1 /*
2  * z_Linux_util.cpp -- platform specific routines.
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 //                     The LLVM Compiler Infrastructure
8 //
9 // This file is dual licensed under the MIT and the University of Illinois Open
10 // Source Licenses. See LICENSE.txt for details.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "kmp.h"
15 #include "kmp_affinity.h"
16 #include "kmp_i18n.h"
17 #include "kmp_io.h"
18 #include "kmp_itt.h"
19 #include "kmp_lock.h"
20 #include "kmp_stats.h"
21 #include "kmp_str.h"
22 #include "kmp_wait_release.h"
23 #include "kmp_wrapper_getpid.h"
24 
25 #if !KMP_OS_DRAGONFLY && !KMP_OS_FREEBSD && !KMP_OS_NETBSD && !KMP_OS_OPENBSD
26 #include <alloca.h>
27 #endif
28 #include <math.h> // HUGE_VAL.
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 && !KMP_OS_CNK
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 <pthread_np.h>
55 #elif KMP_OS_NETBSD
56 #include <sys/types.h>
57 #include <sys/sysctl.h>
58 #endif
59 
60 #include <ctype.h>
61 #include <dirent.h>
62 #include <fcntl.h>
63 
64 #include "tsan_annotations.h"
65 
66 struct kmp_sys_timer {
67   struct timespec start;
68 };
69 
70 // Convert timespec to nanoseconds.
71 #define TS2NS(timespec) (((timespec).tv_sec * 1e9) + (timespec).tv_nsec)
72 
73 static struct kmp_sys_timer __kmp_sys_timer_data;
74 
75 #if KMP_HANDLE_SIGNALS
76 typedef void (*sig_func_t)(int);
77 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG];
78 static sigset_t __kmp_sigset;
79 #endif
80 
81 static int __kmp_init_runtime = FALSE;
82 
83 static int __kmp_fork_count = 0;
84 
85 static pthread_condattr_t __kmp_suspend_cond_attr;
86 static pthread_mutexattr_t __kmp_suspend_mutex_attr;
87 
88 static kmp_cond_align_t __kmp_wait_cv;
89 static kmp_mutex_align_t __kmp_wait_mx;
90 
91 kmp_uint64 __kmp_ticks_per_msec = 1000000;
92 
93 #ifdef DEBUG_SUSPEND
94 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
95   KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
96                cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
97                cond->c_cond.__c_waiting);
98 }
99 #endif
100 
101 #if (KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED)
102 
103 /* Affinity support */
104 
105 void __kmp_affinity_bind_thread(int which) {
106   KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
107               "Illegal set affinity operation when not capable");
108 
109   kmp_affin_mask_t *mask;
110   KMP_CPU_ALLOC_ON_STACK(mask);
111   KMP_CPU_ZERO(mask);
112   KMP_CPU_SET(which, mask);
113   __kmp_set_system_affinity(mask, TRUE);
114   KMP_CPU_FREE_FROM_STACK(mask);
115 }
116 
117 /* Determine if we can access affinity functionality on this version of
118  * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
119  * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
120 void __kmp_affinity_determine_capable(const char *env_var) {
121 // Check and see if the OS supports thread affinity.
122 
123 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
124 
125   int gCode;
126   int sCode;
127   unsigned char *buf;
128   buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
129 
130   // If Linux* OS:
131   // If the syscall fails or returns a suggestion for the size,
132   // then we don't have to search for an appropriate size.
133   gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_SIZE_LIMIT, buf);
134   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
135                 "initial getaffinity call returned %d errno = %d\n",
136                 gCode, errno));
137 
138   // if ((gCode < 0) && (errno == ENOSYS))
139   if (gCode < 0) {
140     // System call not supported
141     if (__kmp_affinity_verbose ||
142         (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
143          (__kmp_affinity_type != affinity_default) &&
144          (__kmp_affinity_type != affinity_disabled))) {
145       int error = errno;
146       kmp_msg_t err_code = KMP_ERR(error);
147       __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
148                 err_code, __kmp_msg_null);
149       if (__kmp_generate_warnings == kmp_warnings_off) {
150         __kmp_str_free(&err_code.str);
151       }
152     }
153     KMP_AFFINITY_DISABLE();
154     KMP_INTERNAL_FREE(buf);
155     return;
156   }
157   if (gCode > 0) { // Linux* OS only
158     // The optimal situation: the OS returns the size of the buffer it expects.
159     //
160     // A verification of correct behavior is that Isetaffinity on a NULL
161     // buffer with the same size fails with errno set to EFAULT.
