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 #if !KMP_OS_AIX
33 #include <sys/syscall.h>
34 #endif
35 #include <sys/time.h>
36 #include <sys/times.h>
37 #include <unistd.h>
38
39 #if KMP_OS_LINUX
40 #include <sys/sysinfo.h>
41 #if KMP_USE_FUTEX
42 // We should really include <futex.h>, but that causes compatibility problems on
43 // different Linux* OS distributions that either require that you include (or
44 // break when you try to include) <pci/types.h>. Since all we need is the two
45 // macros below (which are part of the kernel ABI, so can't change) we just
46 // define the constants here and don't include <futex.h>
47 #ifndef FUTEX_WAIT
48 #define FUTEX_WAIT 0
49 #endif
50 #ifndef FUTEX_WAKE
51 #define FUTEX_WAKE 1
52 #endif
53 #endif
54 #elif KMP_OS_DARWIN
55 #include <mach/mach.h>
56 #include <sys/sysctl.h>
57 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD
58 #include <sys/types.h>
59 #include <sys/sysctl.h>
60 #include <sys/user.h>
61 #include <pthread_np.h>
62 #elif KMP_OS_NETBSD || KMP_OS_OPENBSD
63 #include <sys/types.h>
64 #include <sys/sysctl.h>
65 #elif KMP_OS_SOLARIS
66 #include <sys/loadavg.h>
67 #endif
68
69 #include <ctype.h>
70 #include <dirent.h>
71 #include <fcntl.h>
72
73 struct kmp_sys_timer {
74 struct timespec start;
75 };
76
77 #ifndef TIMEVAL_TO_TIMESPEC
78 // Convert timeval to timespec.
79 #define TIMEVAL_TO_TIMESPEC(tv, ts) \
80 do { \
81 (ts)->tv_sec = (tv)->tv_sec; \
82 (ts)->tv_nsec = (tv)->tv_usec * 1000; \
83 } while (0)
84 #endif
85
86 // Convert timespec to nanoseconds.
87 #define TS2NS(timespec) \
88 (((timespec).tv_sec * (long int)1e9) + (timespec).tv_nsec)
89
90 static struct kmp_sys_timer __kmp_sys_timer_data;
91
92 #if KMP_HANDLE_SIGNALS
93 typedef void (*sig_func_t)(int);
94 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG];
95 static sigset_t __kmp_sigset;
96 #endif
97
98 static int __kmp_init_runtime = FALSE;
99
100 static int __kmp_fork_count = 0;
101
102 static pthread_condattr_t __kmp_suspend_cond_attr;
103 static pthread_mutexattr_t __kmp_suspend_mutex_attr;
104
105 static kmp_cond_align_t __kmp_wait_cv;
106 static kmp_mutex_align_t __kmp_wait_mx;
107
108 kmp_uint64 __kmp_ticks_per_msec = 1000000;
109 kmp_uint64 __kmp_ticks_per_usec = 1000;
110
111 #ifdef DEBUG_SUSPEND
__kmp_print_cond(char * buffer,kmp_cond_align_t * cond)112 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
113 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
114 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
115 cond->c_cond.__c_waiting);
116 }
117 #endif
118
119 #if ((KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_AIX) && KMP_AFFINITY_SUPPORTED)
120
121 /* Affinity support */
122
__kmp_affinity_bind_thread(int which)123 void __kmp_affinity_bind_thread(int which) {
124 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
125 "Illegal set affinity operation when not capable");
126
127 kmp_affin_mask_t *mask;
128 KMP_CPU_ALLOC_ON_STACK(mask);
129 KMP_CPU_ZERO(mask);
130 KMP_CPU_SET(which, mask);
131 __kmp_set_system_affinity(mask, TRUE);
132 KMP_CPU_FREE_FROM_STACK(mask);
133 }
134
135 #if KMP_OS_AIX
__kmp_affinity_determine_capable(const char * env_var)136 void __kmp_affinity_determine_capable(const char *env_var) {
137 // All versions of AIX support bindprocessor().
138
139 size_t mask_size = __kmp_xproc / CHAR_BIT;
140 // Round up to byte boundary.
141 if (__kmp_xproc % CHAR_BIT)
142 ++mask_size;
143
144 // Round up to the mask_size_type boundary.
145 if (mask_size % sizeof(__kmp_affin_mask_size))
146 mask_size += sizeof(__kmp_affin_mask_size) -
147 mask_size % sizeof(__kmp_affin_mask_size);
148 KMP_AFFINITY_ENABLE(mask_size);
149 KA_TRACE(10,
150 ("__kmp_affinity_determine_capable: "
151 "AIX OS affinity interface bindprocessor functional (mask size = "
152 "%" KMP_SIZE_T_SPEC ").\n",
153 __kmp_affin_mask_size));
154 }
155
156 #else // !KMP_OS_AIX
157
158 /* Determine if we can access affinity functionality on this version of
159 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
160 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
__kmp_affinity_determine_capable(const char * env_var)161 void __kmp_affinity_determine_capable(const char *env_var) {
162 // Check and see if the OS supports thread affinity.
163
164 #if KMP_OS_LINUX
165 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
166 #define KMP_CPU_SET_TRY_SIZE CACHE_LINE
167 #elif KMP_OS_FREEBSD
168 #define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t))
169 #endif
170
171 int verbose = __kmp_affinity.flags.verbose;
172 int warnings = __kmp_affinity.flags.warnings;
173 enum affinity_type type = __kmp_affinity.type;
174
175 #if KMP_OS_LINUX
176 long gCode;
177 unsigned char *buf;
178 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
179
180 // If the syscall returns a suggestion for the size,
181 // then we don't have to search for an appropriate size.
182 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf);
183 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
184 "initial getaffinity call returned %ld errno = %d\n",
185 gCode, errno));
186
187 if (gCode < 0 && errno != EINVAL) {
188 // System call not supported
189 if (verbose ||
190 (warnings && (type != affinity_none) && (type != affinity_default) &&
191 (type != affinity_disabled))) {
192 int error = errno;
193 kmp_msg_t err_code = KMP_ERR(error);
194 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
195 err_code, __kmp_msg_null);
196 if (__kmp_generate_warnings == kmp_warnings_off) {
197 __kmp_str_free(&err_code.str);
198 }
199 }
200 KMP_AFFINITY_DISABLE();
201 KMP_INTERNAL_FREE(buf);
202 return;
203 } else if (gCode > 0) {
204 // The optimal situation: the OS returns the size of the buffer it expects.
205 KMP_AFFINITY_ENABLE(gCode);
206 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
207 "affinity supported (mask size %d)\n",
208 (int)__kmp_affin_mask_size));
209 KMP_INTERNAL_FREE(buf);
210 return;
211 }
212
213 // Call the getaffinity system call repeatedly with increasing set sizes
214 // until we succeed, or reach an upper bound on the search.
215 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
216 "searching for proper set size\n"));
217 int size;
218 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
219 gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
220 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
221 "getaffinity for mask size %ld returned %ld errno = %d\n",
222 size, gCode, errno));
223
224 if (gCode < 0) {
225 if (errno == ENOSYS) {
226 // We shouldn't get here
227 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
228 "inconsistent OS call behavior: errno == ENOSYS for mask "
229 "size %d\n",
230 size));
231 if (verbose ||
232 (warnings && (type != affinity_none) &&
233 (type != affinity_default) && (type != affinity_disabled))) {
234 int error = errno;
235 kmp_msg_t err_code = KMP_ERR(error);
236 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
237 err_code, __kmp_msg_null);
238 if (__kmp_generate_warnings == kmp_warnings_off) {
239 __kmp_str_free(&err_code.str);
240 }
241 }
242 KMP_AFFINITY_DISABLE();
243 KMP_INTERNAL_FREE(buf);
244 return;
245 }
246 continue;
247 }
248
249 KMP_AFFINITY_ENABLE(gCode);
250 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
251 "affinity supported (mask size %d)\n",
252 (int)__kmp_affin_mask_size));
253 KMP_INTERNAL_FREE(buf);
254 return;
255 }
256 #elif KMP_OS_FREEBSD
257 long gCode;
258 unsigned char *buf;
259 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
260 gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT,
261 reinterpret_cast<cpuset_t *>(buf));
262 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
263 "initial getaffinity call returned %d errno = %d\n",
264 gCode, errno));
265 if (gCode == 0) {
266 KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT);
267 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
268 "affinity supported (mask size %d)\n",
269 (int)__kmp_affin_mask_size));
270 KMP_INTERNAL_FREE(buf);
271 return;
272 }
273 #endif
274 KMP_INTERNAL_FREE(buf);
275
276 // Affinity is not supported
277 KMP_AFFINITY_DISABLE();
278 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
279 "cannot determine mask size - affinity not supported\n"));
280 if (verbose || (warnings && (type != affinity_none) &&
281 (type != affinity_default) && (type != affinity_disabled))) {
282 KMP_WARNING(AffCantGetMaskSize, env_var);
283 }
284 }
285 #endif // KMP_OS_AIX
286 #endif // (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_AIX) &&
287 // KMP_AFFINITY_SUPPORTED
288
289 #if KMP_USE_FUTEX
290
__kmp_futex_determine_capable()291 int __kmp_futex_determine_capable() {
292 int loc = 0;
293 long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
294 int retval = (rc == 0) || (errno != ENOSYS);
295
296 KA_TRACE(10,
297 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
298 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
299 retval ? "" : " not"));
300
301 return retval;
302 }
303
304 #endif // KMP_USE_FUTEX
305
306 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_WASM) && (!KMP_ASM_INTRINS)
307 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
308 use compare_and_store for these routines */
309
__kmp_test_then_or8(volatile kmp_int8 * p,kmp_int8 d)310 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
311 kmp_int8 old_value, new_value;
312
313 old_value = TCR_1(*p);
314 new_value = old_value | d;
315
316 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
317 KMP_CPU_PAUSE();
318 old_value = TCR_1(*p);
319 new_value = old_value | d;
320 }
321 return old_value;
322 }
323
__kmp_test_then_and8(volatile kmp_int8 * p,kmp_int8 d)324 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
325 kmp_int8 old_value, new_value;
326
327 old_value = TCR_1(*p);
328 new_value = old_value & d;
329
330 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
331 KMP_CPU_PAUSE();
332 old_value = TCR_1(*p);
333 new_value = old_value & d;
334 }
335 return old_value;
336 }
337
__kmp_test_then_or32(volatile kmp_uint32 * p,kmp_uint32 d)338 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
339 kmp_uint32 old_value, new_value;
340
341 old_value = TCR_4(*p);
342 new_value = old_value | d;
343
344 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
345 KMP_CPU_PAUSE();
346 old_value = TCR_4(*p);
347 new_value = old_value | d;
348 }
349 return old_value;
350 }
351
__kmp_test_then_and32(volatile kmp_uint32 * p,kmp_uint32 d)352 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
353 kmp_uint32 old_value, new_value;
354
355 old_value = TCR_4(*p);
356 new_value = old_value & d;
357
358 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
359 KMP_CPU_PAUSE();
360 old_value = TCR_4(*p);
361 new_value = old_value & d;
362 }
363 return old_value;
364 }
365
366 #if KMP_ARCH_X86 || KMP_ARCH_WASM
__kmp_test_then_add8(volatile kmp_int8 * p,kmp_int8 d)367 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
368 kmp_int8 old_value, new_value;
369
370 old_value = TCR_1(*p);
371 new_value = old_value + d;
372
373 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
374 KMP_CPU_PAUSE();
375 old_value = TCR_1(*p);
376 new_value = old_value + d;
377 }
378 return old_value;
379 }
380
__kmp_test_then_add64(volatile kmp_int64 * p,kmp_int64 d)381 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
382 kmp_int64 old_value, new_value;
383
384 old_value = TCR_8(*p);
385 new_value = old_value + d;
386
387 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
388 KMP_CPU_PAUSE();
389 old_value = TCR_8(*p);
390 new_value = old_value + d;
391 }
392 return old_value;
393 }
394 #endif /* KMP_ARCH_X86 */
395
__kmp_test_then_or64(volatile kmp_uint64 * p,kmp_uint64 d)396 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
397 kmp_uint64 old_value, new_value;
398
399 old_value = TCR_8(*p);
400 new_value = old_value | d;
401 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
402 KMP_CPU_PAUSE();
403 old_value = TCR_8(*p);
404 new_value = old_value | d;
405 }
406 return old_value;
407 }
408
__kmp_test_then_and64(volatile kmp_uint64 * p,kmp_uint64 d)409 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
410 kmp_uint64 old_value, new_value;
411
412 old_value = TCR_8(*p);
413 new_value = old_value & d;
414 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
415 KMP_CPU_PAUSE();
416 old_value = TCR_8(*p);
417 new_value = old_value & d;
418 }
419 return old_value;
420 }
421
422 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
423
__kmp_terminate_thread(int gtid)424 void __kmp_terminate_thread(int gtid) {
425 int status;
426 kmp_info_t *th = __kmp_threads[gtid];
427
428 if (!th)
429 return;
430
431 #ifdef KMP_CANCEL_THREADS
432 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
433 status = pthread_cancel(th->th.th_info.ds.ds_thread);
434 if (status != 0 && status != ESRCH) {
435 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
436 __kmp_msg_null);
437 }
438 #endif
439 KMP_YIELD(TRUE);
440 } //
441
442 /* Set thread stack info according to values returned by pthread_getattr_np().
