK_SYSFAIL("pthread_attr_init", status);
426#if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
427 status = pthread_attr_get_np(pthread_self(), &attr);
428 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
429#else
430 status = pthread_getattr_np(pthread_self(), &attr);
431 KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
432#endif
433 status = pthread_attr_getstack(&attr, &addr, &size);
434 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
435 KA_TRACE(60,
436 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
437 " %lu, low addr: %p\n",
438 gtid, size, addr));
439 status = pthread_attr_destroy(&attr);
440 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
441 }
442
443 if (size != 0 && addr != 0) { // was stack parameter determination successful?
444 /* Store the correct base and size */
445 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
446 TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
447 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
448 return TRUE;
449 }
450#endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD \
451 || KMP_OS_HURD */
452 /* Use incremental refinement starting from initial conservative estimate */
453 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
454 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
455 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
456 return FALSE;
457}
458
459static void *__kmp_launch_worker(void *thr) {
460 int status, old_type, old_state;
461#ifdef KMP_BLOCK_SIGNALS
462 sigset_t new_set, old_set;
463#endif /* KMP_BLOCK_SIGNALS */
464 void *exit_val;
465#if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
466 KMP_OS_OPENBSD || KMP_OS_HURD
467 void *volatile padding = 0;
468#endif
469 int gtid;
470
471 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
472 __kmp_gtid_set_specific(gtid);
473#ifdef KMP_TDATA_GTID
474 __kmp_gtid = gtid;
475#endif
476#if KMP_STATS_ENABLED
477 // set thread local index to point to thread-specific stats
478 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
479 __kmp_stats_thread_ptr->startLife();
480 KMP_SET_THREAD_STATE(IDLE);
482#endif
483
484#if USE_ITT_BUILD
485 __kmp_itt_thread_name(gtid);
486#endif /* USE_ITT_BUILD */
487
488#if KMP_AFFINITY_SUPPORTED
489 __kmp_affinity_bind_init_mask(gtid);
490#endif
491
492#ifdef KMP_CANCEL_THREADS
493 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
494 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
495 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
496 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
497 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
498#endif
499
500#if KMP_ARCH_X86 || KMP_ARCH_X86_64
501 // Set FP control regs to be a copy of the parallel initialization thread's.
502 __kmp_clear_x87_fpu_status_word();
503 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
504 __kmp_load_mxcsr(&__kmp_init_mxcsr);
505#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
506
507#ifdef KMP_BLOCK_SIGNALS
508 status = sigfillset(&new_set);
509 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
510 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
511 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
512#endif /* KMP_BLOCK_SIGNALS */
513
514#if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
515 KMP_OS_OPENBSD
516 if (__kmp_stkoffset > 0 && gtid > 0) {
517 padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
518 (void)padding;
519 }
520#endif
521
522 KMP_MB();
523 __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
524
525 __kmp_check_stack_overlap((kmp_info_t *)thr);
526
527 exit_val = __kmp_launch_thread((kmp_info_t *)thr);
528
529#ifdef KMP_BLOCK_SIGNALS
530 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
531 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
532#endif /* KMP_BLOCK_SIGNALS */
533
534 return exit_val;
535}
536
537#if KMP_USE_MONITOR
538/* The monitor thread controls all of the threads in the complex */
539
540static void *__kmp_launch_monitor(void *thr) {
541 int status, old_type, old_state;
542#ifdef KMP_BLOCK_SIGNALS
543 sigset_t new_set;
544#endif /* KMP_BLOCK_SIGNALS */
545 struct timespec interval;
546
547 KMP_MB(); /* Flush all pending memory write invalidates. */
548
549 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
550
551 /* register us as the monitor thread */
552 __kmp_gtid_set_specific(KMP_GTID_MONITOR);
553#ifdef KMP_TDATA_GTID
554 __kmp_gtid = KMP_GTID_MONITOR;
555#endif
556
557 KMP_MB();
558
559#if USE_ITT_BUILD
560 // Instruct Intel(R) Threading Tools to ignore monitor thread.
561 __kmp_itt_thread_ignore();
562#endif /* USE_ITT_BUILD */
563
564 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
565 (kmp_info_t *)thr);
566
567 __kmp_check_stack_overlap((kmp_info_t *)thr);
568
569#ifdef KMP_CANCEL_THREADS
570 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
571 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
572 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
573 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
574 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
575#endif
576
577#if KMP_REAL_TIME_FIX
578 // This is a potential fix which allows application with real-time scheduling
579 // policy work. However, decision about the fix is not made yet, so it is
580 // disabled by default.
581 { // Are program started with real-time scheduling policy?
582 int sched = sched_getscheduler(0);
583 if (sched == SCHED_FIFO || sched == SCHED_RR) {
584 // Yes, we are a part of real-time application. Try to increase the
585 // priority of the monitor.
586 struct sched_param param;
587 int max_priority = sched_get_priority_max(sched);
588 int rc;
589 KMP_WARNING(RealTimeSchedNotSupported);
590 sched_getparam(0, &param);
591 if (param.sched_priority < max_priority) {
592 param.sched_priority += 1;
593 rc = sched_setscheduler(0, sched, &param);
594 if (rc != 0) {
595 int error = errno;
596 kmp_msg_t err_code = KMP_ERR(error);
597 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
598 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
599 if (__kmp_generate_warnings == kmp_warnings_off) {
600 __kmp_str_free(&err_code.str);
601 }
602 }
603 } else {
604 // We cannot abort here, because number of CPUs may be enough for all
605 // the threads, including the monitor thread, so application could
606 // potentially work...
607 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
608 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
609 __kmp_msg_null);
610 }
611 }
612 // AC: free thread that waits for monitor started
613 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
614 }
615#endif // KMP_REAL_TIME_FIX
616
617 KMP_MB(); /* Flush all pending memory write invalidates. */
618
619 if (__kmp_monitor_wakeups == 1) {
620 interval.tv_sec = 1;
621 interval.tv_nsec = 0;
622 } else {
623 interval.tv_sec = 0;
624 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
625 }
626
627 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
628
629 while (!TCR_4(__kmp_global.g.g_done)) {
630 struct timespec now;
631 struct timeval tval;
632
633 /* This thread monitors the state of the system */
634
635 KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
636
637 status = gettimeofday(&tval, NULL);
638 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
639 TIMEVAL_TO_TIMESPEC(&tval, &now);
640
641 now.tv_sec += interval.tv_sec;
642 now.tv_nsec += interval.tv_nsec;
643
644 if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
645 now.tv_sec += 1;
646 now.tv_nsec -= KMP_NSEC_PER_SEC;
647 }
648
649 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
650 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
651 // AC: the monitor should not fall asleep if g_done has been set
652 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
653 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
654 &__kmp_wait_mx.m_mutex, &now);
655 if (status != 0) {
656 if (status != ETIMEDOUT && status != EINTR) {
657 KMP_SYSFAIL("pthread_cond_timedwait", status);
658 }
659 }
660 }
661 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
662 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
663
664 TCW_4(__kmp_global.g.g_time.dt.t_value,
665 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
666
667 KMP_MB(); /* Flush all pending memory write invalidates. */
668 }
669
670 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
671
672#ifdef KMP_BLOCK_SIGNALS
673 status = sigfillset(&new_set);
674 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
675 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
676 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
677#endif /* KMP_BLOCK_SIGNALS */
678
679 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
680
681 if (__kmp_global.g.g_abort != 0) {
682 /* now we need to terminate the worker threads */
683 /* the value of t_abort is the signal we caught */
684
685 int gtid;
686
687 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
688 __kmp_global.g.g_abort));
689
690 /* terminate the OpenMP worker threads */
691 /* TODO this is not valid for sibling threads!!
692 * the uber master might not be 0 anymore.. */
693 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
694 __kmp_terminate_thread(gtid);
695
696 __kmp_cleanup();
697
698 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
699 __kmp_global.g.g_abort));
700
701 if (__kmp_global.g.g_abort > 0)
702 raise(__kmp_global.g.g_abort);
703 }
704
705 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
706
707 return thr;
708}
709#endif // KMP_USE_MONITOR
710
711void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
712 pthread_t handle;
713 pthread_attr_t thread_attr;
714 int status;
715
716 th->th.th_info.ds.ds_gtid = gtid;
717
718#if KMP_STATS_ENABLED
719 // sets up worker thread stats
720 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
721
722 // th->th.th_stats is used to transfer thread-specific stats-pointer to
723 // __kmp_launch_worker. So when thread is created (goes into
724 // __kmp_launch_worker) it will set its thread local pointer to
725 // th->th.th_stats
726 if (!KMP_UBER_GTID(gtid)) {
727 th->th.th_stats = __kmp_stats_list->push_back(gtid);
728 } else {
729 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
730 // so set the th->th.th_stats field to it.