162     sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
163     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
164                   "setaffinity for mask size %d returned %d errno = %d\n",
165                   gCode, sCode, errno));
166     if (sCode < 0) {
167       if (errno == ENOSYS) {
168         if (__kmp_affinity_verbose ||
169             (__kmp_affinity_warnings &&
170              (__kmp_affinity_type != affinity_none) &&
171              (__kmp_affinity_type != affinity_default) &&
172              (__kmp_affinity_type != affinity_disabled))) {
173           int error = errno;
174           kmp_msg_t err_code = KMP_ERR(error);
175           __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
176                     err_code, __kmp_msg_null);
177           if (__kmp_generate_warnings == kmp_warnings_off) {
178             __kmp_str_free(&err_code.str);
179           }
180         }
181         KMP_AFFINITY_DISABLE();
182         KMP_INTERNAL_FREE(buf);
183       }
184       if (errno == EFAULT) {
185         KMP_AFFINITY_ENABLE(gCode);
186         KA_TRACE(10, ("__kmp_affinity_determine_capable: "
187                       "affinity supported (mask size %d)\n",
188                       (int)__kmp_affin_mask_size));
189         KMP_INTERNAL_FREE(buf);
190         return;
191       }
192     }
193   }
194 
195   // Call the getaffinity system call repeatedly with increasing set sizes
196   // until we succeed, or reach an upper bound on the search.
197   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
198                 "searching for proper set size\n"));
199   int size;
200   for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
201     gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
202     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
203                   "getaffinity for mask size %d returned %d errno = %d\n",
204                   size, gCode, errno));
205 
206     if (gCode < 0) {
207       if (errno == ENOSYS) {
208         // We shouldn't get here
209         KA_TRACE(30, ("__kmp_affinity_determine_capable: "
210                       "inconsistent OS call behavior: errno == ENOSYS for mask "
211                       "size %d\n",
212                       size));
213         if (__kmp_affinity_verbose ||
214             (__kmp_affinity_warnings &&
215              (__kmp_affinity_type != affinity_none) &&
216              (__kmp_affinity_type != affinity_default) &&
217              (__kmp_affinity_type != affinity_disabled))) {
218           int error = errno;
219           kmp_msg_t err_code = KMP_ERR(error);
220           __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
221                     err_code, __kmp_msg_null);
222           if (__kmp_generate_warnings == kmp_warnings_off) {
223             __kmp_str_free(&err_code.str);
224           }
225         }
226         KMP_AFFINITY_DISABLE();
227         KMP_INTERNAL_FREE(buf);
228         return;
229       }
230       continue;
231     }
232 
233     sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
234     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
235                   "setaffinity for mask size %d returned %d errno = %d\n",
236                   gCode, sCode, errno));
237     if (sCode < 0) {
238       if (errno == ENOSYS) { // Linux* OS only
239         // We shouldn't get here
240         KA_TRACE(30, ("__kmp_affinity_determine_capable: "
241                       "inconsistent OS call behavior: errno == ENOSYS for mask "
242                       "size %d\n",
243                       size));
244         if (__kmp_affinity_verbose ||
245             (__kmp_affinity_warnings &&
246              (__kmp_affinity_type != affinity_none) &&
247              (__kmp_affinity_type != affinity_default) &&
248              (__kmp_affinity_type != affinity_disabled))) {
249           int error = errno;
250           kmp_msg_t err_code = KMP_ERR(error);
251           __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
252                     err_code, __kmp_msg_null);
253           if (__kmp_generate_warnings == kmp_warnings_off) {
254             __kmp_str_free(&err_code.str);
255           }
256         }
257         KMP_AFFINITY_DISABLE();
258         KMP_INTERNAL_FREE(buf);
259         return;
260       }
261       if (errno == EFAULT) {
262         KMP_AFFINITY_ENABLE(gCode);
263         KA_TRACE(10, ("__kmp_affinity_determine_capable: "
264                       "affinity supported (mask size %d)\n",
265                       (int)__kmp_affin_mask_size));
266         KMP_INTERNAL_FREE(buf);
267         return;
268       }
269     }
270   }
271   // save uncaught error code
272   // int error = errno;
273   KMP_INTERNAL_FREE(buf);
274   // restore uncaught error code, will be printed at the next KMP_WARNING below
275   // errno = error;
276 
277   // Affinity is not supported
278   KMP_AFFINITY_DISABLE();
279   KA_TRACE(10, ("__kmp_affinity_determine_capable: "
280                 "cannot determine mask size - affinity not supported\n"));
281   if (__kmp_affinity_verbose ||
282       (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
283        (__kmp_affinity_type != affinity_default) &&
284        (__kmp_affinity_type != affinity_disabled))) {
285     KMP_WARNING(AffCantGetMaskSize, env_var);
286   }
287 }
288 
289 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
290 
291 #if KMP_USE_FUTEX
292 
293 int __kmp_futex_determine_capable() {
294   int loc = 0;
295   int rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
296   int retval = (rc == 0) || (errno != ENOSYS);
297 
298   KA_TRACE(10,
299            ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
