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 #include <sys/resource.h>
29 #include <sys/syscall.h>
30 #include <sys/time.h>
31 #include <sys/times.h>
32 #include <unistd.h>
33
34 #if KMP_OS_LINUX && !KMP_OS_CNK
35 #include <sys/sysinfo.h>
36 #if KMP_USE_FUTEX
37 // We should really include <futex.h>, but that causes compatibility problems on
38 // different Linux* OS distributions that either require that you include (or
39 // break when you try to include) <pci/types.h>. Since all we need is the two
40 // macros below (which are part of the kernel ABI, so can't change) we just
41 // define the constants here and don't include <futex.h>
42 #ifndef FUTEX_WAIT
43 #define FUTEX_WAIT 0
44 #endif
45 #ifndef FUTEX_WAKE
46 #define FUTEX_WAKE 1
47 #endif
48 #endif
49 #elif KMP_OS_DARWIN
50 #include <mach/mach.h>
51 #include <sys/sysctl.h>
52 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD
53 #include <pthread_np.h>
54 #elif KMP_OS_NETBSD
55 #include <sys/types.h>
56 #include <sys/sysctl.h>
57 #endif
58
59 #include <ctype.h>
60 #include <dirent.h>
61 #include <fcntl.h>
62
63 #include "tsan_annotations.h"
64
65 struct kmp_sys_timer {
66 struct timespec start;
67 };
68
69 // Convert timespec to nanoseconds.
70 #define TS2NS(timespec) (((timespec).tv_sec * 1e9) + (timespec).tv_nsec)
71
72 static struct kmp_sys_timer __kmp_sys_timer_data;
73
74 #if KMP_HANDLE_SIGNALS
75 typedef void (*sig_func_t)(int);
76 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG];
77 static sigset_t __kmp_sigset;
78 #endif
79
80 static int __kmp_init_runtime = FALSE;
81
82 static int __kmp_fork_count = 0;
83
84 static pthread_condattr_t __kmp_suspend_cond_attr;
85 static pthread_mutexattr_t __kmp_suspend_mutex_attr;
86
87 static kmp_cond_align_t __kmp_wait_cv;
88 static kmp_mutex_align_t __kmp_wait_mx;
89
90 kmp_uint64 __kmp_ticks_per_msec = 1000000;
91
92 #ifdef DEBUG_SUSPEND
__kmp_print_cond(char * buffer,kmp_cond_align_t * cond)93 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
94 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
95 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
96 cond->c_cond.__c_waiting);
97 }
98 #endif
99
100 #if (KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED)
101
102 /* Affinity support */
103
__kmp_affinity_bind_thread(int which)104 void __kmp_affinity_bind_thread(int which) {
105 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
106 "Illegal set affinity operation when not capable");
107
108 kmp_affin_mask_t *mask;
109 KMP_CPU_ALLOC_ON_STACK(mask);
110 KMP_CPU_ZERO(mask);
111 KMP_CPU_SET(which, mask);
112 __kmp_set_system_affinity(mask, TRUE);
113 KMP_CPU_FREE_FROM_STACK(mask);
114 }
115
116 /* Determine if we can access affinity functionality on this version of
117 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
118 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
__kmp_affinity_determine_capable(const char * env_var)119 void __kmp_affinity_determine_capable(const char *env_var) {
120 // Check and see if the OS supports thread affinity.
121
122 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
123
124 int gCode;
125 int sCode;
126 unsigned char *buf;
127 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
128
129 // If Linux* OS:
130 // If the syscall fails or returns a suggestion for the size,
131 // then we don't have to search for an appropriate size.
132 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_SIZE_LIMIT, buf);
133 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
134 "initial getaffinity call returned %d errno = %d\n",
135 gCode, errno));
136
137 // if ((gCode < 0) && (errno == ENOSYS))
138 if (gCode < 0) {
139 // System call not supported
140 if (__kmp_affinity_verbose ||
141 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
142 (__kmp_affinity_type != affinity_default) &&
143 (__kmp_affinity_type != affinity_disabled))) {
144 int error = errno;
145 kmp_msg_t err_code = KMP_ERR(error);
146 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
147 err_code, __kmp_msg_null);
148 if (__kmp_generate_warnings == kmp_warnings_off) {
149 __kmp_str_free(&err_code.str);
150 }
151 }
152 KMP_AFFINITY_DISABLE();
153 KMP_INTERNAL_FREE(buf);
154 return;
155 }
156 if (gCode > 0) { // Linux* OS only
157 // The optimal situation: the OS returns the size of the buffer it expects.
158 //
159 // A verification of correct behavior is that Isetaffinity on a NULL
160 // buffer with the same size fails with errno set to EFAULT.
161 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
162 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
163 "setaffinity for mask size %d returned %d errno = %d\n",
164 gCode, sCode, errno));
165 if (sCode < 0) {
166 if (errno == ENOSYS) {
167 if (__kmp_affinity_verbose ||
168 (__kmp_affinity_warnings &&
169 (__kmp_affinity_type != affinity_none) &&
170 (__kmp_affinity_type != affinity_default) &&
171 (__kmp_affinity_type != affinity_disabled))) {
172 int error = errno;
173 kmp_msg_t err_code = KMP_ERR(error);
174 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
175 err_code, __kmp_msg_null);
176 if (__kmp_generate_warnings == kmp_warnings_off) {
177 __kmp_str_free(&err_code.str);
178 }
179 }
180 KMP_AFFINITY_DISABLE();
181 KMP_INTERNAL_FREE(buf);
182 }
183 if (errno == EFAULT) {
184 KMP_AFFINITY_ENABLE(gCode);
185 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
186 "affinity supported (mask size %d)\n",
187 (int)__kmp_affin_mask_size));
188 KMP_INTERNAL_FREE(buf);
189 return;
190 }
191 }
192 }
193
194 // Call the getaffinity system call repeatedly with increasing set sizes
195 // until we succeed, or reach an upper bound on the search.