731 th->th.th_stats = __kmp_stats_thread_ptr;
732 }
733 __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
734
735#endif // KMP_STATS_ENABLED
736
737 if (KMP_UBER_GTID(gtid)) {
738 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
739 th->th.th_info.ds.ds_thread = pthread_self();
740 __kmp_set_stack_info(gtid, th);
741 __kmp_check_stack_overlap(th);
742 return;
743 }
744
745 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
746
747 KMP_MB(); /* Flush all pending memory write invalidates. */
748
749#ifdef KMP_THREAD_ATTR
750 status = pthread_attr_init(&thread_attr);
751 if (status != 0) {
752 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
753 }
754 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
755 if (status != 0) {
756 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
757 }
758
759 /* Set stack size for this thread now.
760 The multiple of 2 is there because on some machines, requesting an unusual
761 stacksize causes the thread to have an offset before the dummy alloca()
762 takes place to create the offset. Since we want the user to have a
763 sufficient stacksize AND support a stack offset, we alloca() twice the
764 offset so that the upcoming alloca() does not eliminate any premade offset,
765 and also gives the user the stack space they requested for all threads */
766 stack_size += gtid * __kmp_stkoffset * 2;
767
768 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
769 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
770 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
771
772#ifdef _POSIX_THREAD_ATTR_STACKSIZE
773 status = pthread_attr_setstacksize(&thread_attr, stack_size);
774#ifdef KMP_BACKUP_STKSIZE
775 if (status != 0) {
776 if (!__kmp_env_stksize) {
777 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
778 __kmp_stksize = KMP_BACKUP_STKSIZE;
779 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
780 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
781 "bytes\n",
782 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
783 status = pthread_attr_setstacksize(&thread_attr, stack_size);
784 }
785 }
786#endif /* KMP_BACKUP_STKSIZE */
787 if (status != 0) {
788 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
789 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
790 }
791#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
792
793#endif /* KMP_THREAD_ATTR */
794
795 status =
796 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
797 if (status != 0 || !handle) { // ??? Why do we check handle??
798#ifdef _POSIX_THREAD_ATTR_STACKSIZE
799 if (status == EINVAL) {
800 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
801 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
802 }
803 if (status == ENOMEM) {
804 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
805 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
806 }
807#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
808 if (status == EAGAIN) {
809 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
810 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
811 }
812 KMP_SYSFAIL("pthread_create", status);
813 }
814
815 th->th.th_info.ds.ds_thread = handle;
816
817#ifdef KMP_THREAD_ATTR
818 status = pthread_attr_destroy(&thread_attr);
819 if (status) {
820 kmp_msg_t err_code = KMP_ERR(status);
821 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
822 __kmp_msg_null);
823 if (__kmp_generate_warnings == kmp_warnings_off) {
824 __kmp_str_free(&err_code.str);
825 }
826 }
827#endif /* KMP_THREAD_ATTR */
828
829 KMP_MB(); /* Flush all pending memory write invalidates. */
830
831 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
832
833} // __kmp_create_worker
834
835#if KMP_USE_MONITOR
836void __kmp_create_monitor(kmp_info_t *th) {
837 pthread_t handle;
838 pthread_attr_t thread_attr;
839 size_t size;
840 int status;
841 int auto_adj_size = FALSE;
842
843 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
844 // We don't need monitor thread in case of MAX_BLOCKTIME
845 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
846 "MAX blocktime\n"));
847 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
848 th->th.th_info.ds.ds_gtid = 0;
849 return;
850 }
851 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
852
853 KMP_MB(); /* Flush all pending memory write invalidates. */
854
855 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
856 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
857#if KMP_REAL_TIME_FIX
858 TCW_4(__kmp_global.g.g_time.dt.t_value,
859 -1); // Will use it for synchronization a bit later.
860#else
861 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
862#endif // KMP_REAL_TIME_FIX
863
864#ifdef KMP_THREAD_ATTR
865 if (__kmp_monitor_stksize == 0) {
866 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
867 auto_adj_size = TRUE;
868 }
869 status = pthread_attr_init(&thread_attr);
870 if (status != 0) {
871 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
872 }
873 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
874 if (status != 0) {
875 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
876 }
877
878#ifdef _POSIX_THREAD_ATTR_STACKSIZE
879 status = pthread_attr_getstacksize(&thread_attr, &size);
880 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
881#else
882 size = __kmp_sys_min_stksize;
883#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
884#endif /* KMP_THREAD_ATTR */
885
886 if (__kmp_monitor_stksize == 0) {
887 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
888 }
889 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
890 __kmp_monitor_stksize = __kmp_sys_min_stksize;
891 }
892
893 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
894 "requested stacksize = %lu bytes\n",
895 size, __kmp_monitor_stksize));
896
897retry:
898
899/* Set stack size for this thread now. */
900#ifdef _POSIX_THREAD_ATTR_STACKSIZE
901 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
902 __kmp_monitor_stksize));
903 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
904 if (status != 0) {
905 if (auto_adj_size) {
906 __kmp_monitor_stksize *= 2;
907 goto retry;
908 }
909 kmp_msg_t err_code = KMP_ERR(status);
910 __kmp_msg(kmp_ms_warning, // should this be fatal? BB
911 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
912 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
913 if (__kmp_generate_warnings == kmp_warnings_off) {
914 __kmp_str_free(&err_code.str);
915 }
916 }
917#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
918
919 status =
920 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
921
922 if (status != 0) {
923#ifdef _POSIX_THREAD_ATTR_STACKSIZE
924 if (status == EINVAL) {
925 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
926 __kmp_monitor_stksize *= 2;
927 goto retry;
928 }
929 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
930 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
931 __kmp_msg_null);
932 }
933 if (status == ENOMEM) {
934 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
935 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
936 __kmp_msg_null);
937 }
938#endif /* _POSIX_THREAD_ATTR_STACKSIZE */
939 if (status == EAGAIN) {
940 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
941 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
942 }
943 KMP_SYSFAIL("pthread_create", status);
944 }
945
946 th->th.th_info.ds.ds_thread = handle;
947
948#if KMP_REAL_TIME_FIX
949 // Wait for the monitor thread is really started and set its *priority*.
950 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
951 sizeof(__kmp_global.g.g_time.dt.t_value));
952 __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1,
953 &__kmp_neq_4, NULL);
954#endif // KMP_REAL_TIME_FIX
955
956#ifdef KMP_THREAD_ATTR
957 status = pthread_attr_destroy(&thread_attr);
958 if (status != 0) {
959 kmp_msg_t err_code = KMP_ERR(status);
960 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
961 __kmp_msg_null);
962 if (__kmp_generate_warnings == kmp_warnings_off) {
963 __kmp_str_free(&err_code.str);
964 }
965 }
966#endif
967
968 KMP_MB(); /* Flush all pending memory write invalidates. */
969
970 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
971 th->th.th_info.ds.ds_thread));
972
973} // __kmp_create_monitor
974#endif // KMP_USE_MONITOR
975
976void __kmp_exit_thread(int exit_status) {
977 pthread_exit((void *)(intptr_t)exit_status);
978} // __kmp_exit_thread
979
980#if KMP_USE_MONITOR
981void __kmp_resume_monitor();
982
983extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
984 int status;
985 void *exit_val;
986
987 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
988 " %#.8lx\n",
989 th->th.th_info.ds.ds_thread));
990
991 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
992 // If both tid and gtid are 0, it means the monitor did not ever start.
993 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
994 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
995 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
996 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
997 return;
998 }
999
1000 KMP_MB(); /* Flush all pending memory write invalidates. */
1001
1002 /* First, check to see whether the monitor thread exists to wake it up. This
1003 is to avoid performance problem when the monitor sleeps during
1004 blocktime-size interval */
1005
1006 status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1007 if (status != ESRCH) {
1008 __kmp_resume_monitor(); // Wake up the monitor thread
1009 }
1010 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1011 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1012 if (exit_val != th) {
1013 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1014 }
1015
1016 th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1017 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1018
1019 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1020 " %#.8lx\n",
1021 th->th.th_info.ds.ds_thread));
1022
1023 KMP_MB(); /* Flush all pending memory write invalidates. */
1024}
1025#else
1026// Empty symbol to export (see exports_so.txt) when
1027// monitor thread feature is disabled
1028extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
1029 (void)th;
1030}
1031#endif // KMP_USE_MONITOR
1032
1033void __kmp_reap_worker(kmp_info_t *th) {
1034 int status;
1035 void *exit_val;
1036
1037 KMP_MB(); /* Flush all pending memory write invalidates. */
1038
1039 KA_TRACE(
1040 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1041
1042 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1043#ifdef KMP_DEBUG
1044 /* Don't expose these to the user until we understand when they trigger */
1045 if (status != 0) {
1046 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1047 }
1048 if (exit_val != th) {
1049 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1050 "exit_val = %p\n",
1051 th->th.th_info.ds.ds_gtid, exit_val));
1052 }
1053#else
1054 (void)status; // unused variable
1055#endif /* KMP_DEBUG */
1056
1057 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1058 th->th.th_info.ds.ds_gtid));
1059
1060 KMP_MB(); /* Flush all pending memory write invalidates. */
1061}
1062
1063#if KMP_HANDLE_SIGNALS
1064
1065static void __kmp_null_handler(int signo) {
1066 // Do nothing, for doing SIG_IGN-type actions.