300   KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
301                 retval ? "" : " not"));
302 
303   return retval;
304 }
305 
306 #endif // KMP_USE_FUTEX
307 
308 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS)
309 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
310    use compare_and_store for these routines */
311 
312 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
313   kmp_int8 old_value, new_value;
314 
315   old_value = TCR_1(*p);
316   new_value = old_value | d;
317 
318   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
319     KMP_CPU_PAUSE();
320     old_value = TCR_1(*p);
321     new_value = old_value | d;
322   }
323   return old_value;
324 }
325 
326 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
327   kmp_int8 old_value, new_value;
328 
329   old_value = TCR_1(*p);
330   new_value = old_value & d;
331 
332   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
333     KMP_CPU_PAUSE();
334     old_value = TCR_1(*p);
335     new_value = old_value & d;
336   }
337   return old_value;
338 }
339 
340 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
341   kmp_uint32 old_value, new_value;
342 
343   old_value = TCR_4(*p);
344   new_value = old_value | d;
345 
346   while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
347     KMP_CPU_PAUSE();
348     old_value = TCR_4(*p);
349     new_value = old_value | d;
350   }
351   return old_value;
352 }
353 
354 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
355   kmp_uint32 old_value, new_value;
356 
357   old_value = TCR_4(*p);
358   new_value = old_value & d;
359 
360   while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
361     KMP_CPU_PAUSE();
362     old_value = TCR_4(*p);
363     new_value = old_value & d;
364   }
365   return old_value;
366 }
367 
368 #if KMP_ARCH_X86
369 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
370   kmp_int8 old_value, new_value;
371 
372   old_value = TCR_1(*p);
373   new_value = old_value + d;
374 
375   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
376     KMP_CPU_PAUSE();
377     old_value = TCR_1(*p);
378     new_value = old_value + d;
379   }
380   return old_value;
381 }
382 
383 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
384   kmp_int64 old_value, new_value;
385 
386   old_value = TCR_8(*p);
387   new_value = old_value + d;
388 
389   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
390     KMP_CPU_PAUSE();
391     old_value = TCR_8(*p);
392     new_value = old_value + d;
393   }
394   return old_value;
395 }
396 #endif /* KMP_ARCH_X86 */
397 
398 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
399   kmp_uint64 old_value, new_value;
400 
401   old_value = TCR_8(*p);
402   new_value = old_value | d;
403   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
404     KMP_CPU_PAUSE();
405     old_value = TCR_8(*p);
406     new_value = old_value | d;
407   }
408   return old_value;
409 }
410 
411 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
412   kmp_uint64 old_value, new_value;
413 
414   old_value = TCR_8(*p);
415   new_value = old_value & d;
416   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
417     KMP_CPU_PAUSE();
418     old_value = TCR_8(*p);
419     new_value = old_value & d;
420   }
421   return old_value;
422 }
423 
424 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
425 
426 void __kmp_terminate_thread(int gtid) {
427   int status;
428   kmp_info_t *th = __kmp_threads[gtid];
429 
430   if (!th)
431     return;
432 
433 #ifdef KMP_CANCEL_THREADS
434   KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
435   status = pthread_cancel(th->th.th_info.ds.ds_thread);
436   if (status != 0 && status != ESRCH) {
437     __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
438                 __kmp_msg_null);
439   }
440 #endif
441   __kmp_yield(TRUE);
442 } //
443 
444 /* Set thread stack info according to values returned by pthread_getattr_np().
445    If values are unreasonable, assume call failed and use incremental stack
446    refinement method instead. Returns TRUE if the stack parameters could be
447    determined exactly, FALSE if incremental refinement is necessary. */
448 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
449   int stack_data;
450 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
451         KMP_OS_HURD
452   pthread_attr_t attr;
453   int status;
454   size_t size = 0;
455   void *addr = 0;
456 
457   /* Always do incremental stack refinement for ubermaster threads since the
458      initial thread stack range can be reduced by sibling thread creation so
459      pthread_attr_getstack may cause thread gtid aliasing */
460   if (!KMP_UBER_GTID(gtid)) {
461 
462     /* Fetch the real thread attributes */
463     status = pthread_attr_init(&attr);
464     KMP_CHECK_SYSFAIL("pthread_attr_init", status);
465 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
466     status = pthread_attr_get_np(pthread_self(), &attr);
467     KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
468 #else
469     status = pthread_getattr_np(pthread_self(), &attr);
470     KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
471 #endif
472     status = pthread_attr_getstack(&attr, &addr, &size);
473     KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
474     KA_TRACE(60,
475              ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
476               " %lu, low addr: %p\n",
477               gtid, size, addr));
478     status = pthread_attr_destroy(&attr);
479     KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
480   }
481 
482   if (size != 0 && addr != 0) { // was stack parameter determination successful?