196 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
197 "searching for proper set size\n"));
198 int size;
199 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
200 gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
201 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
202 "getaffinity for mask size %d returned %d errno = %d\n",
203 size, gCode, errno));
204
205 if (gCode < 0) {
206 if (errno == ENOSYS) {
207 // We shouldn't get here
208 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
209 "inconsistent OS call behavior: errno == ENOSYS for mask "
210 "size %d\n",
211 size));
212 if (__kmp_affinity_verbose ||
213 (__kmp_affinity_warnings &&
214 (__kmp_affinity_type != affinity_none) &&
215 (__kmp_affinity_type != affinity_default) &&
216 (__kmp_affinity_type != affinity_disabled))) {
217 int error = errno;
218 kmp_msg_t err_code = KMP_ERR(error);
219 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
220 err_code, __kmp_msg_null);
221 if (__kmp_generate_warnings == kmp_warnings_off) {
222 __kmp_str_free(&err_code.str);
223 }
224 }
225 KMP_AFFINITY_DISABLE();
226 KMP_INTERNAL_FREE(buf);
227 return;
228 }
229 continue;
230 }
231
232 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
233 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
234 "setaffinity for mask size %d returned %d errno = %d\n",
235 gCode, sCode, errno));
236 if (sCode < 0) {
237 if (errno == ENOSYS) { // Linux* OS only
238 // We shouldn't get here
239 KA_TRACE(30, ("__kmp_affinity_determine_capable: "
240 "inconsistent OS call behavior: errno == ENOSYS for mask "
241 "size %d\n",
242 size));
243 if (__kmp_affinity_verbose ||
244 (__kmp_affinity_warnings &&
245 (__kmp_affinity_type != affinity_none) &&
246 (__kmp_affinity_type != affinity_default) &&
247 (__kmp_affinity_type != affinity_disabled))) {
248 int error = errno;
249 kmp_msg_t err_code = KMP_ERR(error);
250 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
251 err_code, __kmp_msg_null);
252 if (__kmp_generate_warnings == kmp_warnings_off) {
253 __kmp_str_free(&err_code.str);
254 }
255 }
256 KMP_AFFINITY_DISABLE();
257 KMP_INTERNAL_FREE(buf);
258 return;
259 }
260 if (errno == EFAULT) {
261 KMP_AFFINITY_ENABLE(gCode);
262 KA_TRACE(10, ("__kmp_affinity_determine_capable: "
263 "affinity supported (mask size %d)\n",
264 (int)__kmp_affin_mask_size));
265 KMP_INTERNAL_FREE(buf);
266 return;
267 }
268 }
269 }
270 // save uncaught error code
271 // int error = errno;
272 KMP_INTERNAL_FREE(buf);
273 // restore uncaught error code, will be printed at the next KMP_WARNING below
274 // errno = error;
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 (__kmp_affinity_verbose ||
281 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
282 (__kmp_affinity_type != affinity_default) &&
283 (__kmp_affinity_type != affinity_disabled))) {
284 KMP_WARNING(AffCantGetMaskSize, env_var);
285 }
286 }
287
288 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
289
290 #if KMP_USE_FUTEX
291
__kmp_futex_determine_capable()292 int __kmp_futex_determine_capable() {
293 int loc = 0;
294 int rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
295 int retval = (rc == 0) || (errno != ENOSYS);
296
297 KA_TRACE(10,
298 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
299 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
300 retval ? "" : " not"));
301
302 return retval;
303 }
304
305 #endif // KMP_USE_FUTEX
306
307 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS)
308 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
309 use compare_and_store for these routines */
310
__kmp_test_then_or8(volatile kmp_int8 * p,kmp_int8 d)311 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
312 kmp_int8 old_value, new_value;
313
314 old_value = TCR_1(*p);
315 new_value = old_value | d;
316
317 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
318 KMP_CPU_PAUSE();
319 old_value = TCR_1(*p);
320 new_value = old_value | d;
321 }
322 return old_value;
323 }
324
__kmp_test_then_and8(volatile kmp_int8 * p,kmp_int8 d)325 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
326 kmp_int8 old_value, new_value;
327
328 old_value = TCR_1(*p);
329 new_value = old_value & d;
330
331 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
332 KMP_CPU_PAUSE();
333 old_value = TCR_1(*p);
334 new_value = old_value & d;
335 }
336 return old_value;
337 }
338
__kmp_test_then_or32(volatile kmp_uint32 * p,kmp_uint32 d)339 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
340 kmp_uint32 old_value, new_value;
341
342 old_value = TCR_4(*p);
343 new_value = old_value | d;
344
345 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
346 KMP_CPU_PAUSE();
347 old_value = TCR_4(*p);
348 new_value = old_value | d;
349 }
350 return old_value;
351 }
352
__kmp_test_then_and32(volatile kmp_uint32 * p,kmp_uint32 d)353 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
354 kmp_uint32 old_value, new_value;
355
356 old_value = TCR_4(*p);
357 new_value = old_value & d;
358
359 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
360 KMP_CPU_PAUSE();
361 old_value = TCR_4(*p);
362 new_value = old_value & d;
363 }
364 return old_value;
365 }
366
367 #if KMP_ARCH_X86
__kmp_test_then_add8(volatile kmp_int8 * p,kmp_int8 d)368 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
369 kmp_int8 old_value, new_value;
370
371 old_value = TCR_1(*p);
372 new_value = old_value + d;
373
374 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
375 KMP_CPU_PAUSE();
376 old_value = TCR_1(*p);
377 new_value = old_value + d;
378 }
379 return old_value;
380 }
381
__kmp_test_then_add64(volatile kmp_int64 * p,kmp_int64 d)382 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
383 kmp_int64 old_value, new_value;
384
385 old_value = TCR_8(*p);
386 new_value = old_value + d;
387
388 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
389 KMP_CPU_PAUSE();
390 old_value = TCR_8(*p);
391 new_value = old_value + d;
392 }
393 return old_value;
394 }
395 #endif /* KMP_ARCH_X86 */
396
__kmp_test_then_or64(volatile kmp_uint64 * p,kmp_uint64 d)397 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
398 kmp_uint64 old_value, new_value;
399
400 old_value = TCR_8(*p);
401 new_value = old_value | d;
402 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
403 KMP_CPU_PAUSE();
404 old_value = TCR_8(*p);
405 new_value = old_value | d;
406 }
407 return old_value;
408 }
409
__kmp_test_then_and64(volatile kmp_uint64 * p,kmp_uint64 d)410 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
411 kmp_uint64 old_value, new_value;
412
413 old_value = TCR_8(*p);
414 new_value = old_value & d;
415 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
416 KMP_CPU_PAUSE();
417 old_value = TCR_8(*p);
418 new_value = old_value & d;
419 }
420 return old_value;
421 }
422
423 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
424
__kmp_terminate_thread(int gtid)425 void __kmp_terminate_thread(int gtid) {
426 int status;
427 kmp_info_t *th = __kmp_threads[gtid];
428
429 if (!th)
430 return;
431
432 #ifdef KMP_CANCEL_THREADS
433 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
434 status = pthread_cancel(th->th.th_info.ds.ds_thread);
435 if (status != 0 && status != ESRCH) {
436 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
437 __kmp_msg_null);
438 }
439 #endif
440 KMP_YIELD(TRUE);
441 } //
442
443 /* Set thread stack info according to values returned by pthread_getattr_np().
444 If values are unreasonable, assume call failed and use incremental stack
445 refinement method instead. Returns TRUE if the stack parameters could be
446 determined exactly, FALSE if incremental refinement is necessary. */
__kmp_set_stack_info(int gtid,kmp_info_t * th)447 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
448 int stack_data;
449 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
450 KMP_OS_HURD
451 pthread_attr_t attr;
452 int status;
453 size_t size = 0;
454 void *addr = 0;
455
456 /* Always do incremental stack refinement for ubermaster threads since the
457 initial thread stack range can be reduced by sibling thread creation so
458 pthread_attr_getstack may cause thread gtid aliasing */
459 if (!KMP_UBER_GTID(gtid)) {
460
461 /* Fetch the real thread attributes */
462 status = pthread_attr_init(&attr);
463 KMP_CHECK_SYSFAIL("pthread_attr_init", status);
464 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
465 status = pthread_attr_get_np(pthread_self(), &attr);
466 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
467 #else
468 status = pthread_getattr_np(pthread_self(), &attr);
469 KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
470 #endif
471 status = pthread_attr_getstack(&attr, &addr, &size);
472 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
473 KA_TRACE(60,
474 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
475 " %lu, low addr: %p\n",
476 gtid, size, addr));
477 status = pthread_attr_destroy(&attr);
478 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
479 }
480
481 if (size != 0 && addr != 0) { // was stack parameter determination successful?
482 /* Store the correct base and size */
483 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
484 TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
485 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
486 return TRUE;
487 }
488 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
489 KMP_OS_HURD */
490 /* Use incremental refinement starting from initial conservative estimate */
491 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
492 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
493 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
494 return FALSE;
495 }
496
__kmp_launch_worker(void * thr)497 static void *__kmp_launch_worker(void *thr) {
498 int status, old_type, old_state;
499 #ifdef KMP_BLOCK_SIGNALS
500 sigset_t new_set, old_set;
501 #endif /* KMP_BLOCK_SIGNALS */
502 void *exit_val;
503 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
504 KMP_OS_OPENBSD || KMP_OS_HURD
505 void *volatile padding = 0;
506 #endif
507 int gtid;
508
509 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
510 __kmp_gtid_set_specific(gtid);
511 #ifdef KMP_TDATA_GTID
512 __kmp_gtid = gtid;
513 #endif
514 #if KMP_STATS_ENABLED
515 // set thread local index to point to thread-specific stats
516 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
517 __kmp_stats_thread_ptr->startLife();
518 KMP_SET_THREAD_STATE(IDLE);
519 KMP_INIT_PARTITIONED_TIMERS(OMP_idle);
520 #endif
521
522 #if USE_ITT_BUILD
523 __kmp_itt_thread_name(gtid);
524 #endif /* USE_ITT_BUILD */
525
526 #if KMP_AFFINITY_SUPPORTED
527 __kmp_affinity_set_init_mask(gtid, FALSE);
528 #endif
529
530 #ifdef KMP_CANCEL_THREADS
531 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
532 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
533 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
534 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
535 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
536 #endif
537
538 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
539 // Set FP control regs to be a copy of the parallel initialization thread's.
