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