1067} // __kmp_null_handler
1068
1069static void __kmp_team_handler(int signo) {
1070 if (__kmp_global.g.g_abort == 0) {
1071/* Stage 1 signal handler, let's shut down all of the threads */
1072#ifdef KMP_DEBUG
1073 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1074#endif
1075 switch (signo) {
1076 case SIGHUP:
1077 case SIGINT:
1078 case SIGQUIT:
1079 case SIGILL:
1080 case SIGABRT:
1081 case SIGFPE:
1082 case SIGBUS:
1083 case SIGSEGV:
1084#ifdef SIGSYS
1085 case SIGSYS:
1086#endif
1087 case SIGTERM:
1088 if (__kmp_debug_buf) {
1089 __kmp_dump_debug_buffer();
1090 }
1091 __kmp_unregister_library(); // cleanup shared memory
1092 KMP_MB(); // Flush all pending memory write invalidates.
1093 TCW_4(__kmp_global.g.g_abort, signo);
1094 KMP_MB(); // Flush all pending memory write invalidates.
1095 TCW_4(__kmp_global.g.g_done, TRUE);
1096 KMP_MB(); // Flush all pending memory write invalidates.
1097 break;
1098 default:
1099#ifdef KMP_DEBUG
1100 __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1101#endif
1102 break;
1103 }
1104 }
1105} // __kmp_team_handler
1106
1107static void __kmp_sigaction(int signum, const struct sigaction *act,
1108 struct sigaction *oldact) {
1109 int rc = sigaction(signum, act, oldact);
1110 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1111}
1112
1113static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1114 int parallel_init) {
1115 KMP_MB(); // Flush all pending memory write invalidates.
1116 KB_TRACE(60,
1117 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1118 if (parallel_init) {
1119 struct sigaction new_action;
1120 struct sigaction old_action;
1121 new_action.sa_handler = handler_func;
1122 new_action.sa_flags = 0;
1123 sigfillset(&new_action.sa_mask);
1124 __kmp_sigaction(sig, &new_action, &old_action);
1125 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1126 sigaddset(&__kmp_sigset, sig);
1127 } else {
1128 // Restore/keep user's handler if one previously installed.
1129 __kmp_sigaction(sig, &old_action, NULL);
1130 }
1131 } else {
1132 // Save initial/system signal handlers to see if user handlers installed.
1133 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1134 }
1135 KMP_MB(); // Flush all pending memory write invalidates.
1136} // __kmp_install_one_handler
1137
1138static void __kmp_remove_one_handler(int sig) {
1139 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1140 if (sigismember(&__kmp_sigset, sig)) {
1141 struct sigaction old;
1142 KMP_MB(); // Flush all pending memory write invalidates.
1143 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1144 if ((old.sa_handler != __kmp_team_handler) &&
1145 (old.sa_handler != __kmp_null_handler)) {
1146 // Restore the users signal handler.
1147 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1148 "restoring: sig=%d\n",
1149 sig));
1150 __kmp_sigaction(sig, &old, NULL);
1151 }
1152 sigdelset(&__kmp_sigset, sig);
1153 KMP_MB(); // Flush all pending memory write invalidates.
1154 }
1155} // __kmp_remove_one_handler
1156
1157void __kmp_install_signals(int parallel_init) {
1158 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1159 if (__kmp_handle_signals || !parallel_init) {
1160 // If ! parallel_init, we do not install handlers, just save original
1161 // handlers. Let us do it even __handle_signals is 0.
1162 sigemptyset(&__kmp_sigset);
1163 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1164 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1165 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1166 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1167 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1168 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1169 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1170 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1171#ifdef SIGSYS
1172 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1173#endif // SIGSYS
1174 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1175#ifdef SIGPIPE
1176 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1177#endif // SIGPIPE
1178 }
1179} // __kmp_install_signals
1180
1181void __kmp_remove_signals(void) {
1182 int sig;
1183 KB_TRACE(10, ("__kmp_remove_signals()\n"));
1184 for (sig = 1; sig < NSIG; ++sig) {
1185 __kmp_remove_one_handler(sig);
1186 }
1187} // __kmp_remove_signals
1188
1189#endif // KMP_HANDLE_SIGNALS
1190
1191void __kmp_enable(int new_state) {
1192#ifdef KMP_CANCEL_THREADS
1193 int status, old_state;
1194 status = pthread_setcancelstate(new_state, &old_state);
1195 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1196 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1197#endif
1198}
1199
1200void __kmp_disable(int *old_state) {
1201#ifdef KMP_CANCEL_THREADS
1202 int status;
1203 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1204 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1205#endif
1206}
1207
1208static void __kmp_atfork_prepare(void) {
1209 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1210 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1211}
1212
1213static void __kmp_atfork_parent(void) {
1214 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1215 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1216}
1217
1218/* Reset the library so execution in the child starts "all over again" with
1219 clean data structures in initial states. Don't worry about freeing memory
1220 allocated by parent, just abandon it to be safe. */
1221static void __kmp_atfork_child(void) {
1222 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1223 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1224 /* TODO make sure this is done right for nested/sibling */
1225 // ATT: Memory leaks are here? TODO: Check it and fix.
1226 /* KMP_ASSERT( 0 ); */
1227
1228 ++__kmp_fork_count;
1229
1230#if KMP_AFFINITY_SUPPORTED
1231#if KMP_OS_LINUX || KMP_OS_FREEBSD
1232 // reset the affinity in the child to the initial thread
1233 // affinity in the parent
1234 kmp_set_thread_affinity_mask_initial();
1235#endif
1236 // Set default not to bind threads tightly in the child (we're expecting
1237 // over-subscription after the fork and this can improve things for
1238 // scripting languages that use OpenMP inside process-parallel code).
1239 if (__kmp_nested_proc_bind.bind_types != NULL) {
1240 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1241 }
1242 for (kmp_affinity_t *affinity : __kmp_affinities)
1243 *affinity = KMP_AFFINITY_INIT(affinity->env_var);
1244 __kmp_affin_fullMask = nullptr;
1245 __kmp_affin_origMask = nullptr;
1246 __kmp_topology = nullptr;
1247#endif // KMP_AFFINITY_SUPPORTED
1248
1249#if KMP_USE_MONITOR
1250 __kmp_init_monitor = 0;
1251#endif
1252 __kmp_init_parallel = FALSE;
1253 __kmp_init_middle = FALSE;
1254 __kmp_init_serial = FALSE;
1255 TCW_4(__kmp_init_gtid, FALSE);
1256 __kmp_init_common = FALSE;
1257
1258 TCW_4(__kmp_init_user_locks, FALSE);
1259#if !KMP_USE_DYNAMIC_LOCK
1260 __kmp_user_lock_table.used = 1;
1261 __kmp_user_lock_table.allocated = 0;
1262 __kmp_user_lock_table.table = NULL;
1263 __kmp_lock_blocks = NULL;
1264#endif
1265
1266 __kmp_all_nth = 0;
1267 TCW_4(__kmp_nth, 0);
1268
1269 __kmp_thread_pool = NULL;
1270 __kmp_thread_pool_insert_pt = NULL;
1271 __kmp_team_pool = NULL;
1272
1273 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1274 here so threadprivate doesn't use stale data */
1275 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1276 __kmp_threadpriv_cache_list));
1277
1278 while (__kmp_threadpriv_cache_list != NULL) {
1279
1280 if (*__kmp_threadpriv_cache_list->addr != NULL) {
1281 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1282 &(*__kmp_threadpriv_cache_list->addr)));
1283
1284 *__kmp_threadpriv_cache_list->addr = NULL;
1285 }
1286 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1287 }
1288
1289 __kmp_init_runtime = FALSE;
1290
1291 /* reset statically initialized locks */
1292 __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1293 __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1294 __kmp_init_bootstrap_lock(&__kmp_console_lock);
1295 __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1296
1297#if USE_ITT_BUILD
1298 __kmp_itt_reset(); // reset ITT's global state
1299#endif /* USE_ITT_BUILD */
1300
1301 {
1302 // Child process often get terminated without any use of OpenMP. That might
1303 // cause mapped shared memory file to be left unattended. Thus we postpone
1304 // library registration till middle initialization in the child process.