540 __kmp_clear_x87_fpu_status_word();
541 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
542 __kmp_load_mxcsr(&__kmp_init_mxcsr);
543 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
544
545 #ifdef KMP_BLOCK_SIGNALS
546 status = sigfillset(&new_set);
547 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
548 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
549 KMP_CHECK_SYSFAIL("pthread_sigmask", status);
550 #endif /* KMP_BLOCK_SIGNALS */
551
552 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
553 KMP_OS_OPENBSD
554 if (__kmp_stkoffset > 0 && gtid > 0) {
555 padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
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 pthread_exit((void *)(intptr_t)exit_status);
1015 } // __kmp_exit_thread
1016
1017 #if KMP_USE_MONITOR
1018 void __kmp_resume_monitor();
1019
__kmp_reap_monitor(kmp_info_t * th)1020 void __kmp_reap_monitor(kmp_info_t *th) {
1021 int status;
1022 void *exit_val;
1023
1024 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
1025 " %#.8lx\n",
1026 th->th.th_info.ds.ds_thread));
1027
1028 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1029 // If both tid and gtid are 0, it means the monitor did not ever start.
1030 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1031 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1032 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1033 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1034 return;
1035 }
1036
1037 KMP_MB(); /* Flush all pending memory write invalidates. */
1038
1039 /* First, check to see whether the monitor thread exists to wake it up. This
1040 is to avoid performance problem when the monitor sleeps during
1041 blocktime-size interval */
1042
1043 status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1044 if (status != ESRCH) {
1045 __kmp_resume_monitor(); // Wake up the monitor thread
1046 }
1047 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1048 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1049 if (exit_val != th) {
1050 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1051 }
1052
1053 th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1054 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1055
1056 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1057 " %#.8lx\n",
1058 th->th.th_info.ds.ds_thread));
1059
1060 KMP_MB(); /* Flush all pending memory write invalidates. */
1061 }
1062 #endif // KMP_USE_MONITOR
1063
__kmp_reap_worker(kmp_info_t * th)1064 void __kmp_reap_worker(kmp_info_t *th) {
1065 int status;
1066 void *exit_val;
1067
1068 KMP_MB(); /* Flush all pending memory write invalidates. */
1069
1070 KA_TRACE(
1071 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1072
1073 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1074 #ifdef KMP_DEBUG
1075 /* Don't expose these to the user until we understand when they trigger */
1076 if (status != 0) {
1077 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1078 }
1079 if (exit_val != th) {
1080 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1081 "exit_val = %p\n",
1082 th->th.th_info.ds.ds_gtid, exit_val));
1083 }
1084 #endif /* KMP_DEBUG */
1085
1086 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1087 th->th.th_info.ds.ds_gtid));
1088
1089 KMP_MB(); /* Flush all pending memory write invalidates. */
1090 }
1091
1092 #if KMP_HANDLE_SIGNALS
1093
__kmp_null_handler(int signo)1094 static void __kmp_null_handler(int signo) {
1095 // Do nothing, for doing SIG_IGN-type actions.
1096 } // __kmp_null_handler
1097
__kmp_team_handler(int signo)1098 static void __kmp_team_handler(int signo) {
1099 if (__kmp_global.g.g_abort == 0) {
1100 /* Stage 1 signal handler, let's shut down all of the threads */
1101 #ifdef KMP_DEBUG
1102 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1103 #endif
1104 switch (signo) {
1105 case SIGHUP:
1106 case SIGINT:
1107 case SIGQUIT:
1108 case SIGILL:
1109 case SIGABRT:
1110 case SIGFPE:
1111 case SIGBUS:
1112 case SIGSEGV:
1113 #ifdef SIGSYS
1114 case SIGSYS:
1115 #endif
1116 case SIGTERM:
1117 if (__kmp_debug_buf) {
1118 __kmp_dump_debug_buffer();
1119 }
1120 KMP_MB(); // Flush all pending memory write invalidates.
1121 TCW_4(__kmp_global.g.g_abort, signo);
1122 KMP_MB(); // Flush all pending memory write invalidates.
1123 TCW_4(__kmp_global.g.g_done, TRUE);
1124 KMP_MB(); // Flush all pending memory write invalidates.
1125 break;
1126 default:
1127 #ifdef KMP_DEBUG
1128 __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1129 #endif
1130 break;
1131 }
1132 }
1133 } // __kmp_team_handler
1134
__kmp_sigaction(int signum,const struct sigaction * act,struct sigaction * oldact)1135 static void __kmp_sigaction(int signum, const struct sigaction *act,
1136 struct sigaction *oldact) {
1137 int rc = sigaction(signum, act, oldact);
1138 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1139 }
1140
__kmp_install_one_handler(int sig,sig_func_t handler_func,int parallel_init)1141 static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1142 int parallel_init) {
1143 KMP_MB(); // Flush all pending memory write invalidates.
1144 KB_TRACE(60,
1145 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1146 if (parallel_init) {
1147 struct sigaction new_action;
1148 struct sigaction old_action;
1149 new_action.sa_handler = handler_func;
1150 new_action.sa_flags = 0;
1151 sigfillset(&new_action.sa_mask);
1152 __kmp_sigaction(sig, &new_action, &old_action);
1153 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1154 sigaddset(&__kmp_sigset, sig);
1155 } else {
1156 // Restore/keep user's handler if one previously installed.
1157 __kmp_sigaction(sig, &old_action, NULL);
1158 }
1159 } else {
1160 // Save initial/system signal handlers to see if user handlers installed.
1161 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1162 }
1163 KMP_MB(); // Flush all pending memory write invalidates.
1164 } // __kmp_install_one_handler
1165
__kmp_remove_one_handler(int sig)1166 static void __kmp_remove_one_handler(int sig) {
1167 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1168 if (sigismember(&__kmp_sigset, sig)) {
1169 struct sigaction old;
1170 KMP_MB(); // Flush all pending memory write invalidates.
1171 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1172 if ((old.sa_handler != __kmp_team_handler) &&
1173 (old.sa_handler != __kmp_null_handler)) {
1174 // Restore the users signal handler.
1175 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1176 "restoring: sig=%d\n",
1177 sig));
1178 __kmp_sigaction(sig, &old, NULL);
1179 }
1180 sigdelset(&__kmp_sigset, sig);
1181 KMP_MB(); // Flush all pending memory write invalidates.
1182 }
1183 } // __kmp_remove_one_handler
1184
__kmp_install_signals(int parallel_init)1185 void __kmp_install_signals(int parallel_init) {
1186 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1187 if (__kmp_handle_signals || !parallel_init) {
1188 // If ! parallel_init, we do not install handlers, just save original
1189 // handlers. Let us do it even __handle_signals is 0.
