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