1305 __kmp_need_register_serial = FALSE;
1306 __kmp_serial_initialize();
1307 }
1308
1309 /* This is necessary to make sure no stale data is left around */
1310 /* AC: customers complain that we use unsafe routines in the atfork
1311 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1312 in dynamic_link when check the presence of shared tbbmalloc library.
1313 Suggestion is to make the library initialization lazier, similar
1314 to what done for __kmpc_begin(). */
1315 // TODO: synchronize all static initializations with regular library
1316 // startup; look at kmp_global.cpp and etc.
1317 //__kmp_internal_begin ();
1318}
1319
1320void __kmp_register_atfork(void) {
1321 if (__kmp_need_register_atfork) {
1322 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1323 __kmp_atfork_child);
1324 KMP_CHECK_SYSFAIL("pthread_atfork", status);
1325 __kmp_need_register_atfork = FALSE;
1326 }
1327}
1328
1329void __kmp_suspend_initialize(void) {
1330 int status;
1331 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1332 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1333 status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1334 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1335}
1336
1337void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1338 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count);
1339 int new_value = __kmp_fork_count + 1;
1340 // Return if already initialized
1341 if (old_value == new_value)
1342 return;
1343 // Wait, then return if being initialized
1344 if (old_value == -1 || !__kmp_atomic_compare_store(
1345 &th->th.th_suspend_init_count, old_value, -1)) {
1346 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) {
1347 KMP_CPU_PAUSE();
1348 }
1349 } else {
1350 // Claim to be the initializer and do initializations
1351 int status;
1352 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1353 &__kmp_suspend_cond_attr);
1354 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1355 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1356 &__kmp_suspend_mutex_attr);
1357 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1358 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value);
1359 }
1360}
1361
1362void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1363 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) {
1364 /* this means we have initialize the suspension pthread objects for this
1365 thread in this instance of the process */
1366 int status;
1367
1368 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1369 if (status != 0 && status != EBUSY) {
1370 KMP_SYSFAIL("pthread_cond_destroy", status);
1371 }
1372 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1373 if (status != 0 && status != EBUSY) {
1374 KMP_SYSFAIL("pthread_mutex_destroy", status);
1375 }
1376 --th->th.th_suspend_init_count;
1377 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) ==
1378 __kmp_fork_count);
1379 }
1380}
1381
1382// return true if lock obtained, false otherwise
1383int __kmp_try_suspend_mx(kmp_info_t *th) {
1384 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1385}
1386
1387void __kmp_lock_suspend_mx(kmp_info_t *th) {
1388 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1389 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1390}
1391
1392void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1393 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1394 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1395}
1396
1397/* This routine puts the calling thread to sleep after setting the
1398 sleep bit for the indicated flag variable to true. */
1399template <class C>
1400static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1401 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1402 kmp_info_t *th = __kmp_threads[th_gtid];
1403 int status;
1404 typename C::flag_t old_spin;
1405
1406 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1407 flag->get()));
1408
1409 __kmp_suspend_initialize_thread(th);
1410
1411 __kmp_lock_suspend_mx(th);
1412
1413 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1414 th_gtid, flag->get()));
1415
1416 /* TODO: shouldn't this use release semantics to ensure that
1417 __kmp_suspend_initialize_thread gets called first? */
1418 old_spin = flag->set_sleeping();
1419 TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1420 th->th.th_sleep_loc_type = flag->get_type();
1421 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1422 __kmp_pause_status != kmp_soft_paused) {
1423 flag->unset_sleeping();
1424 TCW_PTR(th->th.th_sleep_loc, NULL);
1425 th->th.th_sleep_loc_type = flag_unset;
1426 __kmp_unlock_suspend_mx(th);
1427 return;
1428 }
1429 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1430 " was %x\n",
1431 th_gtid, flag->get(), flag->load(), old_spin));
1432
1433 if (flag->done_check_val(old_spin) || flag->done_check()) {
1434 flag->unset_sleeping();
1435 TCW_PTR(th->th.th_sleep_loc, NULL);
1436 th->th.th_sleep_loc_type = flag_unset;
1437 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1438 "for spin(%p)\n",
1439 th_gtid, flag->get()));
1440 } else {
1441 /* Encapsulate in a loop as the documentation states that this may
1442 "with low probability" return when the condition variable has
1443 not been signaled or broadcast */
1444 int deactivated = FALSE;
1445
1446 while (flag->is_sleeping()) {
1447#ifdef DEBUG_SUSPEND
1448 char buffer[128];
1449 __kmp_suspend_count++;
1450 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1451 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1452 buffer);
1453#endif
1454 // Mark the thread as no longer active (only in the first iteration of the
1455 // loop).
1456 if (!deactivated) {
1457 th->th.th_active = FALSE;
1458 if (th->th.th_active_in_pool) {
1459 th->th.th_active_in_pool = FALSE;
1460 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1461 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1462 }
1463 deactivated = TRUE;
1464 }
1465
1466 KMP_DEBUG_ASSERT(th->th.th_sleep_loc);
1467 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type);
1468
1469#if USE_SUSPEND_TIMEOUT
1470 struct timespec now;
1471 struct timeval tval;
1472 int msecs;
1473
1474 status = gettimeofday(&tval, NULL);
1475 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1476 TIMEVAL_TO_TIMESPEC(&tval, &now);
1477
1478 msecs = (4 * __kmp_dflt_blocktime) + 200;
1479 now.tv_sec += msecs / 1000;
1480 now.tv_nsec += (msecs % 1000) * 1000;
1481
1482 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1483 "pthread_cond_timedwait\n",
1484 th_gtid));
1485 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1486 &th->th.th_suspend_mx.m_mutex, &now);
1487#else
1488 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1489 " pthread_cond_wait\n",
1490 th_gtid));
1491 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1492 &th->th.th_suspend_mx.m_mutex);
1493#endif // USE_SUSPEND_TIMEOUT
1494
1495 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1496 KMP_SYSFAIL("pthread_cond_wait", status);
1497 }
1498
1499 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type());
1500
1501 if (!flag->is_sleeping() &&
1502 ((status == EINTR) || (status == ETIMEDOUT))) {
1503 // if interrupt or timeout, and thread is no longer sleeping, we need to
1504 // make sure sleep_loc gets reset; however, this shouldn't be needed if
1505 // we woke up with resume
1506 flag->unset_sleeping();
1507 TCW_PTR(th->th.th_sleep_loc, NULL);
1508 th->th.th_sleep_loc_type = flag_unset;
1509 }
1510#ifdef KMP_DEBUG
1511 if (status == ETIMEDOUT) {
1512 if (flag->is_sleeping()) {
1513 KF_TRACE(100,
1514 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1515 } else {
1516 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1517 "not set!\n",
1518 th_gtid));
1519 TCW_PTR(th->th.th_sleep_loc, NULL);
1520 th->th.th_sleep_loc_type = flag_unset;
1521 }
1522 } else if (flag->is_sleeping()) {
1523 KF_TRACE(100,
1524 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1525 }
1526#endif
1527 } // while
1528
1529 // Mark the thread as active again (if it was previous marked as inactive)
1530 if (deactivated) {
1531 th->th.th_active = TRUE;
1532 if (TCR_4(th->th.th_in_pool)) {
1533 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1534 th->th.th_active_in_pool = TRUE;
1535 }
1536 }
1537 }
1538 // We may have had the loop variable set before entering the loop body;
1539 // so we need to reset sleep_loc.
1540 TCW_PTR(th->th.th_sleep_loc, NULL);
1541 th->th.th_sleep_loc_type = flag_unset;
1542
1543 KMP_DEBUG_ASSERT(!flag->is_sleeping());
1544 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc);
1545#ifdef DEBUG_SUSPEND
1546 {
1547 char buffer[128];
1548 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1549 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1550 buffer);
1551 }
1552#endif
1553
1554 __kmp_unlock_suspend_mx(th);
1555 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1556}
1557
1558template <bool C, bool S>
1559void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) {
1560 __kmp_suspend_template(th_gtid, flag);
1561}
1562template <bool C, bool S>
1563void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) {
1564 __kmp_suspend_template(th_gtid, flag);
1565}
1566template <bool C, bool S>
1567void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) {
1568 __kmp_suspend_template(th_gtid, flag);
1569}
1570void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1571 __kmp_suspend_template(th_gtid, flag);
1572}
1573
1574template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *);
1575template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *);
1576template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *);
1577template void
1578__kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1579template void
1580__kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *);
1581
1582/* This routine signals the thread specified by target_gtid to wake up
1583 after setting the sleep bit indicated by the flag argument to FALSE.