1190 sigemptyset(&__kmp_sigset);
1191 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1192 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1193 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1194 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1195 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1196 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1197 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1198 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1199 #ifdef SIGSYS
1200 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1201 #endif // SIGSYS
1202 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1203 #ifdef SIGPIPE
1204 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1205 #endif // SIGPIPE
1206 }
1207 } // __kmp_install_signals
1208
__kmp_remove_signals(void)1209 void __kmp_remove_signals(void) {
1210 int sig;
1211 KB_TRACE(10, ("__kmp_remove_signals()\n"));
1212 for (sig = 1; sig < NSIG; ++sig) {
1213 __kmp_remove_one_handler(sig);
1214 }
1215 } // __kmp_remove_signals
1216
1217 #endif // KMP_HANDLE_SIGNALS
1218
__kmp_enable(int new_state)1219 void __kmp_enable(int new_state) {
1220 #ifdef KMP_CANCEL_THREADS
1221 int status, old_state;
1222 status = pthread_setcancelstate(new_state, &old_state);
1223 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1224 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1225 #endif
1226 }
1227
__kmp_disable(int * old_state)1228 void __kmp_disable(int *old_state) {
1229 #ifdef KMP_CANCEL_THREADS
1230 int status;
1231 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1232 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1233 #endif
1234 }
1235
__kmp_atfork_prepare(void)1236 static void __kmp_atfork_prepare(void) {
1237 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1238 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1239 }
1240
__kmp_atfork_parent(void)1241 static void __kmp_atfork_parent(void) {
1242 __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1243 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1244 }
1245
1246 /* Reset the library so execution in the child starts "all over again" with
1247 clean data structures in initial states. Don't worry about freeing memory
1248 allocated by parent, just abandon it to be safe. */
__kmp_atfork_child(void)1249 static void __kmp_atfork_child(void) {
1250 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1251 /* TODO make sure this is done right for nested/sibling */
1252 // ATT: Memory leaks are here? TODO: Check it and fix.
1253 /* KMP_ASSERT( 0 ); */
1254
1255 ++__kmp_fork_count;
1256
1257 #if KMP_AFFINITY_SUPPORTED
1258 #if KMP_OS_LINUX
1259 // reset the affinity in the child to the initial thread
1260 // affinity in the parent
1261 kmp_set_thread_affinity_mask_initial();
1262 #endif
1263 // Set default not to bind threads tightly in the child (we’re expecting
1264 // over-subscription after the fork and this can improve things for
1265 // scripting languages that use OpenMP inside process-parallel code).
1266 __kmp_affinity_type = affinity_none;
1267 if (__kmp_nested_proc_bind.bind_types != NULL) {
1268 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1269 }
1270 #endif // KMP_AFFINITY_SUPPORTED
1271
1272 __kmp_init_runtime = FALSE;
1273 #if KMP_USE_MONITOR
1274 __kmp_init_monitor = 0;
1275 #endif
1276 __kmp_init_parallel = FALSE;
1277 __kmp_init_middle = FALSE;
1278 __kmp_init_serial = FALSE;
1279 TCW_4(__kmp_init_gtid, FALSE);
1280 __kmp_init_common = FALSE;
1281
1282 TCW_4(__kmp_init_user_locks, FALSE);
1283 #if !KMP_USE_DYNAMIC_LOCK
1284 __kmp_user_lock_table.used = 1;
1285 __kmp_user_lock_table.allocated = 0;
1286 __kmp_user_lock_table.table = NULL;
1287 __kmp_lock_blocks = NULL;
1288 #endif
1289
1290 __kmp_all_nth = 0;
1291 TCW_4(__kmp_nth, 0);
1292
1293 __kmp_thread_pool = NULL;
1294 __kmp_thread_pool_insert_pt = NULL;
1295 __kmp_team_pool = NULL;
1296
1297 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1298 here so threadprivate doesn't use stale data */
1299 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1300 __kmp_threadpriv_cache_list));
1301
1302 while (__kmp_threadpriv_cache_list != NULL) {
1303
1304 if (*__kmp_threadpriv_cache_list->addr != NULL) {
1305 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1306 &(*__kmp_threadpriv_cache_list->addr)));
1307
1308 *__kmp_threadpriv_cache_list->addr = NULL;
1309 }
1310 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1311 }
1312
1313 __kmp_init_runtime = FALSE;
1314
1315 /* reset statically initialized locks */
1316 __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1317 __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1318 __kmp_init_bootstrap_lock(&__kmp_console_lock);
1319 __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1320
1321 #if USE_ITT_BUILD
1322 __kmp_itt_reset(); // reset ITT's global state
1323 #endif /* USE_ITT_BUILD */
1324
1325 /* This is necessary to make sure no stale data is left around */
1326 /* AC: customers complain that we use unsafe routines in the atfork
1327 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1328 in dynamic_link when check the presence of shared tbbmalloc library.
1329 Suggestion is to make the library initialization lazier, similar
1330 to what done for __kmpc_begin(). */
1331 // TODO: synchronize all static initializations with regular library
1332 // startup; look at kmp_global.cpp and etc.
1333 //__kmp_internal_begin ();
1334 }
1335
__kmp_register_atfork(void)1336 void __kmp_register_atfork(void) {
1337 if (__kmp_need_register_atfork) {
1338 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1339 __kmp_atfork_child);
1340 KMP_CHECK_SYSFAIL("pthread_atfork", status);
1341 __kmp_need_register_atfork = FALSE;
1342 }
1343 }
1344
__kmp_suspend_initialize(void)1345 void __kmp_suspend_initialize(void) {
1346 int status;
1347 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1348 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1349 status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1350 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1351 }
1352
__kmp_suspend_initialize_thread(kmp_info_t * th)1353 void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1354 ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count);
1355 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count);
1356 int new_value = __kmp_fork_count + 1;
1357 // Return if already initialized
1358 if (old_value == new_value)
1359 return;
1360 // Wait, then return if being initialized
1361 if (old_value == -1 ||
1362 !__kmp_atomic_compare_store(&th->th.th_suspend_init_count, old_value,
1363 -1)) {
1364 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) {
1365 KMP_CPU_PAUSE();
1366 }
1367 } else {
1368 // Claim to be the initializer and do initializations
1369 int status;
1370 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1371 &__kmp_suspend_cond_attr);
1372 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1373 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1374 &__kmp_suspend_mutex_attr);
1375 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1376 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value);
1377 ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count);
1378 }
1379 }
1380
__kmp_suspend_uninitialize_thread(kmp_info_t * th)1381 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1382 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) {
1383 /* this means we have initialize the suspension pthread objects for this
1384 thread in this instance of the process */
1385 int status;
1386
1387 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1388 if (status != 0 && status != EBUSY) {
1389 KMP_SYSFAIL("pthread_cond_destroy", status);
1390 }
1391 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1392 if (status != 0 && status != EBUSY) {
1393 KMP_SYSFAIL("pthread_mutex_destroy", status);
1394 }
1395 --th->th.th_suspend_init_count;
1396 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) ==
1397 __kmp_fork_count);
1398 }
1399 }
1400
1401 // return true if lock obtained, false otherwise
__kmp_try_suspend_mx(kmp_info_t * th)1402 int __kmp_try_suspend_mx(kmp_info_t *th) {
1403 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1404 }
1405
__kmp_lock_suspend_mx(kmp_info_t * th)1406 void __kmp_lock_suspend_mx(kmp_info_t *th) {
1407 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1408 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1409 }
1410
__kmp_unlock_suspend_mx(kmp_info_t * th)1411 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1412 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1413 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1414 }
1415
1416 /* This routine puts the calling thread to sleep after setting the
1417 sleep bit for the indicated flag variable to true. */
1418 template <class C>
__kmp_suspend_template(int th_gtid,C * flag)1419 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1420 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1421 kmp_info_t *th = __kmp_threads[th_gtid];
1422 int status;
1423 typename C::flag_t old_spin;
1424
1425 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1426 flag->get()));
1427
1428 __kmp_suspend_initialize_thread(th);
1429
1430 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1431 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1432
1433 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1434 th_gtid, flag->get()));
1435
1436 /* TODO: shouldn't this use release semantics to ensure that
1437 __kmp_suspend_initialize_thread gets called first? */
1438 old_spin = flag->set_sleeping();
1439 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1440 __kmp_pause_status != kmp_soft_paused) {
1441 flag->unset_sleeping();
1442 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1443 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1444 return;
1445 }
1446 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1447 " was %x\n",
1448 th_gtid, flag->get(), flag->load(), old_spin));
1449
1450 if (flag->done_check_val(old_spin)) {
1451 old_spin = flag->unset_sleeping();
1452 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1453 "for spin(%p)\n",
1454 th_gtid, flag->get()));
1455 } else {
1456 /* Encapsulate in a loop as the documentation states that this may
1457 "with low probability" return when the condition variable has
1458 not been signaled or broadcast */
1459 int deactivated = FALSE;
1460 TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1461
1462 while (flag->is_sleeping()) {
1463 #ifdef DEBUG_SUSPEND
1464 char buffer[128];
1465 __kmp_suspend_count++;
1466 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1467 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1468 buffer);
1469 #endif
1470 // Mark the thread as no longer active (only in the first iteration of the
1471 // loop).