483     /* Store the correct base and size */
484     TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
485     TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
486     TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
487     return TRUE;
488   }
489 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
490               KMP_OS_HURD */
491   /* Use incremental refinement starting from initial conservative estimate */
492   TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
493   TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
494   TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
495   return FALSE;
496 }
497 
498 static void *__kmp_launch_worker(void *thr) {
499   int status, old_type, old_state;
500 #ifdef KMP_BLOCK_SIGNALS
501   sigset_t new_set, old_set;
502 #endif /* KMP_BLOCK_SIGNALS */
503   void *exit_val;
504 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
505         KMP_OS_OPENBSD || KMP_OS_HURD
506   void *volatile padding = 0;
507 #endif
508   int gtid;
509 
510   gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
511   __kmp_gtid_set_specific(gtid);
512 #ifdef KMP_TDATA_GTID
513   __kmp_gtid = gtid;
514 #endif
515 #if KMP_STATS_ENABLED
516   // set thread local index to point to thread-specific stats
517   __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
518   __kmp_stats_thread_ptr->startLife();
519   KMP_SET_THREAD_STATE(IDLE);
520   KMP_INIT_PARTITIONED_TIMERS(OMP_idle);
521 #endif
522 
523 #if USE_ITT_BUILD
524   __kmp_itt_thread_name(gtid);
525 #endif /* USE_ITT_BUILD */
526 
527 #if KMP_AFFINITY_SUPPORTED
528   __kmp_affinity_set_init_mask(gtid, FALSE);
529 #endif
530 
531 #ifdef KMP_CANCEL_THREADS
532   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
533   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
534   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
535   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
536   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
537 #endif
538 
539 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
540   // Set FP control regs to be a copy of the parallel initialization thread's.
541   __kmp_clear_x87_fpu_status_word();
542   __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
543   __kmp_load_mxcsr(&__kmp_init_mxcsr);
544 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
545 
546 #ifdef KMP_BLOCK_SIGNALS
547   status = sigfillset(&new_set);
548   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
549   status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
550   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
551 #endif /* KMP_BLOCK_SIGNALS */
552 
553 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
554         KMP_OS_OPENBSD
555   if (__kmp_stkoffset > 0 && gtid > 0) {
556     padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
557   }
558 #endif
559 
560   KMP_MB();
561   __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
562 
563   __kmp_check_stack_overlap((kmp_info_t *)thr);
564 
565   exit_val = __kmp_launch_thread((kmp_info_t *)thr);
566 
567 #ifdef KMP_BLOCK_SIGNALS
568   status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
569   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
570 #endif /* KMP_BLOCK_SIGNALS */
571 
572   return exit_val;
573 }
574 
575 #if KMP_USE_MONITOR
576 /* The monitor thread controls all of the threads in the complex */
577 
578 static void *__kmp_launch_monitor(void *thr) {
579   int status, old_type, old_state;
580 #ifdef KMP_BLOCK_SIGNALS
581   sigset_t new_set;
582 #endif /* KMP_BLOCK_SIGNALS */
583   struct timespec interval;
584   int yield_count;
585   int yield_cycles = 0;
586 
587   KMP_MB(); /* Flush all pending memory write invalidates.  */
588 
589   KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
590 
591   /* register us as the monitor thread */
592   __kmp_gtid_set_specific(KMP_GTID_MONITOR);
593 #ifdef KMP_TDATA_GTID
594   __kmp_gtid = KMP_GTID_MONITOR;
595 #endif
596 
597   KMP_MB();
598 
599 #if USE_ITT_BUILD
600   // Instruct Intel(R) Threading Tools to ignore monitor thread.
601   __kmp_itt_thread_ignore();
602 #endif /* USE_ITT_BUILD */
603 
604   __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
605                        (kmp_info_t *)thr);
606 
607   __kmp_check_stack_overlap((kmp_info_t *)thr);
608 
609 #ifdef KMP_CANCEL_THREADS
610   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
611   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
612   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
613   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
614   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
615 #endif
616 
617 #if KMP_REAL_TIME_FIX
618   // This is a potential fix which allows application with real-time scheduling
619   // policy work. However, decision about the fix is not made yet, so it is
620   // disabled by default.