443 If values are unreasonable, assume call failed and use incremental stack
444 refinement method instead. Returns TRUE if the stack parameters could be
445 determined exactly, FALSE if incremental refinement is necessary. */
__kmp_set_stack_info(int gtid,kmp_info_t * th)446 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
447 int stack_data;
448 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
449 KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_AIX
450 pthread_attr_t attr;
451 int status;
452 size_t size = 0;
453 void *addr = 0;
454
455 /* Always do incremental stack refinement for ubermaster threads since the
456 initial thread stack range can be reduced by sibling thread creation so
457 pthread_attr_getstack may cause thread gtid aliasing */
458 if (!KMP_UBER_GTID(gtid)) {
459
460 /* Fetch the real thread attributes */
461 status = pthread_attr_init(&attr);
462 KMP_CHECK_SYSFAIL("pthread_attr_init", status);
463 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
464 status = pthread_attr_get_np(pthread_self(), &attr);
465 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
466 #else
467 status = pthread_getattr_np(pthread_self(), &attr);
468 KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
469 #endif
470 status = pthread_attr_getstack(&attr, &addr, &size);
471 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
472 KA_TRACE(60,
473 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
474 " %lu, low addr: %p\n",
475 gtid, size, addr));
476 status = pthread_attr_destroy(&attr);
477 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
478 }
479
480 if (size != 0 && addr != 0) { // was stack parameter determination successful?
481 /* Store the correct base and size */
482 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
483 TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
484 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
485 return TRUE;
486 }
487 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD \
488 || KMP_OS_HURD || KMP_OS_SOLARIS */
489 /* Use incremental refinement starting from initial conservative estimate */
490 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
491 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
492 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
493 return FALSE;
494 }
495
__kmp_launch_worker(void * thr)496 static void *__kmp_launch_worker(void *thr) {
497 int status, old_type, old_state;
498 #ifdef KMP_BLOCK_SIGNALS
499 sigset_t new_set, old_set;
500 #endif /* KMP_BLOCK_SIGNALS */
501 void *exit_val;
502 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
503 KMP_OS_OPENBSD || KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_AIX
504 void *volatile padding = 0;
505 #endif
506 int gtid;
507
508 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
509 __kmp_gtid_set_specific(gtid);
510 #ifdef KMP_TDATA_GTID
511 __kmp_gtid = gtid;
512 #endif
513 #if KMP_STATS_ENABLED
514 // set thread local index to point to thread-specific stats
515 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
516 __kmp_stats_thread_ptr->startLife();
517 KMP_SET_THREAD_STATE(IDLE);
518 KMP_INIT_PARTITIONED_TIMERS(OMP_idle);
519 #endif
520
521 #if USE_ITT_BUILD
522 __kmp_itt_thread_name(gtid);
523 #endif /* USE_ITT_BUILD */
524
525 #if KMP_AFFINITY_SUPPORTED
526 __kmp_affinity_bind_init_mask(gtid);
527 #endif
528
529 #ifdef KMP_CANCEL_THREADS
530 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
531 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
532 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
533 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
534 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
535 #endif
536
537 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
538 // Set FP control regs to be a copy of the parallel initialization thread's.
539 __kmp_clear_x87_fpu_status_word();
540 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
541 __kmp_load_mxcsr(&__kmp_init_mxcsr);
542 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
543
544 #ifdef KMP_BLOCK_SIGNALS
545 status = sigfillset(&new_set);
546 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
547 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
548 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
549 #endif /* KMP_BLOCK_SIGNALS */
550
551 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
552 KMP_OS_OPENBSD || KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_AIX
553 if (__kmp_stkoffset > 0 && gtid > 0) {
554 padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
555 (void)padding;
556 }
557 #endif
558
559 KMP_MB();
560 __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
561
562 __kmp_check_stack_overlap((kmp_info_t *)thr);
563
564 exit_val = __kmp_launch_thread((kmp_info_t *)thr);
565
566 #ifdef KMP_BLOCK_SIGNALS
567 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
568 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
569 #endif /* KMP_BLOCK_SIGNALS */
570
571 return exit_val;
572 }
573
574 #if KMP_USE_MONITOR
575 /* The monitor thread controls all of the threads in the complex */
576
__kmp_launch_monitor(void * thr)577 static void *__kmp_launch_monitor(void *thr) {
578 int status, old_type, old_state;
579 #ifdef KMP_BLOCK_SIGNALS
580 sigset_t new_set;
581 #endif /* KMP_BLOCK_SIGNALS */
582 struct timespec interval;
583
584 KMP_MB(); /* Flush all pending memory write invalidates. */
585
586 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
587
588 /* register us as the monitor thread */
589 __kmp_gtid_set_specific(KMP_GTID_MONITOR);
590 #ifdef KMP_TDATA_GTID
591 __kmp_gtid = KMP_GTID_MONITOR;
592 #endif
593
594 KMP_MB();
595
596 #if USE_ITT_BUILD
597 // Instruct Intel(R) Threading Tools to ignore monitor thread.
598 __kmp_itt_thread_ignore();
599 #endif /* USE_ITT_BUILD */
600
601 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
602 (kmp_info_t *)thr);
603
604 __kmp_check_stack_overlap((kmp_info_t *)thr);
605
606 #ifdef KMP_CANCEL_THREADS
607 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
608 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
609 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
610 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
611 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
612 #endif
613
614 #if KMP_REAL_TIME_FIX
615 // This is a potential fix which allows application with real-time scheduling
616 // policy work. However, decision about the fix is not made yet, so it is
617 // disabled by default.
618 { // Are program started with real-time scheduling policy?
619 int sched = sched_getscheduler(0);
620 if (sched == SCHED_FIFO || sched == SCHED_RR) {
621 // Yes, we are a part of real-time application. Try to increase the
622 // priority of the monitor.
623 struct sched_param param;
624 int max_priority = sched_get_priority_max(sched);
625 int rc;
626 KMP_WARNING(RealTimeSchedNotSupported);
627 sched_getparam(0, ¶m);
628 if (param.sched_priority < max_priority) {
629 param.sched_priority += 1;
630 rc = sched_setscheduler(0, sched, ¶m);
631 if (rc != 0) {
632 int error = errno;
633 kmp_msg_t err_code = KMP_ERR(error);
634 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
635 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
636 if (__kmp_generate_warnings == kmp_warnings_off) {
637 __kmp_str_free(&err_code.str);
638 }
639 }
640 } else {
641 // We cannot abort here, because number of CPUs may be enough for all
642 // the threads, including the monitor thread, so application could
643 // potentially work...
644 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
645 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
646 __kmp_msg_null);
647 }
648 }
649 // AC: free thread that waits for monitor started
650 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
651 }
652 #endif // KMP_REAL_TIME_FIX
653
654 KMP_MB(); /* Flush all pending memory write invalidates. */
655
656 if (__kmp_monitor_wakeups == 1) {
657 interval.tv_sec = 1;
658 interval.tv_nsec = 0;
659 } else {
660 interval.tv_sec = 0;
661 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
662 }
663
664 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
665
666 while (!TCR_4(__kmp_global.g.g_done)) {
667 struct timespec now;
668 struct timeval tval;
669
670 /* This thread monitors the state of the system */
671
672 KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
673
674 status = gettimeofday(&tval, NULL);
675 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
676 TIMEVAL_TO_TIMESPEC(&tval, &now);
677
678 now.tv_sec += interval.tv_sec;
679 now.tv_nsec += interval.tv_nsec;
680
681 if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
682 now.tv_sec += 1;
683 now.tv_nsec -= KMP_NSEC_PER_SEC;
684 }
685
686 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
687 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
688 // AC: the monitor should not fall asleep if g_done has been set
689 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
690 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
691 &__kmp_wait_mx.m_mutex, &now);
692 if (status != 0) {
693 if (status != ETIMEDOUT && status != EINTR) {
694 KMP_SYSFAIL("pthread_cond_timedwait", status);
695 }
696 }
697 }
698 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
699 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
700
701 TCW_4(__kmp_global.g.g_time.dt.t_value,
702 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
703
704 KMP_MB(); /* Flush all pending memory write invalidates. */
705 }
706
707 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
708
709 #ifdef KMP_BLOCK_SIGNALS
710 status = sigfillset(&new_set);
711 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
712 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
713 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
714 #endif /* KMP_BLOCK_SIGNALS */
715
716 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
717
718 if (__kmp_global.g.g_abort != 0) {
719 /* now we need to terminate the worker threads */
720 /* the value of t_abort is the signal we caught */
721
722 int gtid;
723
724 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
725 __kmp_global.g.g_abort));
726
727 /* terminate the OpenMP worker threads */
728 /* TODO this is not valid for sibling threads!!