1472 if (!deactivated) {
1473 th->th.th_active = FALSE;
1474 if (th->th.th_active_in_pool) {
1475 th->th.th_active_in_pool = FALSE;
1476 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1477 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1478 }
1479 deactivated = TRUE;
1480 }
1481
1482 #if USE_SUSPEND_TIMEOUT
1483 struct timespec now;
1484 struct timeval tval;
1485 int msecs;
1486
1487 status = gettimeofday(&tval, NULL);
1488 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1489 TIMEVAL_TO_TIMESPEC(&tval, &now);
1490
1491 msecs = (4 * __kmp_dflt_blocktime) + 200;
1492 now.tv_sec += msecs / 1000;
1493 now.tv_nsec += (msecs % 1000) * 1000;
1494
1495 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1496 "pthread_cond_timedwait\n",
1497 th_gtid));
1498 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1499 &th->th.th_suspend_mx.m_mutex, &now);
1500 #else
1501 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1502 " pthread_cond_wait\n",
1503 th_gtid));
1504 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1505 &th->th.th_suspend_mx.m_mutex);
1506 #endif
1507
1508 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1509 KMP_SYSFAIL("pthread_cond_wait", status);
1510 }
1511 #ifdef KMP_DEBUG
1512 if (status == ETIMEDOUT) {
1513 if (flag->is_sleeping()) {
1514 KF_TRACE(100,
1515 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1516 } else {
1517 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1518 "not set!\n",
1519 th_gtid));
1520 }
1521 } else if (flag->is_sleeping()) {
1522 KF_TRACE(100,
1523 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1524 }
1525 #endif
1526 } // while
1527
1528 // Mark the thread as active again (if it was previous marked as inactive)
1529 if (deactivated) {
1530 th->th.th_active = TRUE;
1531 if (TCR_4(th->th.th_in_pool)) {
1532 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1533 th->th.th_active_in_pool = TRUE;
1534 }
1535 }
1536 }
1537 #ifdef DEBUG_SUSPEND
1538 {
1539 char buffer[128];
1540 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1541 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1542 buffer);
1543 }
1544 #endif
1545
1546 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1547 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1548 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1549 }
1550
__kmp_suspend_32(int th_gtid,kmp_flag_32 * flag)1551 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) {
1552 __kmp_suspend_template(th_gtid, flag);
1553 }
__kmp_suspend_64(int th_gtid,kmp_flag_64 * flag)1554 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) {
1555 __kmp_suspend_template(th_gtid, flag);
1556 }
__kmp_suspend_oncore(int th_gtid,kmp_flag_oncore * flag)1557 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1558 __kmp_suspend_template(th_gtid, flag);
1559 }
1560
1561 /* This routine signals the thread specified by target_gtid to wake up
1562 after setting the sleep bit indicated by the flag argument to FALSE.
1563 The target thread must already have called __kmp_suspend_template() */
1564 template <class C>
__kmp_resume_template(int target_gtid,C * flag)1565 static inline void __kmp_resume_template(int target_gtid, C *flag) {
1566 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1567 kmp_info_t *th = __kmp_threads[target_gtid];
1568 int status;
1569
1570 #ifdef KMP_DEBUG
1571 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1572 #endif
1573
1574 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1575 gtid, target_gtid));
1576 KMP_DEBUG_ASSERT(gtid != target_gtid);
1577
1578 __kmp_suspend_initialize_thread(th);
1579
1580 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1581 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1582
1583 if (!flag) { // coming from __kmp_null_resume_wrapper
1584 flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1585 }
1586
1587 // First, check if the flag is null or its type has changed. If so, someone
1588 // else woke it up.
1589 if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
1590 // simply shows what
1591 // flag was cast to
1592 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1593 "awake: flag(%p)\n",
1594 gtid, target_gtid, NULL));
1595 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1596 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1597 return;
1598 } else { // if multiple threads are sleeping, flag should be internally
1599 // referring to a specific thread here
1600 typename C::flag_t old_spin = flag->unset_sleeping();
1601 if (!flag->is_sleeping_val(old_spin)) {
1602 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1603 "awake: flag(%p): "
1604 "%u => %u\n",
1605 gtid, target_gtid, flag->get(), old_spin, flag->load()));
1606 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1607 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1608 return;
1609 }
1610 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1611 "sleep bit for flag's loc(%p): "
1612 "%u => %u\n",
1613 gtid, target_gtid, flag->get(), old_spin, flag->load()));
1614 }
1615 TCW_PTR(th->th.th_sleep_loc, NULL);
1616
1617 #ifdef DEBUG_SUSPEND
1618 {
1619 char buffer[128];
1620 __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1621 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1622 target_gtid, buffer);
1623 }
1624 #endif
1625 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1626 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1627 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1628 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1629 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1630 " for T#%d\n",
1631 gtid, target_gtid));
1632 }
1633
__kmp_resume_32(int target_gtid,kmp_flag_32 * flag)1634 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) {
1635 __kmp_resume_template(target_gtid, flag);
1636 }
__kmp_resume_64(int target_gtid,kmp_flag_64 * flag)1637 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) {
1638 __kmp_resume_template(target_gtid, flag);
1639 }
__kmp_resume_oncore(int target_gtid,kmp_flag_oncore * flag)1640 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1641 __kmp_resume_template(target_gtid, flag);
1642 }
1643
1644 #if KMP_USE_MONITOR
__kmp_resume_monitor()1645 void __kmp_resume_monitor() {
1646 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1647 int status;
1648 #ifdef KMP_DEBUG
1649 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1650 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1651 KMP_GTID_MONITOR));
1652 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1653 #endif
1654 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1655 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1656 #ifdef DEBUG_SUSPEND
1657 {
1658 char buffer[128];
1659 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1660 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1661 KMP_GTID_MONITOR, buffer);
1662 }
1663 #endif
1664 status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1665 KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1666 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1667 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1668 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1669 " for T#%d\n",
1670 gtid, KMP_GTID_MONITOR));
1671 }
1672 #endif // KMP_USE_MONITOR
1673
__kmp_yield()1674 void __kmp_yield() { sched_yield(); }
1675
__kmp_gtid_set_specific(int gtid)1676 void __kmp_gtid_set_specific(int gtid) {
1677 if (__kmp_init_gtid) {
1678 int status;
1679 status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1680 (void *)(intptr_t)(gtid + 1));
1681 KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1682 } else {
1683 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1684 }
1685 }
1686
__kmp_gtid_get_specific()1687 int __kmp_gtid_get_specific() {
1688 int gtid;
1689 if (!__kmp_init_gtid) {
1690 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1691 "KMP_GTID_SHUTDOWN\n"));
1692 return KMP_GTID_SHUTDOWN;
1693 }
1694 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1695 if (gtid == 0) {
1696 gtid = KMP_GTID_DNE;
1697 } else {
1698 gtid--;
1699 }
1700 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1701 __kmp_gtid_threadprivate_key, gtid));
1702 return gtid;
1703 }
1704