621   { // Are program started with real-time scheduling policy?
622     int sched = sched_getscheduler(0);
623     if (sched == SCHED_FIFO || sched == SCHED_RR) {
624       // Yes, we are a part of real-time application. Try to increase the
625       // priority of the monitor.
626       struct sched_param param;
627       int max_priority = sched_get_priority_max(sched);
628       int rc;
629       KMP_WARNING(RealTimeSchedNotSupported);
630       sched_getparam(0, &param);
631       if (param.sched_priority < max_priority) {
632         param.sched_priority += 1;
633         rc = sched_setscheduler(0, sched, &param);
634         if (rc != 0) {
635           int error = errno;
636           kmp_msg_t err_code = KMP_ERR(error);
637           __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
638                     err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
639           if (__kmp_generate_warnings == kmp_warnings_off) {
640             __kmp_str_free(&err_code.str);
641           }
642         }
643       } else {
644         // We cannot abort here, because number of CPUs may be enough for all
645         // the threads, including the monitor thread, so application could
646         // potentially work...
647         __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
648                   KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
649                   __kmp_msg_null);
650       }
651     }
652     // AC: free thread that waits for monitor started
653     TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
654   }
655 #endif // KMP_REAL_TIME_FIX
656 
657   KMP_MB(); /* Flush all pending memory write invalidates.  */
658 
659   if (__kmp_monitor_wakeups == 1) {
660     interval.tv_sec = 1;
661     interval.tv_nsec = 0;
662   } else {
663     interval.tv_sec = 0;
664     interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
665   }
666 
667   KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
668 
669   if (__kmp_yield_cycle) {
670     __kmp_yielding_on = 0; /* Start out with yielding shut off */
671     yield_count = __kmp_yield_off_count;
672   } else {
673     __kmp_yielding_on = 1; /* Yielding is on permanently */
674   }
675 
676   while (!TCR_4(__kmp_global.g.g_done)) {
677     struct timespec now;
678     struct timeval tval;
679 
680     /*  This thread monitors the state of the system */
681 
682     KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
683 
684     status = gettimeofday(&tval, NULL);
685     KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
686     TIMEVAL_TO_TIMESPEC(&tval, &now);
687 
688     now.tv_sec += interval.tv_sec;
689     now.tv_nsec += interval.tv_nsec;
690 
691     if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
692       now.tv_sec += 1;
693       now.tv_nsec -= KMP_NSEC_PER_SEC;
694     }
695 
696     status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
697     KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
698     // AC: the monitor should not fall asleep if g_done has been set
699     if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
700       status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
701                                       &__kmp_wait_mx.m_mutex, &now);
702       if (status != 0) {
703         if (status != ETIMEDOUT && status != EINTR) {
704           KMP_SYSFAIL("pthread_cond_timedwait", status);
705         }
706       }
707     }
708     status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
709     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
710 
711     if (__kmp_yield_cycle) {
712       yield_cycles++;
713       if ((yield_cycles % yield_count) == 0) {
714         if (__kmp_yielding_on) {
715           __kmp_yielding_on = 0; /* Turn it off now */
716           yield_count = __kmp_yield_off_count;
717         } else {
718           __kmp_yielding_on = 1; /* Turn it on now */
719           yield_count = __kmp_yield_on_count;
720         }
721         yield_cycles = 0;
722       }
723     } else {
724       __kmp_yielding_on = 1;
725     }
726 
727     TCW_4(__kmp_global.g.g_time.dt.t_value,
728           TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
729 
730     KMP_MB(); /* Flush all pending memory write invalidates.  */
731   }
732 
733   KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
734 
735 #ifdef KMP_BLOCK_SIGNALS
736   status = sigfillset(&new_set);
737   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
738   status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
739   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
740 #endif /* KMP_BLOCK_SIGNALS */
741 
742   KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
743 
744   if (__kmp_global.g.g_abort != 0) {
745     /* now we need to terminate the worker threads  */
746     /* the value of t_abort is the signal we caught */
747 
748     int gtid;
749 
750     KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
751                   __kmp_global.g.g_abort));
752 
753     /* terminate the OpenMP worker threads */
754     /* TODO this is not valid for sibling threads!!