1584 The target thread must already have called __kmp_suspend_template() */
1585template <class C>
1586static inline void __kmp_resume_template(int target_gtid, C *flag) {
1587 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1588 kmp_info_t *th = __kmp_threads[target_gtid];
1589 int status;
1590
1591#ifdef KMP_DEBUG
1592 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1593#endif
1594
1595 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1596 gtid, target_gtid));
1597 KMP_DEBUG_ASSERT(gtid != target_gtid);
1598
1599 __kmp_suspend_initialize_thread(th);
1600
1601 __kmp_lock_suspend_mx(th);
1602
1603 if (!flag || flag != th->th.th_sleep_loc) {
1604 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a
1605 // different location; wake up at new location
1606 flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1607 }
1608
1609 // First, check if the flag is null or its type has changed. If so, someone
1610 // else woke it up.
1611 if (!flag) { // Thread doesn't appear to be sleeping on anything
1612 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1613 "awake: flag(%p)\n",
1614 gtid, target_gtid, (void *)NULL));
1615 __kmp_unlock_suspend_mx(th);
1616 return;
1617 } else if (flag->get_type() != th->th.th_sleep_loc_type) {
1618 // Flag type does not appear to match this function template; possibly the
1619 // thread is sleeping on something else. Try null resume again.
1620 KF_TRACE(
1621 5,
1622 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), "
1623 "spin(%p) type=%d ptr_type=%d\n",
1624 gtid, target_gtid, flag, flag->get(), flag->get_type(),
1625 th->th.th_sleep_loc_type));
1626 __kmp_unlock_suspend_mx(th);
1627 __kmp_null_resume_wrapper(th);
1628 return;
1629 } else { // if multiple threads are sleeping, flag should be internally
1630 // referring to a specific thread here
1631 if (!flag->is_sleeping()) {
1632 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1633 "awake: flag(%p): %u\n",
1634 gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1635 __kmp_unlock_suspend_mx(th);
1636 return;
1637 }
1638 }
1639 KMP_DEBUG_ASSERT(flag);
1640 flag->unset_sleeping();
1641 TCW_PTR(th->th.th_sleep_loc, NULL);
1642 th->th.th_sleep_loc_type = flag_unset;
1643
1644 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1645 "sleep bit for flag's loc(%p): %u\n",
1646 gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1647
1648#ifdef DEBUG_SUSPEND
1649 {
1650 char buffer[128];
1651 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1652 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1653 target_gtid, buffer);
1654 }
1655#endif
1656 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1657 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1658 __kmp_unlock_suspend_mx(th);
1659 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1660 " for T#%d\n",
1661 gtid, target_gtid));
1662}
1663
1664template <bool C, bool S>
1665void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) {
1666 __kmp_resume_template(target_gtid, flag);
1667}
1668template <bool C, bool S>
1669void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) {
1670 __kmp_resume_template(target_gtid, flag);
1671}
1672template <bool C, bool S>
1673void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) {
1674 __kmp_resume_template(target_gtid, flag);
1675}
1676void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1677 __kmp_resume_template(target_gtid, flag);
1678}
1679
1680template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *);
1681template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *);
1682template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *);
1683template void
1684__kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1685
1686#if KMP_USE_MONITOR
1687void __kmp_resume_monitor() {
1688 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1689 int status;
1690#ifdef KMP_DEBUG
1691 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1692 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1693 KMP_GTID_MONITOR));
1694 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1695#endif
1696 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1697 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1698#ifdef DEBUG_SUSPEND
1699 {
1700 char buffer[128];
1701 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1702 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1703 KMP_GTID_MONITOR, buffer);
1704 }
1705#endif
1706 status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1707 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1708 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1709 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1710 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1711 " for T#%d\n",
1712 gtid, KMP_GTID_MONITOR));
1713}
1714#endif // KMP_USE_MONITOR
1715
1716void __kmp_yield() { sched_yield(); }
1717
1718void __kmp_gtid_set_specific(int gtid) {
1719 if (__kmp_init_gtid) {
1720 int status;
1721 status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1722 (void *)(intptr_t)(gtid + 1));
1723 KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1724 } else {
1725 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1726 }
1727}
1728
1729int __kmp_gtid_get_specific() {
1730 int gtid;
1731 if (!__kmp_init_gtid) {
1732 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1733 "KMP_GTID_SHUTDOWN\n"));
1734 return KMP_GTID_SHUTDOWN;
1735 }
1736 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1737 if (gtid == 0) {
1738 gtid = KMP_GTID_DNE;
1739 } else {
1740 gtid--;
1741 }
1742 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1743 __kmp_gtid_threadprivate_key, gtid));
1744 return gtid;
1745}
1746
1747double __kmp_read_cpu_time(void) {
1748 /*clock_t t;*/
1749 struct tms buffer;
1750
1751 /*t =*/times(&buffer);
1752
1753 return (double)(buffer.tms_utime + buffer.tms_cutime) /
1754 (double)CLOCKS_PER_SEC;
1755}
1756
1757int __kmp_read_system_info(struct kmp_sys_info *info) {
1758 int status;
1759 struct rusage r_usage;
1760
1761 memset(info, 0, sizeof(*info));
1762
1763 status = getrusage(RUSAGE_SELF, &r_usage);
1764 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1765
1766 // The maximum resident set size utilized (in kilobytes)
1767 info->maxrss = r_usage.ru_maxrss;
1768 // The number of page faults serviced without any I/O
1769 info->minflt = r_usage.ru_minflt;
1770 // The number of page faults serviced that required I/O
1771 info->majflt = r_usage.ru_majflt;
1772 // The number of times a process was "swapped" out of memory
1773 info->nswap = r_usage.ru_nswap;
1774 // The number of times the file system had to perform input
1775 info->inblock = r_usage.ru_inblock;
1776 // The number of times the file system had to perform output
1777 info->oublock = r_usage.ru_oublock;
1778 // The number of times a context switch was voluntarily
1779 info->nvcsw = r_usage.ru_nvcsw;
1780 // The number of times a context switch was forced
1781 info->nivcsw = r_usage.ru_nivcsw;
1782
1783 return (status != 0);
1784}
1785
1786void __kmp_read_system_time(double *delta) {
1787 double t_ns;
1788 struct timeval tval;
1789 struct timespec stop;
1790 int status;
1791
1792 status = gettimeofday(&tval, NULL);
1793 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1794 TIMEVAL_TO_TIMESPEC(&tval, &stop);
1795 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start));
1796 *delta = (t_ns * 1e-9);
1797}
1798
1799void __kmp_clear_system_time(void) {
1800 struct timeval tval;
1801 int status;
1802 status = gettimeofday(&tval, NULL);
1803 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1804 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1805}
1806
1807static int __kmp_get_xproc(void) {
1808
1809 int r = 0;
1810
1811#if KMP_OS_LINUX
1812
1813 __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r));
1814
1815#elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \
1816 KMP_OS_HURD
1817
1818 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r));
1819
1820#elif KMP_OS_DARWIN
1821
1822 // Bug C77011 High "OpenMP Threads and number of active cores".
1823
1824 // Find the number of available CPUs.
1825 kern_return_t rc;
1826 host_basic_info_data_t info;
1827 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1828 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1829 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1830 // Cannot use KA_TRACE() here because this code works before trace support
1831 // is initialized.
1832 r = info.avail_cpus;
1833 } else {
1834 KMP_WARNING(CantGetNumAvailCPU);
1835 KMP_INFORM(AssumedNumCPU);
1836 }
1837
1838#else
1839
1840#error "Unknown or unsupported OS."
1841
1842#endif
1843
1844 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1845
1846} // __kmp_get_xproc
1847
1848int __kmp_read_from_file(char const *path, char const *format, ...) {
1849 int result;
1850 va_list args;
1851
1852 va_start(args, format);
1853 FILE *f = fopen(path, "rb");
1854 if (f == NULL) {
1855 va_end(args);
1856 return 0;
1857 }
1858 result = vfscanf(f, format, args);
1859 fclose(f);
1860 va_end(args);
1861
1862 return result;
1863}
1864
1865void __kmp_runtime_initialize(void) {
1866 int status;
1867 pthread_mutexattr_t mutex_attr;
1868 pthread_condattr_t cond_attr;
1869
1870 if (__kmp_init_runtime) {
1871 return;
1872 }
1873
1874#if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1875 if (!__kmp_cpuinfo.initialized) {
1876 __kmp_query_cpuid(&__kmp_cpuinfo);
1877 }
1878#endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1879
1880 __kmp_xproc = __kmp_get_xproc();
1881
1882#if !KMP_32_BIT_ARCH
1883 struct rlimit rlim;
1884 // read stack size of calling thread, save it as default for worker threads;
1885 // this should be done before reading environment variables
1886 status = getrlimit(RLIMIT_STACK, &rlim);
1887 if (status == 0) { // success?