729 * the uber master might not be 0 anymore.. */
730 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
731 __kmp_terminate_thread(gtid);
732
733 __kmp_cleanup();
734
735 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
736 __kmp_global.g.g_abort));
737
738 if (__kmp_global.g.g_abort > 0)
739 raise(__kmp_global.g.g_abort);
740 }
741
742 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
743
744 return thr;
745 }
746 #endif // KMP_USE_MONITOR
747
__kmp_create_worker(int gtid,kmp_info_t * th,size_t stack_size)748 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
749 pthread_t handle;
750 pthread_attr_t thread_attr;
751 int status;
752
753 th->th.th_info.ds.ds_gtid = gtid;
754
755 #if KMP_STATS_ENABLED
756 // sets up worker thread stats
757 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
758
759 // th->th.th_stats is used to transfer thread-specific stats-pointer to
760 // __kmp_launch_worker. So when thread is created (goes into
761 // __kmp_launch_worker) it will set its thread local pointer to
762 // th->th.th_stats
763 if (!KMP_UBER_GTID(gtid)) {
764 th->th.th_stats = __kmp_stats_list->push_back(gtid);
765 } else {
766 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
767 // so set the th->th.th_stats field to it.
768 th->th.th_stats = __kmp_stats_thread_ptr;
769 }
770 __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
771
772 #endif // KMP_STATS_ENABLED
773
774 if (KMP_UBER_GTID(gtid)) {
775 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
776 th->th.th_info.ds.ds_thread = pthread_self();
777 __kmp_set_stack_info(gtid, th);
778 __kmp_check_stack_overlap(th);
779 return;
780 }
781
782 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
783
784 KMP_MB(); /* Flush all pending memory write invalidates. */
785
786 #ifdef KMP_THREAD_ATTR
787 status = pthread_attr_init(&thread_attr);
788 if (status != 0) {
789 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
790 }
791 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
792 if (status != 0) {
793 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
794 }
795
796 /* Set stack size for this thread now.
797 The multiple of 2 is there because on some machines, requesting an unusual
798 stacksize causes the thread to have an offset before the dummy alloca()
799 takes place to create the offset. Since we want the user to have a
800 sufficient stacksize AND support a stack offset, we alloca() twice the
801 offset so that the upcoming alloca() does not eliminate any premade offset,
802 and also gives the user the stack space they requested for all threads */
803 stack_size += gtid * __kmp_stkoffset * 2;
804
805 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
806 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
807 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
808
809 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
810 status = pthread_attr_setstacksize(&thread_attr, stack_size);
811 #ifdef KMP_BACKUP_STKSIZE
812 if (status != 0) {
813 if (!__kmp_env_stksize) {
814 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
815 __kmp_stksize = KMP_BACKUP_STKSIZE;
816 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
817 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
818 "bytes\n",
819 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
820 status = pthread_attr_setstacksize(&thread_attr, stack_size);
821 }
822 }
823 #endif /* KMP_BACKUP_STKSIZE */
824 if (status != 0) {
825 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
826 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
827 }
828 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
829
830 #endif /* KMP_THREAD_ATTR */
831
832 status =
833 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
834 if (status != 0 || !handle) { // ??? Why do we check handle??
835 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
836 if (status == EINVAL) {
837 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
838 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
839 }
840 if (status == ENOMEM) {
841 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
842 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
843 }
844 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
845 if (status == EAGAIN) {
846 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
847 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
848 }
849 KMP_SYSFAIL("pthread_create", status);
850 }
851
852 th->th.th_info.ds.ds_thread = handle;
853
854 #ifdef KMP_THREAD_ATTR
855 status = pthread_attr_destroy(&thread_attr);
856 if (status) {
857 kmp_msg_t err_code = KMP_ERR(status);
858 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
859 __kmp_msg_null);
860 if (__kmp_generate_warnings == kmp_warnings_off) {
861 __kmp_str_free(&err_code.str);
862 }
863 }
864 #endif /* KMP_THREAD_ATTR */
865
866 KMP_MB(); /* Flush all pending memory write invalidates. */
867
868 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
869
870 } // __kmp_create_worker
871
872 #if KMP_USE_MONITOR
__kmp_create_monitor(kmp_info_t * th)873 void __kmp_create_monitor(kmp_info_t *th) {
874 pthread_t handle;
875 pthread_attr_t thread_attr;
876 size_t size;
877 int status;
878 int auto_adj_size = FALSE;
879
880 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
881 // We don't need monitor thread in case of MAX_BLOCKTIME
882 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
883 "MAX blocktime\n"));
884 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
885 th->th.th_info.ds.ds_gtid = 0;
886 return;
887 }
888 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
889
890 KMP_MB(); /* Flush all pending memory write invalidates. */
891
892 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
893 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
894 #if KMP_REAL_TIME_FIX
895 TCW_4(__kmp_global.g.g_time.dt.t_value,
896 -1); // Will use it for synchronization a bit later.
897 #else
898 TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
899 #endif // KMP_REAL_TIME_FIX
900
901 #ifdef KMP_THREAD_ATTR
902 if (__kmp_monitor_stksize == 0) {
903 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
904 auto_adj_size = TRUE;
905 }
906 status = pthread_attr_init(&thread_attr);
907 if (status != 0) {
908 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
909 }
910 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
911 if (status != 0) {
912 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
913 }
914
915 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
916 status = pthread_attr_getstacksize(&thread_attr, &size);
917 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
918 #else
919 size = __kmp_sys_min_stksize;
920 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
921 #endif /* KMP_THREAD_ATTR */
922
923 if (__kmp_monitor_stksize == 0) {
924 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
925 }
926 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
927 __kmp_monitor_stksize = __kmp_sys_min_stksize;
928 }
929
930 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
931 "requested stacksize = %lu bytes\n",
932 size, __kmp_monitor_stksize));
933
934 retry:
935
936 /* Set stack size for this thread now. */
937 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
938 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
939 __kmp_monitor_stksize));
940 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
941 if (status != 0) {
942 if (auto_adj_size) {
943 __kmp_monitor_stksize *= 2;
944 goto retry;
945 }
946 kmp_msg_t err_code = KMP_ERR(status);
947 __kmp_msg(kmp_ms_warning, // should this be fatal? BB
948 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
949 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
950 if (__kmp_generate_warnings == kmp_warnings_off) {
951 __kmp_str_free(&err_code.str);
952 }
953 }
954 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
955
956 status =
957 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
958
959 if (status != 0) {
960 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
961 if (status == EINVAL) {
962 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
963 __kmp_monitor_stksize *= 2;
964 goto retry;
965 }
966 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
967 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
968 __kmp_msg_null);
969 }
970 if (status == ENOMEM) {
971 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
972 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
973 __kmp_msg_null);
974 }
975 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
976 if (status == EAGAIN) {
977 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
978 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
979 }
980 KMP_SYSFAIL("pthread_create", status);
981 }
982
983 th->th.th_info.ds.ds_thread = handle;
984
985 #if KMP_REAL_TIME_FIX
986 // Wait for the monitor thread is really started and set its *priority*.
987 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
988 sizeof(__kmp_global.g.g_time.dt.t_value));
989 __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1,
990 &__kmp_neq_4, NULL);
991 #endif // KMP_REAL_TIME_FIX
992
993 #ifdef KMP_THREAD_ATTR
994 status = pthread_attr_destroy(&thread_attr);
995 if (status != 0) {
996 kmp_msg_t err_code = KMP_ERR(status);
997 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
998 __kmp_msg_null);
999 if (__kmp_generate_warnings == kmp_warnings_off) {
1000 __kmp_str_free(&err_code.str);
1001 }
1002 }
1003 #endif
1004
1005 KMP_MB(); /* Flush all pending memory write invalidates. */
1006
1007 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
1008 th->th.th_info.ds.ds_thread));
1009
1010 } // __kmp_create_monitor
1011 #endif // KMP_USE_MONITOR
1012
__kmp_exit_thread(int exit_status)1013 void __kmp_exit_thread(int exit_status) {
1014 #if KMP_OS_WASI
1015 // TODO: the wasm32-wasi-threads target does not yet support pthread_exit.
1016 #else
1017 pthread_exit((void *)(intptr_t)exit_status);
1018 #endif
1019 } // __kmp_exit_thread
1020
1021 #if KMP_USE_MONITOR
1022 void __kmp_resume_monitor();
1023
__kmp_reap_monitor(kmp_info_t * th)1024 extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
1025 int status;
1026 void *exit_val;
1027
1028 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
1029 " %#.8lx\n",
1030 th->th.th_info.ds.ds_thread));
1031
1032 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1033 // If both tid and gtid are 0, it means the monitor did not ever start.
1034 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1035 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1036 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1037 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1038 return;
1039 }
1040
1041 KMP_MB(); /* Flush all pending memory write invalidates. */
1042
1043 /* First, check to see whether the monitor thread exists to wake it up. This
1044 is to avoid performance problem when the monitor sleeps during
1045 blocktime-size interval */
1046
1047 status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1048 if (status != ESRCH) {
1049 __kmp_resume_monitor(); // Wake up the monitor thread
1050 }
1051 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1052 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1053 if (exit_val != th) {
1054 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1055 }
1056
1057 th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1058 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1059
1060 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1061 " %#.8lx\n",
1062 th->th.th_info.ds.ds_thread));
1063
1064 KMP_MB(); /* Flush all pending memory write invalidates. */
1065 }
1066 #else
1067 // Empty symbol to export (see exports_so.txt) when
1068 // monitor thread feature is disabled
__kmp_reap_monitor(kmp_info_t * th)1069 extern "C" void __kmp_reap_monitor(kmp_info_t *th) {
1070 (void)th;
1071 }
1072 #endif // KMP_USE_MONITOR
1073
__kmp_reap_worker(kmp_info_t * th)1074 void __kmp_reap_worker(kmp_info_t *th) {
1075 int status;
1076 void *exit_val;
1077
1078 KMP_MB(); /* Flush all pending memory write invalidates. */
1079
1080 KA_TRACE(
1081 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1082
1083 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1084 #ifdef KMP_DEBUG
1085 /* Don't expose these to the user until we understand when they trigger */
1086 if (status != 0) {
1087 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1088 }
1089 if (exit_val != th) {
1090 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1091 "exit_val = %p\n",
1092 th->th.th_info.ds.ds_gtid, exit_val));
1093 }
1094 #else
1095 (void)status; // unused variable
1096 #endif /* KMP_DEBUG */
1097
1098 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1099 th->th.th_info.ds.ds_gtid));
1100
1101 KMP_MB(); /* Flush all pending memory write invalidates. */
1102 }
1103
1104 #if KMP_HANDLE_SIGNALS
1105
__kmp_null_handler(int signo)1106 static void __kmp_null_handler(int signo) {
1107 // Do nothing, for doing SIG_IGN-type actions.