__kmp_read_cpu_time(void)1705 double __kmp_read_cpu_time(void) {
1706 /*clock_t t;*/
1707 struct tms buffer;
1708
1709 /*t =*/times(&buffer);
1710
1711 return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC;
1712 }
1713
__kmp_read_system_info(struct kmp_sys_info * info)1714 int __kmp_read_system_info(struct kmp_sys_info *info) {
1715 int status;
1716 struct rusage r_usage;
1717
1718 memset(info, 0, sizeof(*info));
1719
1720 status = getrusage(RUSAGE_SELF, &r_usage);
1721 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1722
1723 // The maximum resident set size utilized (in kilobytes)
1724 info->maxrss = r_usage.ru_maxrss;
1725 // The number of page faults serviced without any I/O
1726 info->minflt = r_usage.ru_minflt;
1727 // The number of page faults serviced that required I/O
1728 info->majflt = r_usage.ru_majflt;
1729 // The number of times a process was "swapped" out of memory
1730 info->nswap = r_usage.ru_nswap;
1731 // The number of times the file system had to perform input
1732 info->inblock = r_usage.ru_inblock;
1733 // The number of times the file system had to perform output
1734 info->oublock = r_usage.ru_oublock;
1735 // The number of times a context switch was voluntarily
1736 info->nvcsw = r_usage.ru_nvcsw;
1737 // The number of times a context switch was forced
1738 info->nivcsw = r_usage.ru_nivcsw;
1739
1740 return (status != 0);
1741 }
1742
__kmp_read_system_time(double * delta)1743 void __kmp_read_system_time(double *delta) {
1744 double t_ns;
1745 struct timeval tval;
1746 struct timespec stop;
1747 int status;
1748
1749 status = gettimeofday(&tval, NULL);
1750 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1751 TIMEVAL_TO_TIMESPEC(&tval, &stop);
1752 t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start);
1753 *delta = (t_ns * 1e-9);
1754 }
1755
__kmp_clear_system_time(void)1756 void __kmp_clear_system_time(void) {
1757 struct timeval tval;
1758 int status;
1759 status = gettimeofday(&tval, NULL);
1760 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1761 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1762 }
1763
__kmp_get_xproc(void)1764 static int __kmp_get_xproc(void) {
1765
1766 int r = 0;
1767
1768 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \
1769 KMP_OS_OPENBSD || KMP_OS_HURD
1770
1771 r = sysconf(_SC_NPROCESSORS_ONLN);
1772
1773 #elif KMP_OS_DARWIN
1774
1775 // Bug C77011 High "OpenMP Threads and number of active cores".
1776
1777 // Find the number of available CPUs.
1778 kern_return_t rc;
1779 host_basic_info_data_t info;
1780 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1781 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1782 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1783 // Cannot use KA_TRACE() here because this code works before trace support
1784 // is initialized.
1785 r = info.avail_cpus;
1786 } else {
1787 KMP_WARNING(CantGetNumAvailCPU);
1788 KMP_INFORM(AssumedNumCPU);
1789 }
1790
1791 #else
1792
1793 #error "Unknown or unsupported OS."
1794
1795 #endif
1796
1797 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1798
1799 } // __kmp_get_xproc
1800
__kmp_read_from_file(char const * path,char const * format,...)1801 int __kmp_read_from_file(char const *path, char const *format, ...) {
1802 int result;
1803 va_list args;
1804
1805 va_start(args, format);
1806 FILE *f = fopen(path, "rb");
1807 if (f == NULL)
1808 return 0;
1809 result = vfscanf(f, format, args);
1810 fclose(f);
1811
1812 return result;
1813 }
1814
__kmp_runtime_initialize(void)1815 void __kmp_runtime_initialize(void) {
1816 int status;
1817 pthread_mutexattr_t mutex_attr;
1818 pthread_condattr_t cond_attr;
1819
1820 if (__kmp_init_runtime) {
1821 return;
1822 }
1823
1824 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1825 if (!__kmp_cpuinfo.initialized) {
1826 __kmp_query_cpuid(&__kmp_cpuinfo);
1827 }
1828 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1829
1830 __kmp_xproc = __kmp_get_xproc();
1831
1832 #if ! KMP_32_BIT_ARCH
1833 struct rlimit rlim;
1834 // read stack size of calling thread, save it as default for worker threads;
1835 // this should be done before reading environment variables
1836 status = getrlimit(RLIMIT_STACK, &rlim);
1837 if (status == 0) { // success?
1838 __kmp_stksize = rlim.rlim_cur;
1839 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed
1840 }
1841 #endif /* KMP_32_BIT_ARCH */
1842
1843 if (sysconf(_SC_THREADS)) {
1844
1845 /* Query the maximum number of threads */
1846 __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX);
1847 if (__kmp_sys_max_nth == -1) {
1848 /* Unlimited threads for NPTL */
1849 __kmp_sys_max_nth = INT_MAX;
1850 } else if (__kmp_sys_max_nth <= 1) {
1851 /* Can't tell, just use PTHREAD_THREADS_MAX */
1852 __kmp_sys_max_nth = KMP_MAX_NTH;
1853 }
1854
1855 /* Query the minimum stack size */
1856 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1857 if (__kmp_sys_min_stksize <= 1) {
1858 __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1859 }
1860 }
1861
1862 /* Set up minimum number of threads to switch to TLS gtid */
1863 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1864
1865 status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1866 __kmp_internal_end_dest);
1867 KMP_CHECK_SYSFAIL("pthread_key_create", status);
1868 status = pthread_mutexattr_init(&mutex_attr);
1869 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1870 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1871 KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1872 status = pthread_condattr_init(&cond_attr);
1873 KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1874 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1875 KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1876 #if USE_ITT_BUILD
1877 __kmp_itt_initialize();
1878 #endif /* USE_ITT_BUILD */
1879
1880 __kmp_init_runtime = TRUE;
1881 }
1882
__kmp_runtime_destroy(void)1883 void __kmp_runtime_destroy(void) {
1884 int status;
1885
1886 if (!__kmp_init_runtime) {
1887 return; // Nothing to do.
1888 }
1889
1890 #if USE_ITT_BUILD
1891 __kmp_itt_destroy();
1892 #endif /* USE_ITT_BUILD */
1893
1894 status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1895 KMP_CHECK_SYSFAIL("pthread_key_delete", status);
1896
1897 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1898 if (status != 0 && status != EBUSY) {
1899 KMP_SYSFAIL("pthread_mutex_destroy", status);
1900 }
1901 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1902 if (status != 0 && status != EBUSY) {
1903 KMP_SYSFAIL("pthread_cond_destroy", status);
1904 }
1905 #if KMP_AFFINITY_SUPPORTED
1906 __kmp_affinity_uninitialize();
1907 #endif
1908
1909 __kmp_init_runtime = FALSE;
1910 }
1911
1912 /* Put the thread to sleep for a time period */
1913 /* NOTE: not currently used anywhere */
__kmp_thread_sleep(int millis)1914 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
1915
1916 /* Calculate the elapsed wall clock time for the user */
__kmp_elapsed(double * t)1917 void __kmp_elapsed(double *t) {
1918 int status;
1919 #ifdef FIX_SGI_CLOCK
1920 struct timespec ts;
1921
1922 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1923 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
1924 *t =
1925 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
1926 #else
1927 struct timeval tv;
1928
1929 status = gettimeofday(&tv, NULL);
1930 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1931 *t =
1932 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
1933 #endif
1934 }
1935
1936 /* Calculate the elapsed wall clock tick for the user */
__kmp_elapsed_tick(double * t)1937 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1938
1939 /* Return the current time stamp in nsec */
__kmp_now_nsec()1940 kmp_uint64 __kmp_now_nsec() {
1941 struct timeval t;
1942 gettimeofday(&t, NULL);
1943 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
1944 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
1945 return nsec;
1946 }
1947
1948 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1949 /* Measure clock ticks per millisecond */
__kmp_initialize_system_tick()1950 void __kmp_initialize_system_tick() {
1951 kmp_uint64 now, nsec2, diff;
1952 kmp_uint64 delay = 100000; // 50~100 usec on most machines.