1888 __kmp_stksize = rlim.rlim_cur;
1889 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed
1890 }
1891#endif /* KMP_32_BIT_ARCH */
1892
1893 if (sysconf(_SC_THREADS)) {
1894
1895 /* Query the maximum number of threads */
1896 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth));
1897 if (__kmp_sys_max_nth == -1) {
1898 /* Unlimited threads for NPTL */
1899 __kmp_sys_max_nth = INT_MAX;
1900 } else if (__kmp_sys_max_nth <= 1) {
1901 /* Can't tell, just use PTHREAD_THREADS_MAX */
1902 __kmp_sys_max_nth = KMP_MAX_NTH;
1903 }
1904
1905 /* Query the minimum stack size */
1906 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1907 if (__kmp_sys_min_stksize <= 1) {
1908 __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1909 }
1910 }
1911
1912 /* Set up minimum number of threads to switch to TLS gtid */
1913 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1914
1915 status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1916 __kmp_internal_end_dest);
1917 KMP_CHECK_SYSFAIL("pthread_key_create", status);
1918 status = pthread_mutexattr_init(&mutex_attr);
1919 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1920 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1921 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1922 status = pthread_mutexattr_destroy(&mutex_attr);
1923 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status);
1924 status = pthread_condattr_init(&cond_attr);
1925 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1926 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1927 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1928 status = pthread_condattr_destroy(&cond_attr);
1929 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status);
1930#if USE_ITT_BUILD
1931 __kmp_itt_initialize();
1932#endif /* USE_ITT_BUILD */
1933
1934 __kmp_init_runtime = TRUE;
1935}
1936
1937void __kmp_runtime_destroy(void) {
1938 int status;
1939
1940 if (!__kmp_init_runtime) {
1941 return; // Nothing to do.
1942 }
1943
1944#if USE_ITT_BUILD
1945 __kmp_itt_destroy();
1946#endif /* USE_ITT_BUILD */
1947
1948 status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1949 KMP_CHECK_SYSFAIL("pthread_key_delete", status);
1950
1951 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1952 if (status != 0 && status != EBUSY) {
1953 KMP_SYSFAIL("pthread_mutex_destroy", status);
1954 }
1955 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1956 if (status != 0 && status != EBUSY) {
1957 KMP_SYSFAIL("pthread_cond_destroy", status);
1958 }
1959#if KMP_AFFINITY_SUPPORTED
1960 __kmp_affinity_uninitialize();
1961#endif
1962
1963 __kmp_init_runtime = FALSE;
1964}
1965
1966/* Put the thread to sleep for a time period */
1967/* NOTE: not currently used anywhere */
1968void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
1969
1970/* Calculate the elapsed wall clock time for the user */
1971void __kmp_elapsed(double *t) {
1972 int status;
1973#ifdef FIX_SGI_CLOCK
1974 struct timespec ts;
1975
1976 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1977 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
1978 *t =
1979 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
1980#else
1981 struct timeval tv;
1982
1983 status = gettimeofday(&tv, NULL);
1984 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1985 *t =
1986 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
1987#endif
1988}
1989
1990/* Calculate the elapsed wall clock tick for the user */
1991void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1992
1993/* Return the current time stamp in nsec */
1994kmp_uint64 __kmp_now_nsec() {
1995 struct timeval t;
1996 gettimeofday(&t, NULL);
1997 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
1998 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
1999 return nsec;
2000}
2001
2002#if KMP_ARCH_X86 || KMP_ARCH_X86_64
2003/* Measure clock ticks per millisecond */
2004void __kmp_initialize_system_tick() {
2005 kmp_uint64 now, nsec2, diff;
2006 kmp_uint64 delay = 1000000; // ~450 usec on most machines.
2007 kmp_uint64 nsec = __kmp_now_nsec();
2008 kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
2009 while ((now = __kmp_hardware_timestamp()) < goal)
2010 ;
2011 nsec2 = __kmp_now_nsec();
2012 diff = nsec2 - nsec;
2013 if (diff > 0) {
2014 double tpus = 1000.0 * (double)(delay + (now - goal)) / (double)diff;
2015 if (tpus > 0.0) {
2016 __kmp_ticks_per_msec = (kmp_uint64)(tpus * 1000.0);
2017 __kmp_ticks_per_usec = (kmp_uint64)tpus;
2018 }
2019 }
2020}
2021#endif
2022
2023/* Determine whether the given address is mapped into the current address
2024 space. */
2025
2026int __kmp_is_address_mapped(void *addr) {
2027
2028 int found = 0;
2029 int rc;
2030
2031#if KMP_OS_LINUX || KMP_OS_HURD
2032
2033 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the
2034 address ranges mapped into the address space. */
2035
2036 char *name = __kmp_str_format("/proc/%d/maps", getpid());
2037 FILE *file = NULL;
2038
2039 file = fopen(name, "r");
2040 KMP_ASSERT(file != NULL);
2041
2042 for (;;) {
2043
2044 void *beginning = NULL;
2045 void *ending = NULL;
2046 char perms[5];
2047
2048 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
2049 if (rc == EOF) {
2050 break;
2051 }
2052 KMP_ASSERT(rc == 3 &&
2053 KMP_STRLEN(perms) == 4); // Make sure all fields are read.
2054
2055 // Ending address is not included in the region, but beginning is.
2056 if ((addr >= beginning) && (addr < ending)) {
2057 perms[2] = 0; // 3th and 4th character does not matter.
2058 if (strcmp(perms, "rw") == 0) {
2059 // Memory we are looking for should be readable and writable.
2060 found = 1;
2061 }
2062 break;
2063 }
2064 }
2065
2066 // Free resources.
2067 fclose(file);
2068 KMP_INTERNAL_FREE(name);
2069#elif KMP_OS_FREEBSD
2070 char *buf;
2071 size_t lstsz;
2072 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()};
2073 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0);
2074 if (rc < 0)
2075 return 0;
2076 // We pass from number of vm entry's semantic
2077 // to size of whole entry map list.
2078 lstsz = lstsz * 4 / 3;
2079 buf = reinterpret_cast<char *>(kmpc_malloc(lstsz));
2080 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0);
2081 if (rc < 0) {
2082 kmpc_free(buf);
2083 return 0;
2084 }
2085
2086 char *lw = buf;
2087 char *up = buf + lstsz;
2088
2089 while (lw < up) {
2090 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw);
2091 size_t cursz = cur->kve_structsize;
2092 if (cursz == 0)
2093 break;
2094 void *start = reinterpret_cast<void *>(cur->kve_start);
2095 void *end = reinterpret_cast<void *>(cur->kve_end);
2096 // Readable/Writable addresses within current map entry
2097 if ((addr >= start) && (addr < end)) {
2098 if ((cur->kve_protection & KVME_PROT_READ) != 0 &&
2099 (cur->kve_protection & KVME_PROT_WRITE) != 0) {
2100 found = 1;
2101 break;
2102 }
2103 }
2104 lw += cursz;
2105 }
2106 kmpc_free(buf);
2107
2108#elif KMP_OS_DARWIN
2109
2110 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2111 using vm interface. */
2112
2113 int buffer;
2114 vm_size_t count;
2115 rc = vm_read_overwrite(
2116 mach_task_self(), // Task to read memory of.
2117 (vm_address_t)(addr), // Address to read from.
2118 1, // Number of bytes to be read.
2119 (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2120 &count // Address of var to save number of read bytes in.
2121 );
2122 if (rc == 0) {
2123 // Memory successfully read.