1108 } // __kmp_null_handler
1109
__kmp_team_handler(int signo)1110 static void __kmp_team_handler(int signo) {
1111 if (__kmp_global.g.g_abort == 0) {
1112 /* Stage 1 signal handler, let's shut down all of the threads */
1113 #ifdef KMP_DEBUG
1114 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1115 #endif
1116 switch (signo) {
1117 case SIGHUP:
1118 case SIGINT:
1119 case SIGQUIT:
1120 case SIGILL:
1121 case SIGABRT:
1122 case SIGFPE:
1123 case SIGBUS:
1124 case SIGSEGV:
1125 #ifdef SIGSYS
1126 case SIGSYS:
1127 #endif
1128 case SIGTERM:
1129 if (__kmp_debug_buf) {
1130 __kmp_dump_debug_buffer();
1131 }
1132 __kmp_unregister_library(); // cleanup shared memory
1133 KMP_MB(); // Flush all pending memory write invalidates.
1134 TCW_4(__kmp_global.g.g_abort, signo);
1135 KMP_MB(); // Flush all pending memory write invalidates.
1136 TCW_4(__kmp_global.g.g_done, TRUE);
1137 KMP_MB(); // Flush all pending memory write invalidates.
1138 break;
1139 default:
1140 #ifdef KMP_DEBUG
1141 __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1142 #endif
1143 break;
1144 }
1145 }
1146 } // __kmp_team_handler
1147
__kmp_sigaction(int signum,const struct sigaction * act,struct sigaction * oldact)1148 static void __kmp_sigaction(int signum, const struct sigaction *act,
1149 struct sigaction *oldact) {
1150 int rc = sigaction(signum, act, oldact);
1151 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1152 }
1153
__kmp_install_one_handler(int sig,sig_func_t handler_func,int parallel_init)1154 static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1155 int parallel_init) {
1156 KMP_MB(); // Flush all pending memory write invalidates.
1157 KB_TRACE(60,
1158 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1159 if (parallel_init) {
1160 struct sigaction new_action;
1161 struct sigaction old_action;
1162 new_action.sa_handler = handler_func;
1163 new_action.sa_flags = 0;
1164 sigfillset(&new_action.sa_mask);
1165 __kmp_sigaction(sig, &new_action, &old_action);
1166 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1167 sigaddset(&__kmp_sigset, sig);
1168 } else {
1169 // Restore/keep user's handler if one previously installed.
1170 __kmp_sigaction(sig, &old_action, NULL);
1171 }
1172 } else {
1173 // Save initial/system signal handlers to see if user handlers installed.
1174 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1175 }
1176 KMP_MB(); // Flush all pending memory write invalidates.
1177 } // __kmp_install_one_handler
1178
__kmp_remove_one_handler(int sig)1179 static void __kmp_remove_one_handler(int sig) {
1180 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1181 if (sigismember(&__kmp_sigset, sig)) {
1182 struct sigaction old;
1183 KMP_MB(); // Flush all pending memory write invalidates.
1184 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1185 if ((old.sa_handler != __kmp_team_handler) &&
1186 (old.sa_handler != __kmp_null_handler)) {
1187 // Restore the users signal handler.
1188 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1189 "restoring: sig=%d\n",
1190 sig));
1191 __kmp_sigaction(sig, &old, NULL);
1192 }
1193 sigdelset(&__kmp_sigset, sig);
1194 KMP_MB(); // Flush all pending memory write invalidates.
1195 }
1196 } // __kmp_remove_one_handler
1197
__kmp_install_signals(int parallel_init)1198 void __kmp_install_signals(int parallel_init) {
1199 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1200 if (__kmp_handle_signals || !parallel_init) {
1201 // If ! parallel_init, we do not install handlers, just save original
1202 // handlers. Let us do it even __handle_signals is 0.
1203 sigemptyset(&__kmp_sigset);
1204 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1205 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1206 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1207 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1208 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1209 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1210 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1211 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1212 #ifdef SIGSYS
1213 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1214 #endif // SIGSYS
1215 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1216 #ifdef SIGPIPE
1217 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1218 #endif // SIGPIPE
1219 }
1220 } // __kmp_install_signals
1221
__kmp_remove_signals(void)1222 void __kmp_remove_signals(void) {
1223 int sig;
1224 KB_TRACE(10, ("__kmp_remove_signals()\n"));
1225 for (sig = 1; sig < NSIG; ++sig) {
1226 __kmp_remove_one_handler(sig);
1227 }
1228 } // __kmp_remove_signals
1229
1230 #endif // KMP_HANDLE_SIGNALS
1231
__kmp_enable(int new_state)1232 void __kmp_enable(int new_state) {
1233 #ifdef KMP_CANCEL_THREADS
1234 int status, old_state;
1235 status = pthread_setcancelstate(new_state, &old_state);
1236 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1237 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1238 #endif
1239 }
1240
__kmp_disable(int * old_state)1241 void __kmp_disable(int *old_state) {
1242 #ifdef KMP_CANCEL_THREADS
1243 int status;
1244 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1245 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1246 #endif
1247 }
1248
__kmp_atfork_prepare(void)1249 static void __kmp_atfork_prepare(void) {
1250 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1251 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1252 }
1253
__kmp_atfork_parent(void)1254 static void __kmp_atfork_parent(void) {
1255 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1256 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1257 }
1258
1259 /* Reset the library so execution in the child starts "all over again" with
1260 clean data structures in initial states. Don't worry about freeing memory
1261 allocated by parent, just abandon it to be safe. */
__kmp_atfork_child(void)1262 static void __kmp_atfork_child(void) {
1263 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1264 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1265 /* TODO make sure this is done right for nested/sibling */
1266 // ATT: Memory leaks are here? TODO: Check it and fix.
1267 /* KMP_ASSERT( 0 ); */
1268
1269 ++__kmp_fork_count;
1270
1271 #if KMP_AFFINITY_SUPPORTED
1272 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_AIX
1273 // reset the affinity in the child to the initial thread
1274 // affinity in the parent
1275 kmp_set_thread_affinity_mask_initial();
1276 #endif
1277 // Set default not to bind threads tightly in the child (we're expecting
1278 // over-subscription after the fork and this can improve things for
1279 // scripting languages that use OpenMP inside process-parallel code).
1280 if (__kmp_nested_proc_bind.bind_types != NULL) {
1281 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1282 }
1283 for (kmp_affinity_t *affinity : __kmp_affinities)
1284 *affinity = KMP_AFFINITY_INIT(affinity->env_var);
1285 __kmp_affin_fullMask = nullptr;
1286 __kmp_affin_origMask = nullptr;
1287 __kmp_topology = nullptr;
1288 #endif // KMP_AFFINITY_SUPPORTED
1289
1290 #if KMP_USE_MONITOR
1291 __kmp_init_monitor = 0;
1292 #endif
1293 __kmp_init_parallel = FALSE;
1294 __kmp_init_middle = FALSE;
1295 __kmp_init_serial = FALSE;
1296 TCW_4(__kmp_init_gtid, FALSE);
1297 __kmp_init_common = FALSE;
1298
1299 TCW_4(__kmp_init_user_locks, FALSE);
1300 #if !KMP_USE_DYNAMIC_LOCK
1301 __kmp_user_lock_table.used = 1;
1302 __kmp_user_lock_table.allocated = 0;
1303 __kmp_user_lock_table.table = NULL;
1304 __kmp_lock_blocks = NULL;
1305 #endif
1306
1307 __kmp_all_nth = 0;
1308 TCW_4(__kmp_nth, 0);
1309
1310 __kmp_thread_pool = NULL;
1311 __kmp_thread_pool_insert_pt = NULL;
1312 __kmp_team_pool = NULL;
1313
1314 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1315 here so threadprivate doesn't use stale data */
1316 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1317 __kmp_threadpriv_cache_list));
1318
1319 while (__kmp_threadpriv_cache_list != NULL) {
1320
1321 if (*__kmp_threadpriv_cache_list->addr != NULL) {
1322 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1323 &(*__kmp_threadpriv_cache_list->addr)));
1324
1325 *__kmp_threadpriv_cache_list->addr = NULL;
1326 }
1327 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1328 }
1329
1330 __kmp_init_runtime = FALSE;
1331
1332 /* reset statically initialized locks */
1333 __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1334 __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1335 __kmp_init_bootstrap_lock(&__kmp_console_lock);
1336 __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1337
1338 #if USE_ITT_BUILD
1339 __kmp_itt_reset(); // reset ITT's global state
1340 #endif /* USE_ITT_BUILD */
1341
1342 {
1343 // Child process often get terminated without any use of OpenMP. That might
1344 // cause mapped shared memory file to be left unattended. Thus we postpone
1345 // library registration till middle initialization in the child process.
1346 __kmp_need_register_serial = FALSE;
1347 __kmp_serial_initialize();
1348 }
1349
1350 /* This is necessary to make sure no stale data is left around */
1351 /* AC: customers complain that we use unsafe routines in the atfork
1352 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1353 in dynamic_link when check the presence of shared tbbmalloc library.
1354 Suggestion is to make the library initialization lazier, similar
1355 to what done for __kmpc_begin(). */
1356 // TODO: synchronize all static initializations with regular library
1357 // startup; look at kmp_global.cpp and etc.
1358 //__kmp_internal_begin ();
1359 }
1360
__kmp_register_atfork(void)1361 void __kmp_register_atfork(void) {
1362 if (__kmp_need_register_atfork) {
1363 #if !KMP_OS_WASI
1364 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1365 __kmp_atfork_child);
1366 KMP_CHECK_SYSFAIL("pthread_atfork", status);
1367 #endif
1368 __kmp_need_register_atfork = FALSE;
1369 }
1370 }
1371
__kmp_suspend_initialize(void)1372 void __kmp_suspend_initialize(void) {
1373 int status;
1374 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1375 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1376 status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1377 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1378 }
1379
__kmp_suspend_initialize_thread(kmp_info_t * th)1380 void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1381 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count);
1382 int new_value = __kmp_fork_count + 1;
1383 // Return if already initialized
1384 if (old_value == new_value)
1385 return;
1386 // Wait, then return if being initialized
1387 if (old_value == -1 || !__kmp_atomic_compare_store(
1388 &th->th.th_suspend_init_count, old_value, -1)) {
1389 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) {
1390 KMP_CPU_PAUSE();
1391 }
1392 } else {
1393 // Claim to be the initializer and do initializations
1394 int status;
1395 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1396 &__kmp_suspend_cond_attr);
1397 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1398 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1399 &__kmp_suspend_mutex_attr);
1400 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1401 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value);
1402 }
1403 }
1404
__kmp_suspend_uninitialize_thread(kmp_info_t * th)1405 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1406 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) {
1407 /* this means we have initialize the suspension pthread objects for this
1408 thread in this instance of the process */
1409 int status;
1410
1411 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1412 if (status != 0 && status != EBUSY) {
1413 KMP_SYSFAIL("pthread_cond_destroy", status);
1414 }
1415 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1416 if (status != 0 && status != EBUSY) {
1417 KMP_SYSFAIL("pthread_mutex_destroy", status);
1418 }
1419 --th->th.th_suspend_init_count;
1420 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) ==
1421 __kmp_fork_count);
1422 }
1423 }
1424
1425 // return true if lock obtained, false otherwise
__kmp_try_suspend_mx(kmp_info_t * th)1426 int __kmp_try_suspend_mx(kmp_info_t *th) {
1427 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1428 }
1429
__kmp_lock_suspend_mx(kmp_info_t * th)1430 void __kmp_lock_suspend_mx(kmp_info_t *th) {
1431 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1432 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1433 }
1434
__kmp_unlock_suspend_mx(kmp_info_t * th)1435 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1436 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1437 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1438 }
1439
1440 /* This routine puts the calling thread to sleep after setting the
1441 sleep bit for the indicated flag variable to true. */
1442 template <class C>
__kmp_suspend_template(int th_gtid,C * flag)1443 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1444 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1445 kmp_info_t *th = __kmp_threads[th_gtid];
1446 int status;
1447 typename C::flag_t old_spin;
1448
1449 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1450 flag->get()));
1451
1452 __kmp_suspend_initialize_thread(th);
1453
1454 __kmp_lock_suspend_mx(th);
1455
1456 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1457 th_gtid, flag->get()));
1458
1459 /* TODO: shouldn't this use release semantics to ensure that
1460 __kmp_suspend_initialize_thread gets called first? */
1461 old_spin = flag->set_sleeping();
1462 TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1463 th->th.th_sleep_loc_type = flag->get_type();
1464 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1465 __kmp_pause_status != kmp_soft_paused) {
1466 flag->unset_sleeping();
1467 TCW_PTR(th->th.th_sleep_loc, NULL);
1468 th->th.th_sleep_loc_type = flag_unset;
1469 __kmp_unlock_suspend_mx(th);
1470 return;
1471 }
1472 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1473 " was %x\n",
1474 th_gtid, flag->get(), flag->load(), old_spin));
1475
1476 if (flag->done_check_val(old_spin) || flag->done_check()) {
1477 flag->unset_sleeping();
1478 TCW_PTR(th->th.th_sleep_loc, NULL);
1479 th->th.th_sleep_loc_type = flag_unset;
1480 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1481 "for spin(%p)\n",
1482 th_gtid, flag->get()));
1483 } else {
1484 /* Encapsulate in a loop as the documentation states that this may
1485 "with low probability" return when the condition variable has
1486 not been signaled or broadcast */
1487 int deactivated = FALSE;
1488
1489 while (flag->is_sleeping()) {
1490 #ifdef DEBUG_SUSPEND
1491 char buffer[128];
1492 __kmp_suspend_count++;
1493 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1494 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1495 buffer);
1496 #endif
1497 // Mark the thread as no longer active (only in the first iteration of the
1498 // loop).