755      * the uber master might not be 0 anymore.. */
756     for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
757       __kmp_terminate_thread(gtid);
758 
759     __kmp_cleanup();
760 
761     KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
762                   __kmp_global.g.g_abort));
763 
764     if (__kmp_global.g.g_abort > 0)
765       raise(__kmp_global.g.g_abort);
766   }
767 
768   KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
769 
770   return thr;
771 }
772 #endif // KMP_USE_MONITOR
773 
774 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
775   pthread_t handle;
776   pthread_attr_t thread_attr;
777   int status;
778 
779   th->th.th_info.ds.ds_gtid = gtid;
780 
781 #if KMP_STATS_ENABLED
782   // sets up worker thread stats
783   __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
784 
785   // th->th.th_stats is used to transfer thread-specific stats-pointer to
786   // __kmp_launch_worker. So when thread is created (goes into
787   // __kmp_launch_worker) it will set its thread local pointer to
788   // th->th.th_stats
789   if (!KMP_UBER_GTID(gtid)) {
790     th->th.th_stats = __kmp_stats_list->push_back(gtid);
791   } else {
792     // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
793     // so set the th->th.th_stats field to it.
794     th->th.th_stats = __kmp_stats_thread_ptr;
795   }
796   __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
797 
798 #endif // KMP_STATS_ENABLED
799 
800   if (KMP_UBER_GTID(gtid)) {
801     KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
802     th->th.th_info.ds.ds_thread = pthread_self();
803     __kmp_set_stack_info(gtid, th);
804     __kmp_check_stack_overlap(th);
805     return;
806   }
807 
808   KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
809 
810   KMP_MB(); /* Flush all pending memory write invalidates.  */
811 
812 #ifdef KMP_THREAD_ATTR
813   status = pthread_attr_init(&thread_attr);
814   if (status != 0) {
815     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
816   }
817   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
818   if (status != 0) {
819     __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
820   }
821 
822   /* Set stack size for this thread now.
823      The multiple of 2 is there because on some machines, requesting an unusual
824      stacksize causes the thread to have an offset before the dummy alloca()
825      takes place to create the offset.  Since we want the user to have a
826      sufficient stacksize AND support a stack offset, we alloca() twice the
827      offset so that the upcoming alloca() does not eliminate any premade offset,
828      and also gives the user the stack space they requested for all threads */
829   stack_size += gtid * __kmp_stkoffset * 2;
830 
831   KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
832                 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
833                 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
834 
835 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
836   status = pthread_attr_setstacksize(&thread_attr, stack_size);
837 #ifdef KMP_BACKUP_STKSIZE
838   if (status != 0) {
839     if (!__kmp_env_stksize) {
840       stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
841       __kmp_stksize = KMP_BACKUP_STKSIZE;
842       KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
843                     "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
844                     "bytes\n",
845                     gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
846       status = pthread_attr_setstacksize(&thread_attr, stack_size);
847     }
848   }
849 #endif /* KMP_BACKUP_STKSIZE */
850   if (status != 0) {
851     __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
852                 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
853   }
854 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
855 
856 #endif /* KMP_THREAD_ATTR */
857 
858   status =
859       pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
860   if (status != 0 || !handle) { // ??? Why do we check handle??
861 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
862     if (status == EINVAL) {
863       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
864                   KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
865     }
866     if (status == ENOMEM) {
867       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
868                   KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
869     }
870 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
871     if (status == EAGAIN) {
872       __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
873                   KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
874     }
875     KMP_SYSFAIL("pthread_create", status);
876   }
877 
878   th->th.th_info.ds.ds_thread = handle;
879 
880 #ifdef KMP_THREAD_ATTR
881   status = pthread_attr_destroy(&thread_attr);
882   if (status) {
883     kmp_msg_t err_code = KMP_ERR(status);
884     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
885               __kmp_msg_null);
886     if (__kmp_generate_warnings == kmp_warnings_off) {
887       __kmp_str_free(&err_code.str);
888     }
889   }
890 #endif /* KMP_THREAD_ATTR */
891 
892   KMP_MB(); /* Flush all pending memory write invalidates.  */
893 
894   KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
895 
896 } // __kmp_create_worker
897 
898 #if KMP_USE_MONITOR
899 void __kmp_create_monitor(kmp_info_t *th) {
900   pthread_t handle;
901   pthread_attr_t thread_attr;
902   size_t size;
903   int status;
904   int auto_adj_size = FALSE;
905 
906   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
907     // We don't need monitor thread in case of MAX_BLOCKTIME
908     KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
909                   "MAX blocktime\n"));
910     th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
911     th->th.th_info.ds.ds_gtid = 0;
912     return;
913   }
914   KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
915 
916   KMP_MB(); /* Flush all pending memory write invalidates.  */
917 
918   th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
919   th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
920 #if KMP_REAL_TIME_FIX
921   TCW_4(__kmp_global.g.g_time.dt.t_value,
922         -1); // Will use it for synchronization a bit later.