2124 found = 1;
2125 }
2126
2127#elif KMP_OS_NETBSD
2128
2129 int mib[5];
2130 mib[0] = CTL_VM;
2131 mib[1] = VM_PROC;
2132 mib[2] = VM_PROC_MAP;
2133 mib[3] = getpid();
2134 mib[4] = sizeof(struct kinfo_vmentry);
2135
2136 size_t size;
2137 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2138 KMP_ASSERT(!rc);
2139 KMP_ASSERT(size);
2140
2141 size = size * 4 / 3;
2142 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2143 KMP_ASSERT(kiv);
2144
2145 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2146 KMP_ASSERT(!rc);
2147 KMP_ASSERT(size);
2148
2149 for (size_t i = 0; i < size; i++) {
2150 if (kiv[i].kve_start >= (uint64_t)addr &&
2151 kiv[i].kve_end <= (uint64_t)addr) {
2152 found = 1;
2153 break;
2154 }
2155 }
2156 KMP_INTERNAL_FREE(kiv);
2157#elif KMP_OS_OPENBSD
2158
2159 int mib[3];
2160 mib[0] = CTL_KERN;
2161 mib[1] = KERN_PROC_VMMAP;
2162 mib[2] = getpid();
2163
2164 size_t size;
2165 uint64_t end;
2166 rc = sysctl(mib, 3, NULL, &size, NULL, 0);
2167 KMP_ASSERT(!rc);
2168 KMP_ASSERT(size);
2169 end = size;
2170
2171 struct kinfo_vmentry kiv = {.kve_start = 0};
2172
2173 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) {
2174 KMP_ASSERT(size);
2175 if (kiv.kve_end == end)
2176 break;
2177
2178 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) {
2179 found = 1;
2180 break;
2181 }
2182 kiv.kve_start += 1;
2183 }
2184#elif KMP_OS_DRAGONFLY
2185
2186 // FIXME(DragonFly): Implement this
2187 found = 1;
2188
2189#else
2190
2191#error "Unknown or unsupported OS"
2192
2193#endif
2194
2195 return found;
2196
2197} // __kmp_is_address_mapped
2198
2199#ifdef USE_LOAD_BALANCE
2200
2201#if KMP_OS_DARWIN || KMP_OS_NETBSD
2202
2203// The function returns the rounded value of the system load average
2204// during given time interval which depends on the value of
2205// __kmp_load_balance_interval variable (default is 60 sec, other values
2206// may be 300 sec or 900 sec).
2207// It returns -1 in case of error.
2208int __kmp_get_load_balance(int max) {
2209 double averages[3];
2210 int ret_avg = 0;
2211
2212 int res = getloadavg(averages, 3);
2213
2214 // Check __kmp_load_balance_interval to determine which of averages to use.
2215 // getloadavg() may return the number of samples less than requested that is
2216 // less than 3.
2217 if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2218 ret_avg = (int)averages[0]; // 1 min
2219 } else if ((__kmp_load_balance_interval >= 180 &&
2220 __kmp_load_balance_interval < 600) &&
2221 (res >= 2)) {
2222 ret_avg = (int)averages[1]; // 5 min
2223 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2224 ret_avg = (int)averages[2]; // 15 min
2225 } else { // Error occurred
2226 return -1;
2227 }
2228
2229 return ret_avg;
2230}
2231
2232#else // Linux* OS
2233
2234// The function returns number of running (not sleeping) threads, or -1 in case
2235// of error. Error could be reported if Linux* OS kernel too old (without
2236// "/proc" support). Counting running threads stops if max running threads
2237// encountered.
2238int __kmp_get_load_balance(int max) {
2239 static int permanent_error = 0;
2240 static int glb_running_threads = 0; // Saved count of the running threads for
2241 // the thread balance algorithm
2242 static double glb_call_time = 0; /* Thread balance algorithm call time */
2243
2244 int running_threads = 0; // Number of running threads in the system.
2245
2246 DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2247 struct dirent *proc_entry = NULL;
2248
2249 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2250 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2251 struct dirent *task_entry = NULL;
2252 int task_path_fixed_len;
2253
2254 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2255 int stat_file = -1;
2256 int stat_path_fixed_len;
2257
2258#ifdef KMP_DEBUG
2259 int total_processes = 0; // Total number of processes in system.
2260#endif
2261
2262 double call_time = 0.0;
2263
2264 __kmp_str_buf_init(&task_path);
2265 __kmp_str_buf_init(&stat_path);
2266
2267 __kmp_elapsed(&call_time);
2268
2269 if (glb_call_time &&
2270 (call_time - glb_call_time < __kmp_load_balance_interval)) {
2271 running_threads = glb_running_threads;
2272 goto finish;
2273 }
2274
2275 glb_call_time = call_time;
2276
2277 // Do not spend time on scanning "/proc/" if we have a permanent error.
2278 if (permanent_error) {
2279 running_threads = -1;
2280 goto finish;
2281 }
2282
2283 if (max <= 0) {
2284 max = INT_MAX;
2285 }
2286
2287 // Open "/proc/" directory.
2288 proc_dir = opendir("/proc");
2289 if (proc_dir == NULL) {
2290 // Cannot open "/prroc/". Probably the kernel does not support it. Return an
2291 // error now and in subsequent calls.
2292 running_threads = -1;
2293 permanent_error = 1;
2294 goto finish;
2295 }
2296
2297 // Initialize fixed part of task_path. This part will not change.
2298 __kmp_str_buf_cat(&task_path, "/proc/", 6);
2299 task_path_fixed_len = task_path.used; // Remember number of used characters.
2300
2301 proc_entry = readdir(proc_dir);
2302 while (proc_entry != NULL) {
2303 // Proc entry is a directory and name starts with a digit. Assume it is a
2304 // process' directory.
2305 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2306
2307#ifdef KMP_DEBUG
2308 ++total_processes;
2309#endif
2310 // Make sure init process is the very first in "/proc", so we can replace
2311 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2312 // 1. We are going to check that total_processes == 1 => d_name == "1" is
2313 // true (where "=>" is implication). Since C++ does not have => operator,
2314 // let us replace it with its equivalent: a => b == ! a || b.
2315 KMP_DEBUG_ASSERT(total_processes != 1 ||
2316 strcmp(proc_entry->d_name, "1") == 0);
2317
2318 // Construct task_path.
2319 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2320 __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2321 KMP_STRLEN(proc_entry->d_name));
2322 __kmp_str_buf_cat(&task_path, "/task", 5);
2323
2324 task_dir = opendir(task_path.str);
2325 if (task_dir == NULL) {
2326 // Process can finish between reading "/proc/" directory entry and
2327 // opening process' "task/" directory. So, in general case we should not
2328 // complain, but have to skip this process and read the next one. But on
2329 // systems with no "task/" support we will spend lot of time to scan
2330 // "/proc/" tree again and again without any benefit. "init" process
2331 // (its pid is 1) should exist always, so, if we cannot open
2332 // "/proc/1/task/" directory, it means "task/" is not supported by
2333 // kernel. Report an error now and in the future.
2334 if (strcmp(proc_entry->d_name, "1") == 0) {
2335 running_threads = -1;
2336 permanent_error = 1;
2337 goto finish;
2338 }
2339 } else {
2340 // Construct fixed part of stat file path.
2341 __kmp_str_buf_clear(&stat_path);
2342 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2343 __kmp_str_buf_cat(&stat_path, "/", 1);
2344 stat_path_fixed_len = stat_path.used;
2345
2346 task_entry = readdir(task_dir);
2347 while (task_entry != NULL) {
2348 // It is a directory and name starts with a digit.
2349 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2350
2351 // Construct complete stat file path. Easiest way would be:
2352 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2353 // task_entry->d_name );
2354 // but seriae of __kmp_str_buf_cat works a bit faster.
2355 stat_path.used =
2356 stat_path_fixed_len; // Reset stat path to its fixed part.
2357 __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2358 KMP_STRLEN(task_entry->d_name));
2359 __kmp_str_buf_cat(&stat_path, "/stat", 5);
2360
2361 // Note: Low-level API (open/read/close) is used. High-level API
2362 // (fopen/fclose) works ~ 30 % slower.
2363 stat_file = open(stat_path.str, O_RDONLY);
2364 if (stat_file == -1) {
2365 // We cannot report an error because task (thread) can terminate
2366 // just before reading this file.
2367 } else {
2368 /* Content of "stat" file looks like:
2369 24285 (program) S ...
2370
2371 It is a single line (if program name does not include funny
2372 symbols). First number is a thread id, then name of executable
2373 file name in paretheses, then state of the thread. We need just
2374 thread state.
2375
2376 Good news: Length of program name is 15 characters max. Longer
2377 names are truncated.
2378
2379 Thus, we need rather short buffer: 15 chars for program name +
2380 2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2381
2382 Bad news: Program name may contain special symbols like space,
2383 closing parenthesis, or even new line. This makes parsing
2384 "stat" file not 100 % reliable. In case of fanny program names
2385 parsing may fail (report incorrect thread state).
2386
2387 Parsing "status" file looks more promissing (due to different
2388 file structure and escaping special symbols) but reading and
2389 parsing of "status" file works slower.
2390 -- ln
2391 */
2392 char buffer[65];
2393 ssize_t len;
2394 len = read(stat_file, buffer, sizeof(buffer) - 1);
2395 if (len >= 0) {
2396 buffer[len] = 0;
2397 // Using scanf:
2398 // sscanf( buffer, "%*d (%*s) %c ", & state );
2399 // looks very nice, but searching for a closing parenthesis
2400 // works a bit faster.