1499 if (!deactivated) {
1500 th->th.th_active = FALSE;
1501 if (th->th.th_active_in_pool) {
1502 th->th.th_active_in_pool = FALSE;
1503 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1504 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1505 }
1506 deactivated = TRUE;
1507 }
1508
1509 KMP_DEBUG_ASSERT(th->th.th_sleep_loc);
1510 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type);
1511
1512 #if USE_SUSPEND_TIMEOUT
1513 struct timespec now;
1514 struct timeval tval;
1515 int msecs;
1516
1517 status = gettimeofday(&tval, NULL);
1518 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1519 TIMEVAL_TO_TIMESPEC(&tval, &now);
1520
1521 msecs = (4 * __kmp_dflt_blocktime) + 200;
1522 now.tv_sec += msecs / 1000;
1523 now.tv_nsec += (msecs % 1000) * 1000;
1524
1525 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1526 "pthread_cond_timedwait\n",
1527 th_gtid));
1528 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1529 &th->th.th_suspend_mx.m_mutex, &now);
1530 #else
1531 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1532 " pthread_cond_wait\n",
1533 th_gtid));
1534 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1535 &th->th.th_suspend_mx.m_mutex);
1536 #endif // USE_SUSPEND_TIMEOUT
1537
1538 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1539 KMP_SYSFAIL("pthread_cond_wait", status);
1540 }
1541
1542 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type());
1543
1544 if (!flag->is_sleeping() &&
1545 ((status == EINTR) || (status == ETIMEDOUT))) {
1546 // if interrupt or timeout, and thread is no longer sleeping, we need to
1547 // make sure sleep_loc gets reset; however, this shouldn't be needed if
1548 // we woke up with resume
1549 flag->unset_sleeping();
1550 TCW_PTR(th->th.th_sleep_loc, NULL);
1551 th->th.th_sleep_loc_type = flag_unset;
1552 }
1553 #ifdef KMP_DEBUG
1554 if (status == ETIMEDOUT) {
1555 if (flag->is_sleeping()) {
1556 KF_TRACE(100,
1557 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1558 } else {
1559 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1560 "not set!\n",
1561 th_gtid));
1562 TCW_PTR(th->th.th_sleep_loc, NULL);
1563 th->th.th_sleep_loc_type = flag_unset;
1564 }
1565 } else if (flag->is_sleeping()) {
1566 KF_TRACE(100,
1567 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1568 }
1569 #endif
1570 } // while
1571
1572 // Mark the thread as active again (if it was previous marked as inactive)
1573 if (deactivated) {
1574 th->th.th_active = TRUE;
1575 if (TCR_4(th->th.th_in_pool)) {
1576 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1577 th->th.th_active_in_pool = TRUE;
1578 }
1579 }
1580 }
1581 // We may have had the loop variable set before entering the loop body;
1582 // so we need to reset sleep_loc.
1583 TCW_PTR(th->th.th_sleep_loc, NULL);
1584 th->th.th_sleep_loc_type = flag_unset;
1585
1586 KMP_DEBUG_ASSERT(!flag->is_sleeping());
1587 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc);
1588 #ifdef DEBUG_SUSPEND
1589 {
1590 char buffer[128];
1591 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1592 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1593 buffer);
1594 }
1595 #endif
1596
1597 __kmp_unlock_suspend_mx(th);
1598 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1599 }
1600
1601 template <bool C, bool S>
__kmp_suspend_32(int th_gtid,kmp_flag_32<C,S> * flag)1602 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) {
1603 __kmp_suspend_template(th_gtid, flag);
1604 }
1605 template <bool C, bool S>
__kmp_suspend_64(int th_gtid,kmp_flag_64<C,S> * flag)1606 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) {
1607 __kmp_suspend_template(th_gtid, flag);
1608 }
1609 template <bool C, bool S>
__kmp_atomic_suspend_64(int th_gtid,kmp_atomic_flag_64<C,S> * flag)1610 void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) {
1611 __kmp_suspend_template(th_gtid, flag);
1612 }
__kmp_suspend_oncore(int th_gtid,kmp_flag_oncore * flag)1613 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1614 __kmp_suspend_template(th_gtid, flag);
1615 }
1616
1617 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *);
1618 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *);
1619 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *);
1620 template void
1621 __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1622 template void
1623 __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *);
1624
1625 /* This routine signals the thread specified by target_gtid to wake up
1626 after setting the sleep bit indicated by the flag argument to FALSE.
1627 The target thread must already have called __kmp_suspend_template() */
1628 template <class C>
__kmp_resume_template(int target_gtid,C * flag)1629 static inline void __kmp_resume_template(int target_gtid, C *flag) {
1630 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1631 kmp_info_t *th = __kmp_threads[target_gtid];
1632 int status;
1633
1634 #ifdef KMP_DEBUG
1635 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1636 #endif
1637
1638 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1639 gtid, target_gtid));
1640 KMP_DEBUG_ASSERT(gtid != target_gtid);
1641
1642 __kmp_suspend_initialize_thread(th);
1643
1644 __kmp_lock_suspend_mx(th);
1645
1646 if (!flag || flag != th->th.th_sleep_loc) {
1647 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a
1648 // different location; wake up at new location
1649 flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1650 }
1651
1652 // First, check if the flag is null or its type has changed. If so, someone
1653 // else woke it up.
1654 if (!flag) { // Thread doesn't appear to be sleeping on anything
1655 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1656 "awake: flag(%p)\n",
1657 gtid, target_gtid, (void *)NULL));
1658 __kmp_unlock_suspend_mx(th);
1659 return;
1660 } else if (flag->get_type() != th->th.th_sleep_loc_type) {
1661 // Flag type does not appear to match this function template; possibly the
1662 // thread is sleeping on something else. Try null resume again.
1663 KF_TRACE(
1664 5,
1665 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), "
1666 "spin(%p) type=%d ptr_type=%d\n",
1667 gtid, target_gtid, flag, flag->get(), flag->get_type(),
1668 th->th.th_sleep_loc_type));
1669 __kmp_unlock_suspend_mx(th);
1670 __kmp_null_resume_wrapper(th);
1671 return;
1672 } else { // if multiple threads are sleeping, flag should be internally
1673 // referring to a specific thread here
1674 if (!flag->is_sleeping()) {
1675 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1676 "awake: flag(%p): %u\n",
1677 gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1678 __kmp_unlock_suspend_mx(th);
1679 return;
1680 }
1681 }
1682 KMP_DEBUG_ASSERT(flag);
1683 flag->unset_sleeping();
1684 TCW_PTR(th->th.th_sleep_loc, NULL);
1685 th->th.th_sleep_loc_type = flag_unset;
1686
1687 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1688 "sleep bit for flag's loc(%p): %u\n",
1689 gtid, target_gtid, flag->get(), (unsigned int)flag->load()));
1690
1691 #ifdef DEBUG_SUSPEND
1692 {
1693 char buffer[128];
1694 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1695 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1696 target_gtid, buffer);
1697 }
1698 #endif
1699 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1700 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1701 __kmp_unlock_suspend_mx(th);
1702 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1703 " for T#%d\n",
1704 gtid, target_gtid));
1705 }
1706
1707 template <bool C, bool S>
__kmp_resume_32(int target_gtid,kmp_flag_32<C,S> * flag)1708 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) {
1709 __kmp_resume_template(target_gtid, flag);
1710 }
1711 template <bool C, bool S>
__kmp_resume_64(int target_gtid,kmp_flag_64<C,S> * flag)1712 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) {
1713 __kmp_resume_template(target_gtid, flag);
1714 }
1715 template <bool C, bool S>
__kmp_atomic_resume_64(int target_gtid,kmp_atomic_flag_64<C,S> * flag)1716 void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) {
1717 __kmp_resume_template(target_gtid, flag);
1718 }
__kmp_resume_oncore(int target_gtid,kmp_flag_oncore * flag)1719 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1720 __kmp_resume_template(target_gtid, flag);
1721 }
1722
1723 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *);
1724 template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *);
1725 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *);
1726 template void
1727 __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *);
1728
1729 #if KMP_USE_MONITOR
__kmp_resume_monitor()1730 void __kmp_resume_monitor() {
1731 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1732 int status;
1733 #ifdef KMP_DEBUG
1734 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1735 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1736 KMP_GTID_MONITOR));
1737 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1738 #endif
1739 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1740 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1741 #ifdef DEBUG_SUSPEND
1742 {
1743 char buffer[128];
1744 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1745 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1746 KMP_GTID_MONITOR, buffer);
1747 }
1748 #endif
1749 status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1750 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1751 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1752 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1753 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1754 " for T#%d\n",
1755 gtid, KMP_GTID_MONITOR));
1756 }
1757 #endif // KMP_USE_MONITOR
1758
__kmp_yield()1759 void __kmp_yield() { sched_yield(); }
1760
__kmp_gtid_set_specific(int gtid)1761 void __kmp_gtid_set_specific(int gtid) {
1762 if (__kmp_init_gtid) {
1763 int status;
1764 status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1765 (void *)(intptr_t)(gtid + 1));
1766 KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1767 } else {
1768 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1769 }
1770 }
1771
__kmp_gtid_get_specific()1772 int __kmp_gtid_get_specific() {
1773 int gtid;
1774 if (!__kmp_init_gtid) {
1775 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1776 "KMP_GTID_SHUTDOWN\n"));
1777 return KMP_GTID_SHUTDOWN;
1778 }
1779 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1780 if (gtid == 0) {
1781 gtid = KMP_GTID_DNE;
1782 } else {
1783 gtid--;
1784 }
1785 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1786 __kmp_gtid_threadprivate_key, gtid));
1787 return gtid;
1788 }
1789
__kmp_read_cpu_time(void)1790 double __kmp_read_cpu_time(void) {
1791 /*clock_t t;*/
1792 struct tms buffer;
1793
1794 /*t =*/times(&buffer);
1795
1796 return (double)(buffer.tms_utime + buffer.tms_cutime) /
1797 (double)CLOCKS_PER_SEC;
1798 }
1799
__kmp_read_system_info(struct kmp_sys_info * info)1800 int __kmp_read_system_info(struct kmp_sys_info *info) {
1801 int status;
1802 struct rusage r_usage;
1803
1804 memset(info, 0, sizeof(*info));
1805
1806 status = getrusage(RUSAGE_SELF, &r_usage);
1807 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1808
1809 #if !KMP_OS_WASI
1810 // The maximum resident set size utilized (in kilobytes)
1811 info->maxrss = r_usage.ru_maxrss;
1812 // The number of page faults serviced without any I/O
1813 info->minflt = r_usage.ru_minflt;
1814 // The number of page faults serviced that required I/O
1815 info->majflt = r_usage.ru_majflt;
1816 // The number of times a process was "swapped" out of memory
1817 info->nswap = r_usage.ru_nswap;
1818 // The number of times the file system had to perform input
1819 info->inblock = r_usage.ru_inblock;
1820 // The number of times the file system had to perform output
1821 info->oublock = r_usage.ru_oublock;
1822 // The number of times a context switch was voluntarily
1823 info->nvcsw = r_usage.ru_nvcsw;
1824 // The number of times a context switch was forced
1825 info->nivcsw = r_usage.ru_nivcsw;
1826 #endif
1827
1828 return (status != 0);
1829 }
1830
__kmp_read_system_time(double * delta)1831 void __kmp_read_system_time(double *delta) {
1832 double t_ns;
1833 struct timeval tval;
1834 struct timespec stop;
1835 int status;
1836
1837 status = gettimeofday(&tval, NULL);
1838 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1839 TIMEVAL_TO_TIMESPEC(&tval, &stop);
1840 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start));
1841 *delta = (t_ns * 1e-9);
1842 }
1843
__kmp_clear_system_time(void)1844 void __kmp_clear_system_time(void) {
1845 struct timeval tval;
1846 int status;
1847 status = gettimeofday(&tval, NULL);
1848 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1849 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1850 }
1851
__kmp_get_xproc(void)1852 static int __kmp_get_xproc(void) {
1853
1854 int r = 0;
1855
1856 #if KMP_OS_LINUX
1857
1858 __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r));
1859
1860 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \
1861 KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_WASI || KMP_OS_AIX
1862
1863 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r));
1864
1865 #elif KMP_OS_DARWIN
1866
1867 // Bug C77011 High "OpenMP Threads and number of active cores".