923 #else
924   TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
925 #endif // KMP_REAL_TIME_FIX
926 
927 #ifdef KMP_THREAD_ATTR
928   if (__kmp_monitor_stksize == 0) {
929     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
930     auto_adj_size = TRUE;
931   }
932   status = pthread_attr_init(&thread_attr);
933   if (status != 0) {
934     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
935   }
936   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
937   if (status != 0) {
938     __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
939   }
940 
941 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
942   status = pthread_attr_getstacksize(&thread_attr, &size);
943   KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
944 #else
945   size = __kmp_sys_min_stksize;
946 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
947 #endif /* KMP_THREAD_ATTR */
948 
949   if (__kmp_monitor_stksize == 0) {
950     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
951   }
952   if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
953     __kmp_monitor_stksize = __kmp_sys_min_stksize;
954   }
955 
956   KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
957                 "requested stacksize = %lu bytes\n",
958                 size, __kmp_monitor_stksize));
959 
960 retry:
961 
962 /* Set stack size for this thread now. */
963 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
964   KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
965                 __kmp_monitor_stksize));
966   status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
967   if (status != 0) {
968     if (auto_adj_size) {
969       __kmp_monitor_stksize *= 2;
970       goto retry;
971     }
972     kmp_msg_t err_code = KMP_ERR(status);
973     __kmp_msg(kmp_ms_warning, // should this be fatal?  BB
974               KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
975               err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
976     if (__kmp_generate_warnings == kmp_warnings_off) {
977       __kmp_str_free(&err_code.str);
978     }
979   }
980 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
981 
982   status =
983       pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
984 
985   if (status != 0) {
986 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
987     if (status == EINVAL) {
988       if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
989         __kmp_monitor_stksize *= 2;
990         goto retry;
991       }
992       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
993                   KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
994                   __kmp_msg_null);
995     }
996     if (status == ENOMEM) {
997       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
998                   KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
999                   __kmp_msg_null);
1000     }
1001 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
1002     if (status == EAGAIN) {
1003       __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
1004                   KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
1005     }
1006     KMP_SYSFAIL("pthread_create", status);
1007   }
1008 
1009   th->th.th_info.ds.ds_thread = handle;
1010 
1011 #if KMP_REAL_TIME_FIX
1012   // Wait for the monitor thread is really started and set its *priority*.
1013   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
1014                    sizeof(__kmp_global.g.g_time.dt.t_value));
1015   __kmp_wait_yield_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value,
1016                      -1, &__kmp_neq_4, NULL);
1017 #endif // KMP_REAL_TIME_FIX
1018 
1019 #ifdef KMP_THREAD_ATTR
1020   status = pthread_attr_destroy(&thread_attr);
1021   if (status != 0) {
1022     kmp_msg_t err_code = KMP_ERR(status);
1023     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
1024               __kmp_msg_null);
1025     if (__kmp_generate_warnings == kmp_warnings_off) {
1026       __kmp_str_free(&err_code.str);
1027     }
1028   }
1029 #endif
1030 
1031   KMP_MB(); /* Flush all pending memory write invalidates.  */
1032 
1033   KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
1034                 th->th.th_info.ds.ds_thread));
1035 
1036 } // __kmp_create_monitor
1037 #endif // KMP_USE_MONITOR
1038 
1039 void __kmp_exit_thread(int exit_status) {
1040   pthread_exit((void *)(intptr_t)exit_status);
1041 } // __kmp_exit_thread
1042 
1043 #if KMP_USE_MONITOR
1044 void __kmp_resume_monitor();
1045 
1046 void __kmp_reap_monitor(kmp_info_t *th) {
1047   int status;
1048   void *exit_val;
1049 
1050   KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
1051                 " %#.8lx\n",
1052                 th->th.th_info.ds.ds_thread));
1053 
1054   // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1055   // If both tid and gtid are 0, it means the monitor did not ever start.