2401 char *close_parent = strstr(buffer, ") ");
2402 if (close_parent != NULL) {
2403 char state = *(close_parent + 2);
2404 if (state == 'R') {
2405 ++running_threads;
2406 if (running_threads >= max) {
2407 goto finish;
2408 }
2409 }
2410 }
2411 }
2412 close(stat_file);
2413 stat_file = -1;
2414 }
2415 }
2416 task_entry = readdir(task_dir);
2417 }
2418 closedir(task_dir);
2419 task_dir = NULL;
2420 }
2421 }
2422 proc_entry = readdir(proc_dir);
2423 }
2424
2425 // There _might_ be a timing hole where the thread executing this
2426 // code get skipped in the load balance, and running_threads is 0.
2427 // Assert in the debug builds only!!!
2428 KMP_DEBUG_ASSERT(running_threads > 0);
2429 if (running_threads <= 0) {
2430 running_threads = 1;
2431 }
2432
2433finish: // Clean up and exit.
2434 if (proc_dir != NULL) {
2435 closedir(proc_dir);
2436 }
2437 __kmp_str_buf_free(&task_path);
2438 if (task_dir != NULL) {
2439 closedir(task_dir);
2440 }
2441 __kmp_str_buf_free(&stat_path);
2442 if (stat_file != -1) {
2443 close(stat_file);
2444 }
2445
2446 glb_running_threads = running_threads;
2447
2448 return running_threads;
2449
2450} // __kmp_get_load_balance
2451
2452#endif // KMP_OS_DARWIN
2453
2454#endif // USE_LOAD_BALANCE
2455
2456#if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \
2457 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \
2458 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \
2459 KMP_ARCH_ARM || KMP_ARCH_VE)
2460
2461// we really only need the case with 1 argument, because CLANG always build
2462// a struct of pointers to shared variables referenced in the outlined function
2463int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2464 void *p_argv[]
2465#if OMPT_SUPPORT
2466 ,
2467 void **exit_frame_ptr
2468#endif
2469) {
2470#if OMPT_SUPPORT
2471 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2472#endif
2473
2474 switch (argc) {
2475 default:
2476 fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2477 fflush(stderr);
2478 exit(-1);
2479 case 0:
2480 (*pkfn)(&gtid, &tid);
2481 break;
2482 case 1:
2483 (*pkfn)(&gtid, &tid, p_argv[0]);
2484 break;
2485 case 2:
2486 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1]);
2487 break;
2488 case 3:
2489 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2]);
2490 break;
2491 case 4:
2492 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2493 break;
2494 case 5:
2495 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2496 break;
2497 case 6:
2498 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2499 p_argv[5]);
2500 break;
2501 case 7:
2502 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2503 p_argv[5], p_argv[6]);
2504 break;
2505 case 8:
2506 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2507 p_argv[5], p_argv[6], p_argv[7]);
2508 break;
2509 case 9:
2510 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2511 p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2512 break;
2513 case 10:
2514 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2515 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2516 break;
2517 case 11:
2518 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2519 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2520 break;
2521 case 12:
2522 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2523 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2524 p_argv[11]);
2525 break;
2526 case 13:
2527 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2528 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2529 p_argv[11], p_argv[12]);
2530 break;
2531 case 14:
2532 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2533 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2534 p_argv[11], p_argv[12], p_argv[13]);
2535 break;
2536 case 15:
2537 (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2538 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2539 p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2540 break;
2541 }
2542
2543 return 1;
2544}
2545
2546#endif
2547
2548#if KMP_OS_LINUX
2549// Functions for hidden helper task
2550namespace {
2551// Condition variable for initializing hidden helper team
2552pthread_cond_t hidden_helper_threads_initz_cond_var;
2553pthread_mutex_t hidden_helper_threads_initz_lock;
2554volatile int hidden_helper_initz_signaled = FALSE;
2555
2556// Condition variable for deinitializing hidden helper team
2557pthread_cond_t hidden_helper_threads_deinitz_cond_var;
2558pthread_mutex_t hidden_helper_threads_deinitz_lock;
2559volatile int hidden_helper_deinitz_signaled = FALSE;
2560
2561// Condition variable for the wrapper function of main thread
2562pthread_cond_t hidden_helper_main_thread_cond_var;
2563pthread_mutex_t hidden_helper_main_thread_lock;
2564volatile int hidden_helper_main_thread_signaled = FALSE;
2565
2566// Semaphore for worker threads. We don't use condition variable here in case
2567// that when multiple signals are sent at the same time, only one thread might
2568// be waken.
2569sem_t hidden_helper_task_sem;
2570} // namespace
2571
2572void __kmp_hidden_helper_worker_thread_wait() {
2573 int status = sem_wait(&hidden_helper_task_sem);
2574 KMP_CHECK_SYSFAIL("sem_wait", status);
2575}
2576
2577void __kmp_do_initialize_hidden_helper_threads() {
2578 // Initialize condition variable
2579 int status =
2580 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr);
2581 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2582
2583 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr);
2584 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2585
2586 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr);
2587 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2588
2589 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr);
2590 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2591
2592 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr);
2593 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2594
2595 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr);
2596 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2597
2598 // Initialize the semaphore
2599 status = sem_init(&hidden_helper_task_sem, 0, 0);
2600 KMP_CHECK_SYSFAIL("sem_init", status);
2601
2602 // Create a new thread to finish initialization
2603 pthread_t handle;
2604 status = pthread_create(
2605 &handle, nullptr,
2606 [](void *) -> void * {
2607 __kmp_hidden_helper_threads_initz_routine();
2608 return nullptr;
2609 },
2610 nullptr);
2611 KMP_CHECK_SYSFAIL("pthread_create", status);
2612}
2613
2614void __kmp_hidden_helper_threads_initz_wait() {
2615 // Initial thread waits here for the completion of the initialization. The
2616 // condition variable will be notified by main thread of hidden helper teams.
2617 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2618 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2619
2620 if (!TCR_4(hidden_helper_initz_signaled)) {
2621 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var,
2622 &hidden_helper_threads_initz_lock);
2623 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2624 }
2625
2626 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2627 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2628}
2629
2630void __kmp_hidden_helper_initz_release() {
2631 // After all initialization, reset __kmp_init_hidden_helper_threads to false.
2632 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2633 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2634
2635 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var);
2636 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2637
2638 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE);
2639
2640 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2641 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2642}
2643
2644void __kmp_hidden_helper_main_thread_wait() {
2645 // The main thread of hidden helper team will be blocked here. The
2646 // condition variable can only be signal in the destructor of RTL.
2647 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2648 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2649
2650 if (!TCR_4(hidden_helper_main_thread_signaled)) {
2651 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var,
2652 &hidden_helper_main_thread_lock);
2653 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2654 }
2655
2656 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2657 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2658}
2659
2660void __kmp_hidden_helper_main_thread_release() {
2661 // The initial thread of OpenMP RTL should call this function to wake up the
2662 // main thread of hidden helper team.
2663 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2664 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2665
2666 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var);
2667 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
2668
2669 // The hidden helper team is done here
2670 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE);
2671
2672 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2673 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2674}
2675
2676void __kmp_hidden_helper_worker_thread_signal() {
2677 int status = sem_post(&hidden_helper_task_sem);
2678 KMP_CHECK_SYSFAIL("sem_post", status);
2679}
2680
2681void __kmp_hidden_helper_threads_deinitz_wait() {
2682 // Initial thread waits here for the completion of the deinitialization. The
2683 // condition variable will be notified by main thread of hidden helper teams.
2684 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
2685 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2686
2687 if (!TCR_4(hidden_helper_deinitz_signaled)) {
2688 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var,
2689 &hidden_helper_threads_deinitz_lock);
2690 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2691 }
2692
2693 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
2694 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2695}
2696
2697void __kmp_hidden_helper_threads_deinitz_release() {
2698 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
2699 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2700
2701 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var);
2702 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2703
2704 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE);
2705
2706 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
2707 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2708}
2709#else // KMP_OS_LINUX
2710void __kmp_hidden_helper_worker_thread_wait() {
2711 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2712}
2713
2714void __kmp_do_initialize_hidden_helper_threads() {
2715 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2716}
2717
2718void __kmp_hidden_helper_threads_initz_wait() {
2719 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2720}
2721
2722void __kmp_hidden_helper_initz_release() {
2723 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2724}
2725
2726void __kmp_hidden_helper_main_thread_wait() {
2727 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2728}
2729
2730void __kmp_hidden_helper_main_thread_release() {
2731 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2732}
2733
2734void __kmp_hidden_helper_worker_thread_signal() {
2735 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2736}
2737
2738void __kmp_hidden_helper_threads_deinitz_wait() {
2739 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2740}
2741
2742void __kmp_hidden_helper_threads_deinitz_release() {
2743 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2744}
2745#endif // KMP_OS_LINUX
2746
2747// end of file //
#define KMP_INIT_PARTITIONED_TIMERS(name)
Initializes the partitioned timers to begin with name.
Definition kmp_stats.h:940