1868
1869 // Find the number of available CPUs.
1870 kern_return_t rc;
1871 host_basic_info_data_t info;
1872 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1873 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1874 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1875 // Cannot use KA_TRACE() here because this code works before trace support
1876 // is initialized.
1877 r = info.avail_cpus;
1878 } else {
1879 KMP_WARNING(CantGetNumAvailCPU);
1880 KMP_INFORM(AssumedNumCPU);
1881 }
1882
1883 #else
1884
1885 #error "Unknown or unsupported OS."
1886
1887 #endif
1888
1889 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1890
1891 } // __kmp_get_xproc
1892
__kmp_read_from_file(char const * path,char const * format,...)1893 int __kmp_read_from_file(char const *path, char const *format, ...) {
1894 int result;
1895 va_list args;
1896
1897 va_start(args, format);
1898 FILE *f = fopen(path, "rb");
1899 if (f == NULL) {
1900 va_end(args);
1901 return 0;
1902 }
1903 result = vfscanf(f, format, args);
1904 fclose(f);
1905 va_end(args);
1906
1907 return result;
1908 }
1909
__kmp_runtime_initialize(void)1910 void __kmp_runtime_initialize(void) {
1911 int status;
1912 pthread_mutexattr_t mutex_attr;
1913 pthread_condattr_t cond_attr;
1914
1915 if (__kmp_init_runtime) {
1916 return;
1917 }
1918
1919 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1920 if (!__kmp_cpuinfo.initialized) {
1921 __kmp_query_cpuid(&__kmp_cpuinfo);
1922 }
1923 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1924
1925 __kmp_xproc = __kmp_get_xproc();
1926
1927 #if !KMP_32_BIT_ARCH
1928 struct rlimit rlim;
1929 // read stack size of calling thread, save it as default for worker threads;
1930 // this should be done before reading environment variables
1931 status = getrlimit(RLIMIT_STACK, &rlim);
1932 if (status == 0) { // success?
1933 __kmp_stksize = rlim.rlim_cur;
1934 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed
1935 }
1936 #endif /* KMP_32_BIT_ARCH */
1937
1938 if (sysconf(_SC_THREADS)) {
1939
1940 /* Query the maximum number of threads */
1941 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth));
1942 #ifdef __ve__
1943 if (__kmp_sys_max_nth == -1) {
1944 // VE's pthread supports only up to 64 threads per a VE process.
1945 // So we use that KMP_MAX_NTH (predefined as 64) here.
1946 __kmp_sys_max_nth = KMP_MAX_NTH;
1947 }
1948 #else
1949 if (__kmp_sys_max_nth == -1) {
1950 /* Unlimited threads for NPTL */
1951 __kmp_sys_max_nth = INT_MAX;
1952 } else if (__kmp_sys_max_nth <= 1) {
1953 /* Can't tell, just use PTHREAD_THREADS_MAX */
1954 __kmp_sys_max_nth = KMP_MAX_NTH;
1955 }
1956 #endif
1957
1958 /* Query the minimum stack size */
1959 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1960 if (__kmp_sys_min_stksize <= 1) {
1961 __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1962 }
1963 }
1964
1965 /* Set up minimum number of threads to switch to TLS gtid */
1966 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1967
1968 status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1969 __kmp_internal_end_dest);
1970 KMP_CHECK_SYSFAIL("pthread_key_create", status);
1971 status = pthread_mutexattr_init(&mutex_attr);
1972 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1973 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1974 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1975 status = pthread_mutexattr_destroy(&mutex_attr);
1976 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status);
1977 status = pthread_condattr_init(&cond_attr);
1978 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1979 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1980 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1981 status = pthread_condattr_destroy(&cond_attr);
1982 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status);
1983 #if USE_ITT_BUILD
1984 __kmp_itt_initialize();
1985 #endif /* USE_ITT_BUILD */
1986
1987 __kmp_init_runtime = TRUE;
1988 }
1989
__kmp_runtime_destroy(void)1990 void __kmp_runtime_destroy(void) {
1991 int status;
1992
1993 if (!__kmp_init_runtime) {
1994 return; // Nothing to do.
1995 }
1996
1997 #if USE_ITT_BUILD
1998 __kmp_itt_destroy();
1999 #endif /* USE_ITT_BUILD */
2000
2001 status = pthread_key_delete(__kmp_gtid_threadprivate_key);
2002 KMP_CHECK_SYSFAIL("pthread_key_delete", status);
2003
2004 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
2005 if (status != 0 && status != EBUSY) {
2006 KMP_SYSFAIL("pthread_mutex_destroy", status);
2007 }
2008 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
2009 if (status != 0 && status != EBUSY) {
2010 KMP_SYSFAIL("pthread_cond_destroy", status);
2011 }
2012 #if KMP_AFFINITY_SUPPORTED
2013 __kmp_affinity_uninitialize();
2014 #endif
2015
2016 __kmp_init_runtime = FALSE;
2017 }
2018
2019 /* Put the thread to sleep for a time period */
2020 /* NOTE: not currently used anywhere */
__kmp_thread_sleep(int millis)2021 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
2022
2023 /* Calculate the elapsed wall clock time for the user */
__kmp_elapsed(double * t)2024 void __kmp_elapsed(double *t) {
2025 int status;
2026 #ifdef FIX_SGI_CLOCK
2027 struct timespec ts;
2028
2029 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
2030 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
2031 *t =
2032 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
2033 #else
2034 struct timeval tv;
2035
2036 status = gettimeofday(&tv, NULL);
2037 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
2038 *t =
2039 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
2040 #endif
2041 }
2042
2043 /* Calculate the elapsed wall clock tick for the user */
__kmp_elapsed_tick(double * t)2044 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
2045
2046 /* Return the current time stamp in nsec */
__kmp_now_nsec()2047 kmp_uint64 __kmp_now_nsec() {
2048 struct timeval t;
2049 gettimeofday(&t, NULL);
2050 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
2051 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
2052 return nsec;
2053 }
2054
2055 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2056 /* Measure clock ticks per millisecond */
__kmp_initialize_system_tick()2057 void __kmp_initialize_system_tick() {
2058 kmp_uint64 now, nsec2, diff;
2059 kmp_uint64 delay = 1000000; // ~450 usec on most machines.
2060 kmp_uint64 nsec = __kmp_now_nsec();
2061 kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
2062 while ((now = __kmp_hardware_timestamp()) < goal)
2063 ;
2064 nsec2 = __kmp_now_nsec();
2065 diff = nsec2 - nsec;
2066 if (diff > 0) {
2067 double tpus = 1000.0 * (double)(delay + (now - goal)) / (double)diff;
2068 if (tpus > 0.0) {
2069 __kmp_ticks_per_msec = (kmp_uint64)(tpus * 1000.0);
2070 __kmp_ticks_per_usec = (kmp_uint64)tpus;
2071 }
2072 }
2073 }
2074 #endif
2075
2076 /* Determine whether the given address is mapped into the current address
2077 space. */
2078
__kmp_is_address_mapped(void * addr)2079 int __kmp_is_address_mapped(void *addr) {
2080
2081 int found = 0;
2082 int rc;
2083
2084 #if KMP_OS_LINUX || KMP_OS_HURD
2085
2086 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the
2087 address ranges mapped into the address space. */
2088
2089 char *name = __kmp_str_format("/proc/%d/maps", getpid());
2090 FILE *file = NULL;
2091
2092 file = fopen(name, "r");
2093 KMP_ASSERT(file != NULL);
2094
2095 for (;;) {
2096
2097 void *beginning = NULL;
2098 void *ending = NULL;
2099 char perms[5];
2100
2101 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
2102 if (rc == EOF) {
2103 break;
2104 }
2105 KMP_ASSERT(rc == 3 &&
2106 KMP_STRLEN(perms) == 4); // Make sure all fields are read.
2107
2108 // Ending address is not included in the region, but beginning is.
2109 if ((addr >= beginning) && (addr < ending)) {
2110 perms[2] = 0; // 3th and 4th character does not matter.