1953 kmp_uint64 nsec = __kmp_now_nsec();
1954 kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
1955 while ((now = __kmp_hardware_timestamp()) < goal)
1956 ;
1957 nsec2 = __kmp_now_nsec();
1958 diff = nsec2 - nsec;
1959 if (diff > 0) {
1960 kmp_uint64 tpms = (kmp_uint64)(1e6 * (delay + (now - goal)) / diff);
1961 if (tpms > 0)
1962 __kmp_ticks_per_msec = tpms;
1963 }
1964 }
1965 #endif
1966
1967 /* Determine whether the given address is mapped into the current address
1968 space. */
1969
__kmp_is_address_mapped(void * addr)1970 int __kmp_is_address_mapped(void *addr) {
1971
1972 int found = 0;
1973 int rc;
1974
1975 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_HURD
1976
1977 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the address
1978 ranges mapped into the address space. */
1979
1980 char *name = __kmp_str_format("/proc/%d/maps", getpid());
1981 FILE *file = NULL;
1982
1983 file = fopen(name, "r");
1984 KMP_ASSERT(file != NULL);
1985
1986 for (;;) {
1987
1988 void *beginning = NULL;
1989 void *ending = NULL;
1990 char perms[5];
1991
1992 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
1993 if (rc == EOF) {
1994 break;
1995 }
1996 KMP_ASSERT(rc == 3 &&
1997 KMP_STRLEN(perms) == 4); // Make sure all fields are read.
1998
1999 // Ending address is not included in the region, but beginning is.
2000 if ((addr >= beginning) && (addr < ending)) {
2001 perms[2] = 0; // 3th and 4th character does not matter.
2002 if (strcmp(perms, "rw") == 0) {
2003 // Memory we are looking for should be readable and writable.
2004 found = 1;
2005 }
2006 break;
2007 }
2008 }
2009
2010 // Free resources.
2011 fclose(file);
2012 KMP_INTERNAL_FREE(name);
2013
2014 #elif KMP_OS_DARWIN
2015
2016 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2017 using vm interface. */
2018
2019 int buffer;
2020 vm_size_t count;
2021 rc = vm_read_overwrite(
2022 mach_task_self(), // Task to read memory of.
2023 (vm_address_t)(addr), // Address to read from.
2024 1, // Number of bytes to be read.
2025 (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2026 &count // Address of var to save number of read bytes in.
2027 );
2028 if (rc == 0) {
2029 // Memory successfully read.
2030 found = 1;
2031 }
2032
2033 #elif KMP_OS_NETBSD
2034
2035 int mib[5];
2036 mib[0] = CTL_VM;
2037 mib[1] = VM_PROC;
2038 mib[2] = VM_PROC_MAP;
2039 mib[3] = getpid();
2040 mib[4] = sizeof(struct kinfo_vmentry);
2041
2042 size_t size;
2043 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2044 KMP_ASSERT(!rc);
2045 KMP_ASSERT(size);
2046
2047 size = size * 4 / 3;
2048 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2049 KMP_ASSERT(kiv);
2050
2051 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2052 KMP_ASSERT(!rc);
2053 KMP_ASSERT(size);
2054
2055 for (size_t i = 0; i < size; i++) {
2056 if (kiv[i].kve_start >= (uint64_t)addr &&
2057 kiv[i].kve_end <= (uint64_t)addr) {
2058 found = 1;
2059 break;
2060 }
2061 }
2062 KMP_INTERNAL_FREE(kiv);
2063 #elif KMP_OS_DRAGONFLY || KMP_OS_OPENBSD
2064
2065 // FIXME(DragonFly, OpenBSD): Implement this
2066 found = 1;
2067
2068 #else
2069
2070 #error "Unknown or unsupported OS"
2071
2072 #endif
2073
2074 return found;
2075
2076 } // __kmp_is_address_mapped
2077
2078 #ifdef USE_LOAD_BALANCE
2079
2080 #if KMP_OS_DARWIN || KMP_OS_NETBSD
2081
2082 // The function returns the rounded value of the system load average
2083 // during given time interval which depends on the value of
2084 // __kmp_load_balance_interval variable (default is 60 sec, other values
2085 // may be 300 sec or 900 sec).
2086 // It returns -1 in case of error.
__kmp_get_load_balance(int max)2087 int __kmp_get_load_balance(int max) {
2088 double averages[3];
2089 int ret_avg = 0;
2090
2091 int res = getloadavg(averages, 3);
2092
2093 // Check __kmp_load_balance_interval to determine which of averages to use.
2094 // getloadavg() may return the number of samples less than requested that is
2095 // less than 3.
2096 if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2097 ret_avg = averages[0]; // 1 min
2098 } else if ((__kmp_load_balance_interval >= 180 &&
2099 __kmp_load_balance_interval < 600) &&
2100 (res >= 2)) {
2101 ret_avg = averages[1]; // 5 min
2102 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2103 ret_avg = averages[2]; // 15 min
2104 } else { // Error occurred
2105 return -1;
2106 }
2107
2108 return ret_avg;
2109 }
2110
2111 #else // Linux* OS
2112
2113 // The fuction returns number of running (not sleeping) threads, or -1 in case
2114 // of error. Error could be reported if Linux* OS kernel too old (without
2115 // "/proc" support). Counting running threads stops if max running threads
2116 // encountered.
__kmp_get_load_balance(int max)2117 int __kmp_get_load_balance(int max) {
2118 static int permanent_error = 0;
2119 static int glb_running_threads = 0; // Saved count of the running threads for
2120 // the thread balance algortihm
2121 static double glb_call_time = 0; /* Thread balance algorithm call time */
2122
2123 int running_threads = 0; // Number of running threads in the system.
2124
2125 DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2126 struct dirent *proc_entry = NULL;
2127
2128 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2129 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2130 struct dirent *task_entry = NULL;
2131 int task_path_fixed_len;
2132
2133 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2134 int stat_file = -1;
2135 int stat_path_fixed_len;
2136
2137 int total_processes = 0; // Total number of processes in system.
2138 int total_threads = 0; // Total number of threads in system.
2139
2140 double call_time = 0.0;
2141
2142 __kmp_str_buf_init(&task_path);
2143 __kmp_str_buf_init(&stat_path);
2144
2145 __kmp_elapsed(&call_time);
2146
2147 if (glb_call_time &&
2148 (call_time - glb_call_time < __kmp_load_balance_interval)) {
2149 running_threads = glb_running_threads;
2150 goto finish;
2151 }
2152
2153 glb_call_time = call_time;
2154
2155 // Do not spend time on scanning "/proc/" if we have a permanent error.
2156 if (permanent_error) {
2157 running_threads = -1;
2158 goto finish;
2159 }
2160
2161 if (max <= 0) {
2162 max = INT_MAX;
2163 }
2164
2165 // Open "/proc/" directory.
2166 proc_dir = opendir("/proc");
2167 if (proc_dir == NULL) {
2168 // Cannot open "/prroc/". Probably the kernel does not support it. Return an
2169 // error now and in subsequent calls.
2170 running_threads = -1;
2171 permanent_error = 1;
2172 goto finish;
2173 }
2174
2175 // Initialize fixed part of task_path. This part will not change.