2111 if (strcmp(perms, "rw") == 0) {
2112 // Memory we are looking for should be readable and writable.
2113 found = 1;
2114 }
2115 break;
2116 }
2117 }
2118
2119 // Free resources.
2120 fclose(file);
2121 KMP_INTERNAL_FREE(name);
2122 #elif KMP_OS_FREEBSD
2123 char *buf;
2124 size_t lstsz;
2125 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()};
2126 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0);
2127 if (rc < 0)
2128 return 0;
2129 // We pass from number of vm entry's semantic
2130 // to size of whole entry map list.
2131 lstsz = lstsz * 4 / 3;
2132 buf = reinterpret_cast<char *>(KMP_INTERNAL_MALLOC(lstsz));
2133 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0);
2134 if (rc < 0) {
2135 KMP_INTERNAL_FREE(buf);
2136 return 0;
2137 }
2138
2139 char *lw = buf;
2140 char *up = buf + lstsz;
2141
2142 while (lw < up) {
2143 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw);
2144 size_t cursz = cur->kve_structsize;
2145 if (cursz == 0)
2146 break;
2147 void *start = reinterpret_cast<void *>(cur->kve_start);
2148 void *end = reinterpret_cast<void *>(cur->kve_end);
2149 // Readable/Writable addresses within current map entry
2150 if ((addr >= start) && (addr < end)) {
2151 if ((cur->kve_protection & KVME_PROT_READ) != 0 &&
2152 (cur->kve_protection & KVME_PROT_WRITE) != 0) {
2153 found = 1;
2154 break;
2155 }
2156 }
2157 lw += cursz;
2158 }
2159 KMP_INTERNAL_FREE(buf);
2160
2161 #elif KMP_OS_DARWIN
2162
2163 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2164 using vm interface. */
2165
2166 int buffer;
2167 vm_size_t count;
2168 rc = vm_read_overwrite(
2169 mach_task_self(), // Task to read memory of.
2170 (vm_address_t)(addr), // Address to read from.
2171 1, // Number of bytes to be read.
2172 (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2173 &count // Address of var to save number of read bytes in.
2174 );
2175 if (rc == 0) {
2176 // Memory successfully read.
2177 found = 1;
2178 }
2179
2180 #elif KMP_OS_NETBSD
2181
2182 int mib[5];
2183 mib[0] = CTL_VM;
2184 mib[1] = VM_PROC;
2185 mib[2] = VM_PROC_MAP;
2186 mib[3] = getpid();
2187 mib[4] = sizeof(struct kinfo_vmentry);
2188
2189 size_t size;
2190 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2191 KMP_ASSERT(!rc);
2192 KMP_ASSERT(size);
2193
2194 size = size * 4 / 3;
2195 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2196 KMP_ASSERT(kiv);
2197
2198 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2199 KMP_ASSERT(!rc);
2200 KMP_ASSERT(size);
2201
2202 for (size_t i = 0; i < size; i++) {
2203 if (kiv[i].kve_start >= (uint64_t)addr &&
2204 kiv[i].kve_end <= (uint64_t)addr) {
2205 found = 1;
2206 break;
2207 }
2208 }
2209 KMP_INTERNAL_FREE(kiv);
2210 #elif KMP_OS_OPENBSD
2211
2212 int mib[3];
2213 mib[0] = CTL_KERN;
2214 mib[1] = KERN_PROC_VMMAP;
2215 mib[2] = getpid();
2216
2217 size_t size;
2218 uint64_t end;
2219 rc = sysctl(mib, 3, NULL, &size, NULL, 0);
2220 KMP_ASSERT(!rc);
2221 KMP_ASSERT(size);
2222 end = size;
2223
2224 struct kinfo_vmentry kiv = {.kve_start = 0};
2225
2226 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) {
2227 KMP_ASSERT(size);
2228 if (kiv.kve_end == end)
2229 break;
2230
2231 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) {
2232 found = 1;
2233 break;
2234 }
2235 kiv.kve_start += 1;
2236 }
2237 #elif KMP_OS_WASI
2238 found = (int)addr < (__builtin_wasm_memory_size(0) * PAGESIZE);
2239 #elif KMP_OS_DRAGONFLY || KMP_OS_SOLARIS || KMP_OS_AIX
2240
2241 (void)rc;
2242 // FIXME(DragonFly, Solaris, AIX): Implement this
2243 found = 1;
2244
2245 #else
2246
2247 #error "Unknown or unsupported OS"
2248
2249 #endif
2250
2251 return found;
2252
2253 } // __kmp_is_address_mapped
2254
2255 #ifdef USE_LOAD_BALANCE
2256
2257 #if KMP_OS_DARWIN || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
2258 KMP_OS_OPENBSD || KMP_OS_SOLARIS
2259
2260 // The function returns the rounded value of the system load average
2261 // during given time interval which depends on the value of
2262 // __kmp_load_balance_interval variable (default is 60 sec, other values
2263 // may be 300 sec or 900 sec).
2264 // It returns -1 in case of error.
__kmp_get_load_balance(int max)2265 int __kmp_get_load_balance(int max) {
2266 double averages[3];
2267 int ret_avg = 0;
2268
2269 int res = getloadavg(averages, 3);
2270
2271 // Check __kmp_load_balance_interval to determine which of averages to use.
2272 // getloadavg() may return the number of samples less than requested that is
2273 // less than 3.
2274 if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2275 ret_avg = (int)averages[0]; // 1 min
2276 } else if ((__kmp_load_balance_interval >= 180 &&
2277 __kmp_load_balance_interval < 600) &&
2278 (res >= 2)) {
2279 ret_avg = (int)averages[1]; // 5 min
2280 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2281 ret_avg = (int)averages[2]; // 15 min
2282 } else { // Error occurred
2283 return -1;
2284 }
2285
2286 return ret_avg;
2287 }
2288
2289 #else // Linux* OS
2290
2291 // The function returns number of running (not sleeping) threads, or -1 in case
2292 // of error. Error could be reported if Linux* OS kernel too old (without
2293 // "/proc" support). Counting running threads stops if max running threads
2294 // encountered.
__kmp_get_load_balance(int max)2295 int __kmp_get_load_balance(int max) {
2296 static int permanent_error = 0;
2297 static int glb_running_threads = 0; // Saved count of the running threads for
2298 // the thread balance algorithm
2299 static double glb_call_time = 0; /* Thread balance algorithm call time */
2300
2301 int running_threads = 0; // Number of running threads in the system.
2302
2303 DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2304 struct dirent *proc_entry = NULL;
2305
2306 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2307 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2308 struct dirent *task_entry = NULL;
2309 int task_path_fixed_len;
2310
2311 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2312 int stat_file = -1;
2313 int stat_path_fixed_len;
2314
2315 #ifdef KMP_DEBUG
2316 int total_processes = 0; // Total number of processes in system.
2317 #endif
2318
2319 double call_time = 0.0;
2320
2321 __kmp_str_buf_init(&task_path);
2322 __kmp_str_buf_init(&stat_path);
2323
2324 __kmp_elapsed(&call_time);
2325
2326 if (glb_call_time &&
2327 (call_time - glb_call_time < __kmp_load_balance_interval)) {
2328 running_threads = glb_running_threads;
2329 goto finish;
2330 }
2331
2332 glb_call_time = call_time;
2333
2334 // Do not spend time on scanning "/proc/" if we have a permanent error.
2335 if (permanent_error) {
2336 running_threads = -1;
2337 goto finish;
2338 }
2339
2340 if (max <= 0) {
2341 max = INT_MAX;
2342 }
2343
2344 // Open "/proc/" directory.
2345 proc_dir = opendir("/proc");
2346 if (proc_dir == NULL) {
2347 // Cannot open "/proc/". Probably the kernel does not support it. Return an
2348 // error now and in subsequent calls.
2349 running_threads = -1;
2350 permanent_error = 1;
2351 goto finish;
2352 }
2353
2354 // Initialize fixed part of task_path. This part will not change.
2355 __kmp_str_buf_cat(&task_path, "/proc/", 6);
2356 task_path_fixed_len = task_path.used; // Remember number of used characters.
2357
2358 proc_entry = readdir(proc_dir);
2359 while (proc_entry != NULL) {
2360 #if KMP_OS_AIX
2361 // Proc entry name starts with a digit. Assume it is a process' directory.
2362 if (isdigit(proc_entry->d_name[0])) {
2363 #else
2364 // Proc entry is a directory and name starts with a digit. Assume it is a
2365 // process' directory.
2366 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2367 #endif
2368
2369 #ifdef KMP_DEBUG
2370 ++total_processes;
2371 #endif
2372 // Make sure init process is the very first in "/proc", so we can replace
2373 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2374 // 1. We are going to check that total_processes == 1 => d_name == "1" is
2375 // true (where "=>" is implication). Since C++ does not have => operator,
2376 // let us replace it with its equivalent: a => b == ! a || b.
2377 KMP_DEBUG_ASSERT(total_processes != 1 ||
2378 strcmp(proc_entry->d_name, "1") == 0);
2379
2380 // Construct task_path.
2381 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2382 __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2383 KMP_STRLEN(proc_entry->d_name));
2384 __kmp_str_buf_cat(&task_path, "/task", 5);
2385
2386 task_dir = opendir(task_path.str);
2387 if (task_dir == NULL) {
2388 // Process can finish between reading "/proc/" directory entry and
2389 // opening process' "task/" directory. So, in general case we should not
2390 // complain, but have to skip this process and read the next one. But on
2391 // systems with no "task/" support we will spend lot of time to scan
2392 // "/proc/" tree again and again without any benefit. "init" process
2393 // (its pid is 1) should exist always, so, if we cannot open
2394 // "/proc/1/task/" directory, it means "task/" is not supported by
2395 // kernel. Report an error now and in the future.
2396 if (strcmp(proc_entry->d_name, "1") == 0) {
2397 running_threads = -1;
2398 permanent_error = 1;
2399 goto finish;
2400 }
2401 } else {
2402 // Construct fixed part of stat file path.
2403 __kmp_str_buf_clear(&stat_path);
2404 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2405 __kmp_str_buf_cat(&stat_path, "/", 1);
2406 stat_path_fixed_len = stat_path.used;
2407
2408 task_entry = readdir(task_dir);
2409 while (task_entry != NULL) {
2410 // It is a directory and name starts with a digit.
2411 #if KMP_OS_AIX
2412 if (isdigit(task_entry->d_name[0])) {
2413 #else
2414 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2415 #endif
2416
2417 // Construct complete stat file path. Easiest way would be:
2418 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2419 // task_entry->d_name );
2420 // but seriae of __kmp_str_buf_cat works a bit faster.
2421 stat_path.used =
2422 stat_path_fixed_len; // Reset stat path to its fixed part.
2423 __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2424 KMP_STRLEN(task_entry->d_name));
2425 __kmp_str_buf_cat(&stat_path, "/stat", 5);
2426
2427 // Note: Low-level API (open/read/close) is used. High-level API
2428 // (fopen/fclose) works ~ 30 % slower.
2429 stat_file = open(stat_path.str, O_RDONLY);
2430 if (stat_file == -1) {
2431 // We cannot report an error because task (thread) can terminate
2432 // just before reading this file.
2433 } else {
2434 /* Content of "stat" file looks like:
2435 24285 (program) S ...
2436
2437 It is a single line (if program name does not include funny
2438 symbols). First number is a thread id, then name of executable
2439 file name in paretheses, then state of the thread. We need just
2440 thread state.
2441
2442 Good news: Length of program name is 15 characters max. Longer
2443 names are truncated.
2444
2445 Thus, we need rather short buffer: 15 chars for program name +
2446 2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2447
2448 Bad news: Program name may contain special symbols like space,
2449 closing parenthesis, or even new line. This makes parsing
2450 "stat" file not 100 % reliable. In case of fanny program names
2451 parsing may fail (report incorrect thread state).
2452
2453 Parsing "status" file looks more promissing (due to different
2454 file structure and escaping special symbols) but reading and
2455 parsing of "status" file works slower.
2456 -- ln
2457 */
2458 char buffer[65];
2459 ssize_t len;
2460 len = read(stat_file, buffer, sizeof(buffer) - 1);
2461 if (len >= 0) {
2462 buffer[len] = 0;
2463 // Using scanf:
2464 // sscanf( buffer, "%*d (%*s) %c ", & state );
2465 // looks very nice, but searching for a closing parenthesis
2466 // works a bit faster.
2467 char *close_parent = strstr(buffer, ") ");
2468 if (close_parent != NULL) {
2469 char state = *(close_parent + 2);
2470 if (state == 'R') {
2471 ++running_threads;
2472 if (running_threads >= max) {
2473 goto finish;
2474 }
2475 }
2476 }
2477 }
2478 close(stat_file);
2479 stat_file = -1;
2480 }
2481 }
2482 task_entry = readdir(task_dir);
2483 }
2484 closedir(task_dir);
2485 task_dir = NULL;
2486 }
2487 }
2488 proc_entry = readdir(proc_dir);
2489 }
2490
2491 // There _might_ be a timing hole where the thread executing this
2492 // code get skipped in the load balance, and running_threads is 0.
2493 // Assert in the debug builds only!!!
2494 KMP_DEBUG_ASSERT(running_threads > 0);
2495 if (running_threads <= 0) {
2496 running_threads = 1;
2497 }
2498
2499 finish: // Clean up and exit.
2500 if (proc_dir != NULL) {
2501 closedir(proc_dir);
2502 }
2503 __kmp_str_buf_free(&task_path);
2504 if (task_dir != NULL) {
2505 closedir(task_dir);
2506 }
2507 __kmp_str_buf_free(&stat_path);
2508 if (stat_file != -1) {
2509 close(stat_file);
2510 }
2511
2512 glb_running_threads = running_threads;
2513
2514 return running_threads;
2515
2516 } // __kmp_get_load_balance
2517
2518 #endif // KMP_OS_DARWIN
2519
2520 #endif // USE_LOAD_BALANCE
2521
2522 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \
2523 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \
2524 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \
2525 KMP_ARCH_ARM || KMP_ARCH_VE || KMP_ARCH_S390X || KMP_ARCH_PPC_XCOFF)
2526
2527 // we really only need the case with 1 argument, because CLANG always build
2528 // a struct of pointers to shared variables referenced in the outlined function
2529 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2530 void *p_argv[]
2531 #if OMPT_SUPPORT
2532 ,
2533 void **exit_frame_ptr
2534 #endif
2535 ) {
2536 #if OMPT_SUPPORT
2537 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2538 #endif
2539
2540 switch (argc) {
2541 default:
2542 fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2543 fflush(stderr);
2544 exit(-1);
2545 case 0:
2546 (*pkfn)(>id, &tid);
2547 break;
2548 case 1:
2549 (*pkfn)(>id, &tid, p_argv[0]);
2550 break;
2551 case 2:
2552 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]);
2553 break;
2554 case 3:
2555 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]);
2556 break;
2557 case 4:
2558 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2559 break;
2560 case 5:
2561 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2562 break;
2563 case 6:
2564 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2565 p_argv[5]);
2566 break;
2567 case 7:
2568 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2569 p_argv[5], p_argv[6]);
2570 break;
2571 case 8:
2572 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2573 p_argv[5], p_argv[6], p_argv[7]);
2574 break;
2575 case 9:
2576 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2577 p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2578 break;
2579 case 10:
2580 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2581 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2582 break;
2583 case 11:
2584 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2585 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2586 break;
2587 case 12:
2588 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2589 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2590 p_argv[11]);
2591 break;
2592 case 13:
2593 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2594 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2595 p_argv[11], p_argv[12]);
2596 break;
2597 case 14:
2598 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2599 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2600 p_argv[11], p_argv[12], p_argv[13]);
2601 break;
2602 case 15:
2603 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2604 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2605 p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2606 break;
2607 }
2608
2609 return 1;
2610 }
2611
2612 #endif
2613
2614 #if KMP_OS_LINUX
2615 // Functions for hidden helper task
2616 namespace {
2617 // Condition variable for initializing hidden helper team
2618 pthread_cond_t hidden_helper_threads_initz_cond_var;
2619 pthread_mutex_t hidden_helper_threads_initz_lock;
2620 volatile int hidden_helper_initz_signaled = FALSE;
2621
2622 // Condition variable for deinitializing hidden helper team
2623 pthread_cond_t hidden_helper_threads_deinitz_cond_var;
2624 pthread_mutex_t hidden_helper_threads_deinitz_lock;
2625 volatile int hidden_helper_deinitz_signaled = FALSE;
2626
2627 // Condition variable for the wrapper function of main thread
2628 pthread_cond_t hidden_helper_main_thread_cond_var;
2629 pthread_mutex_t hidden_helper_main_thread_lock;
2630 volatile int hidden_helper_main_thread_signaled = FALSE;
2631
2632 // Semaphore for worker threads. We don't use condition variable here in case
2633 // that when multiple signals are sent at the same time, only one thread might
2634 // be waken.
2635 sem_t hidden_helper_task_sem;
2636 } // namespace
2637
2638 void __kmp_hidden_helper_worker_thread_wait() {
2639 int status = sem_wait(&hidden_helper_task_sem);
2640 KMP_CHECK_SYSFAIL("sem_wait", status);
2641 }
2642
2643 void __kmp_do_initialize_hidden_helper_threads() {
2644 // Initialize condition variable
2645 int status =
2646 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr);
2647 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2648
2649 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr);
2650 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2651
2652 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr);
2653 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
2654
2655 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr);
2656 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2657
2658 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr);
2659 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2660
2661 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr);
2662 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
2663
2664 // Initialize the semaphore
2665 status = sem_init(&hidden_helper_task_sem, 0, 0);
2666 KMP_CHECK_SYSFAIL("sem_init", status);
2667
2668 // Create a new thread to finish initialization
2669 pthread_t handle;
2670 status = pthread_create(
2671 &handle, nullptr,
2672 [](void *) -> void * {
2673 __kmp_hidden_helper_threads_initz_routine();
2674 return nullptr;
2675 },
2676 nullptr);
2677 KMP_CHECK_SYSFAIL("pthread_create", status);
2678 }
2679
2680 void __kmp_hidden_helper_threads_initz_wait() {
2681 // Initial thread waits here for the completion of the initialization. The
2682 // condition variable will be notified by main thread of hidden helper teams.
2683 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2684 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2685
2686 if (!TCR_4(hidden_helper_initz_signaled)) {
2687 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var,
2688 &hidden_helper_threads_initz_lock);
2689 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2690 }
2691
2692 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2693 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2694 }
2695
2696 void __kmp_hidden_helper_initz_release() {
2697 // After all initialization, reset __kmp_init_hidden_helper_threads to false.
2698 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock);
2699 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2700
2701 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var);
2702 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2703
2704 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE);
2705
2706 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock);
2707 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2708 }
2709
2710 void __kmp_hidden_helper_main_thread_wait() {
2711 // The main thread of hidden helper team will be blocked here. The
2712 // condition variable can only be signal in the destructor of RTL.
2713 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2714 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2715
2716 if (!TCR_4(hidden_helper_main_thread_signaled)) {
2717 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var,
2718 &hidden_helper_main_thread_lock);
2719 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2720 }
2721
2722 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2723 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2724 }
2725
2726 void __kmp_hidden_helper_main_thread_release() {
2727 // The initial thread of OpenMP RTL should call this function to wake up the
2728 // main thread of hidden helper team.
2729 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock);
2730 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2731
2732 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var);
2733 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
2734
2735 // The hidden helper team is done here
2736 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE);
2737
2738 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock);
2739 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2740 }
2741
2742 void __kmp_hidden_helper_worker_thread_signal() {
2743 int status = sem_post(&hidden_helper_task_sem);
2744 KMP_CHECK_SYSFAIL("sem_post", status);
2745 }
2746
2747 void __kmp_hidden_helper_threads_deinitz_wait() {
2748 // Initial thread waits here for the completion of the deinitialization. The
2749 // condition variable will be notified by main thread of hidden helper teams.
2750 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
2751 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2752
2753 if (!TCR_4(hidden_helper_deinitz_signaled)) {
2754 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var,
2755 &hidden_helper_threads_deinitz_lock);
2756 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2757 }
2758
2759 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
2760 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2761 }
2762
2763 void __kmp_hidden_helper_threads_deinitz_release() {
2764 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock);
2765 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
2766
2767 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var);
2768 KMP_CHECK_SYSFAIL("pthread_cond_wait", status);
2769
2770 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE);
2771
2772 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock);
2773 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
2774 }
2775 #else // KMP_OS_LINUX
2776 void __kmp_hidden_helper_worker_thread_wait() {
2777 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2778 }
2779
2780 void __kmp_do_initialize_hidden_helper_threads() {
2781 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2782 }
2783
2784 void __kmp_hidden_helper_threads_initz_wait() {
2785 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2786 }
2787
2788 void __kmp_hidden_helper_initz_release() {
2789 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2790 }
2791
2792 void __kmp_hidden_helper_main_thread_wait() {
2793 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2794 }
2795
2796 void __kmp_hidden_helper_main_thread_release() {
2797 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2798 }
2799
2800 void __kmp_hidden_helper_worker_thread_signal() {
2801 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2802 }
2803
2804 void __kmp_hidden_helper_threads_deinitz_wait() {
2805 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2806 }
2807
2808 void __kmp_hidden_helper_threads_deinitz_release() {
2809 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS");
2810 }
2811 #endif // KMP_OS_LINUX
2812
2813 bool __kmp_detect_shm() {
2814 DIR *dir = opendir("/dev/shm");
2815 if (dir) { // /dev/shm exists
2816 closedir(dir);
2817 return true;
2818 } else if (ENOENT == errno) { // /dev/shm does not exist
2819 return false;
2820 } else { // opendir() failed
2821 return false;
2822 }
2823 }
2824
2825 bool __kmp_detect_tmp() {
2826 DIR *dir = opendir("/tmp");
2827 if (dir) { // /tmp exists
2828 closedir(dir);
2829 return true;
2830 } else if (ENOENT == errno) { // /tmp does not exist
2831 return false;
2832 } else { // opendir() failed
2833 return false;
2834 }
2835 }
2836
2837 // end of file //
2838