2176 __kmp_str_buf_cat(&task_path, "/proc/", 6);
2177 task_path_fixed_len = task_path.used; // Remember number of used characters.
2178
2179 proc_entry = readdir(proc_dir);
2180 while (proc_entry != NULL) {
2181 // Proc entry is a directory and name starts with a digit. Assume it is a
2182 // process' directory.
2183 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2184
2185 ++total_processes;
2186 // Make sure init process is the very first in "/proc", so we can replace
2187 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2188 // 1. We are going to check that total_processes == 1 => d_name == "1" is
2189 // true (where "=>" is implication). Since C++ does not have => operator,
2190 // let us replace it with its equivalent: a => b == ! a || b.
2191 KMP_DEBUG_ASSERT(total_processes != 1 ||
2192 strcmp(proc_entry->d_name, "1") == 0);
2193
2194 // Construct task_path.
2195 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2196 __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2197 KMP_STRLEN(proc_entry->d_name));
2198 __kmp_str_buf_cat(&task_path, "/task", 5);
2199
2200 task_dir = opendir(task_path.str);
2201 if (task_dir == NULL) {
2202 // Process can finish between reading "/proc/" directory entry and
2203 // opening process' "task/" directory. So, in general case we should not
2204 // complain, but have to skip this process and read the next one. But on
2205 // systems with no "task/" support we will spend lot of time to scan
2206 // "/proc/" tree again and again without any benefit. "init" process
2207 // (its pid is 1) should exist always, so, if we cannot open
2208 // "/proc/1/task/" directory, it means "task/" is not supported by
2209 // kernel. Report an error now and in the future.
2210 if (strcmp(proc_entry->d_name, "1") == 0) {
2211 running_threads = -1;
2212 permanent_error = 1;
2213 goto finish;
2214 }
2215 } else {
2216 // Construct fixed part of stat file path.
2217 __kmp_str_buf_clear(&stat_path);
2218 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2219 __kmp_str_buf_cat(&stat_path, "/", 1);
2220 stat_path_fixed_len = stat_path.used;
2221
2222 task_entry = readdir(task_dir);
2223 while (task_entry != NULL) {
2224 // It is a directory and name starts with a digit.
2225 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2226 ++total_threads;
2227
2228 // Consruct complete stat file path. Easiest way would be:
2229 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2230 // task_entry->d_name );
2231 // but seriae of __kmp_str_buf_cat works a bit faster.
2232 stat_path.used =
2233 stat_path_fixed_len; // Reset stat path to its fixed part.
2234 __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2235 KMP_STRLEN(task_entry->d_name));
2236 __kmp_str_buf_cat(&stat_path, "/stat", 5);
2237
2238 // Note: Low-level API (open/read/close) is used. High-level API
2239 // (fopen/fclose) works ~ 30 % slower.
2240 stat_file = open(stat_path.str, O_RDONLY);
2241 if (stat_file == -1) {
2242 // We cannot report an error because task (thread) can terminate
2243 // just before reading this file.
2244 } else {
2245 /* Content of "stat" file looks like:
2246 24285 (program) S ...
2247
2248 It is a single line (if program name does not include funny
2249 symbols). First number is a thread id, then name of executable
2250 file name in paretheses, then state of the thread. We need just
2251 thread state.
2252
2253 Good news: Length of program name is 15 characters max. Longer
2254 names are truncated.
2255
2256 Thus, we need rather short buffer: 15 chars for program name +
2257 2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2258
2259 Bad news: Program name may contain special symbols like space,
2260 closing parenthesis, or even new line. This makes parsing
2261 "stat" file not 100 % reliable. In case of fanny program names
2262 parsing may fail (report incorrect thread state).
2263
2264 Parsing "status" file looks more promissing (due to different
2265 file structure and escaping special symbols) but reading and
2266 parsing of "status" file works slower.
2267 -- ln
2268 */
2269 char buffer[65];
2270 int len;
2271 len = read(stat_file, buffer, sizeof(buffer) - 1);
2272 if (len >= 0) {
2273 buffer[len] = 0;
2274 // Using scanf:
2275 // sscanf( buffer, "%*d (%*s) %c ", & state );
2276 // looks very nice, but searching for a closing parenthesis
2277 // works a bit faster.
2278 char *close_parent = strstr(buffer, ") ");
2279 if (close_parent != NULL) {
2280 char state = *(close_parent + 2);
2281 if (state == 'R') {
2282 ++running_threads;
2283 if (running_threads >= max) {
2284 goto finish;
2285 }
2286 }
2287 }
2288 }
2289 close(stat_file);
2290 stat_file = -1;
2291 }
2292 }
2293 task_entry = readdir(task_dir);
2294 }
2295 closedir(task_dir);
2296 task_dir = NULL;
2297 }
2298 }
2299 proc_entry = readdir(proc_dir);
2300 }
2301
2302 // There _might_ be a timing hole where the thread executing this
2303 // code get skipped in the load balance, and running_threads is 0.
2304 // Assert in the debug builds only!!!
2305 KMP_DEBUG_ASSERT(running_threads > 0);
2306 if (running_threads <= 0) {
2307 running_threads = 1;
2308 }
2309
2310 finish: // Clean up and exit.
2311 if (proc_dir != NULL) {
2312 closedir(proc_dir);
2313 }
2314 __kmp_str_buf_free(&task_path);
2315 if (task_dir != NULL) {
2316 closedir(task_dir);
2317 }
2318 __kmp_str_buf_free(&stat_path);
2319 if (stat_file != -1) {
2320 close(stat_file);
2321 }
2322
2323 glb_running_threads = running_threads;
2324
2325 return running_threads;
2326
2327 } // __kmp_get_load_balance
2328
2329 #endif // KMP_OS_DARWIN
2330
2331 #endif // USE_LOAD_BALANCE
2332
2333 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \
2334 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || KMP_ARCH_PPC64)
2335
2336 // we really only need the case with 1 argument, because CLANG always build
2337 // a struct of pointers to shared variables referenced in the outlined function
__kmp_invoke_microtask(microtask_t pkfn,int gtid,int tid,int argc,void * p_argv[],void ** exit_frame_ptr)2338 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2339 void *p_argv[]
2340 #if OMPT_SUPPORT
2341 ,
2342 void **exit_frame_ptr
2343 #endif
2344 ) {
2345 #if OMPT_SUPPORT
2346 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2347 #endif
2348
2349 switch (argc) {
2350 default:
2351 fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2352 fflush(stderr);
2353 exit(-1);
2354 case 0:
2355 (*pkfn)(>id, &tid);
2356 break;
2357 case 1:
2358 (*pkfn)(>id, &tid, p_argv[0]);
2359 break;
2360 case 2:
2361 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]);
2362 break;
2363 case 3:
2364 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]);
2365 break;
2366 case 4:
2367 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2368 break;
2369 case 5:
2370 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2371 break;
2372 case 6:
2373 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2374 p_argv[5]);
2375 break;
2376 case 7:
2377 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2378 p_argv[5], p_argv[6]);
2379 break;
2380 case 8:
2381 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2382 p_argv[5], p_argv[6], p_argv[7]);
2383 break;
2384 case 9:
2385 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2386 p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2387 break;
2388 case 10:
2389 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2390 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2391 break;
2392 case 11:
2393 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2394 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2395 break;
2396 case 12:
2397 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2398 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2399 p_argv[11]);
2400 break;
2401 case 13:
2402 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2403 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2404 p_argv[11], p_argv[12]);
2405 break;
2406 case 14:
2407 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2408 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2409 p_argv[11], p_argv[12], p_argv[13]);
2410 break;
2411 case 15:
2412 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2413 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2414 p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2415 break;
2416 }
2417
2418 #if OMPT_SUPPORT
2419 *exit_frame_ptr = 0;
2420 #endif
2421
2422 return 1;
2423 }
2424
2425 #endif
2426
2427 // end of file //
2428