1056   // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1057   KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1058   if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1059     KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1060     return;
1061   }
1062 
1063   KMP_MB(); /* Flush all pending memory write invalidates.  */
1064 
1065   /* First, check to see whether the monitor thread exists to wake it up. This
1066      is to avoid performance problem when the monitor sleeps during
1067      blocktime-size interval */
1068 
1069   status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1070   if (status != ESRCH) {
1071     __kmp_resume_monitor(); // Wake up the monitor thread
1072   }
1073   KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1074   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1075   if (exit_val != th) {
1076     __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1077   }
1078 
1079   th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1080   th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1081 
1082   KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1083                 " %#.8lx\n",
1084                 th->th.th_info.ds.ds_thread));
1085 
1086   KMP_MB(); /* Flush all pending memory write invalidates.  */
1087 }
1088 #endif // KMP_USE_MONITOR
1089 
1090 void __kmp_reap_worker(kmp_info_t *th) {
1091   int status;
1092   void *exit_val;
1093 
1094   KMP_MB(); /* Flush all pending memory write invalidates.  */
1095 
1096   KA_TRACE(
1097       10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1098 
1099   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1100 #ifdef KMP_DEBUG
1101   /* Don't expose these to the user until we understand when they trigger */
1102   if (status != 0) {
1103     __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1104   }
1105   if (exit_val != th) {
1106     KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1107                   "exit_val = %p\n",
1108                   th->th.th_info.ds.ds_gtid, exit_val));
1109   }
1110 #endif /* KMP_DEBUG */
1111 
1112   KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1113                 th->th.th_info.ds.ds_gtid));
1114 
1115   KMP_MB(); /* Flush all pending memory write invalidates.  */
1116 }
1117 
1118 #if KMP_HANDLE_SIGNALS
1119 
1120 static void __kmp_null_handler(int signo) {
1121   //  Do nothing, for doing SIG_IGN-type actions.
1122 } // __kmp_null_handler
1123 
1124 static void __kmp_team_handler(int signo) {
1125   if (__kmp_global.g.g_abort == 0) {
1126 /* Stage 1 signal handler, let's shut down all of the threads */
1127 #ifdef KMP_DEBUG
1128     __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1129 #endif
1130     switch (signo) {
1131     case SIGHUP:
1132     case SIGINT:
1133     case SIGQUIT:
1134     case SIGILL:
1135     case SIGABRT:
1136     case SIGFPE:
1137     case SIGBUS:
1138     case SIGSEGV:
1139 #ifdef SIGSYS
1140     case SIGSYS:
1141 #endif
1142     case SIGTERM:
1143       if (__kmp_debug_buf) {
1144         __kmp_dump_debug_buffer();
1145       }
1146       KMP_MB(); // Flush all pending memory write invalidates.
1147       TCW_4(__kmp_global.g.g_abort, signo);
1148       KMP_MB(); // Flush all pending memory write invalidates.
1149       TCW_4(__kmp_global.g.g_done, TRUE);
1150       KMP_MB(); // Flush all pending memory write invalidates.
1151       break;
1152     default:
1153 #ifdef KMP_DEBUG
1154       __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1155 #endif
1156       break;
1157     }
1158   }
1159 } // __kmp_team_handler
1160 
1161 static void __kmp_sigaction(int signum, const struct sigaction *act,
1162                             struct sigaction *oldact) {
1163   int rc = sigaction(signum, act, oldact);
1164   KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1165 }
1166 
1167 static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1168                                       int parallel_init) {
1169   KMP_MB(); // Flush all pending memory write invalidates.
1170   KB_TRACE(60,
1171            ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1172   if (parallel_init) {
1173     struct sigaction new_action;
1174     struct sigaction old_action;
1175     new_action.sa_handler = handler_func;
1176     new_action.sa_flags = 0;
1177     sigfillset(&new_action.sa_mask);
1178     __kmp_sigaction(sig, &new_action, &old_action);
1179     if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1180       sigaddset(&__kmp_sigset, sig);
1181     } else {
1182       // Restore/keep user's handler if one previously installed.
1183       __kmp_sigaction(sig, &old_action, NULL);
1184     }
1185   } else {
1186     // Save initial/system signal handlers to see if user handlers installed.
1187     __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1188   }
1189   KMP_MB(); // Flush all pending memory write invalidates.
1190 } // __kmp_install_one_handler
1191 
1192 static void __kmp_remove_one_handler(int sig) {
1193   KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1194   if (sigismember(&__kmp_sigset, sig)) {
1195     struct sigaction old;
1196     KMP_MB(); // Flush all pending memory write invalidates.
1197     __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1198     if ((old.sa_handler != __kmp_team_handler) &&
1199         (old.sa_handler != __kmp_null_handler)) {
1200       // Restore the users signal handler.
1201       KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1202                     "restoring: sig=%d\n",
1203                     sig));
1204       __kmp_sigaction(sig, &old, NULL);
1205     }
1206     sigdelset(&__kmp_sigset, sig);
1207     KMP_MB(); // Flush all pending memory write invalidates.
1208   }
1209 } // __kmp_remove_one_handler
1210 
1211 void __kmp_install_signals(int parallel_init) {
1212   KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1213   if (__kmp_handle_signals || !parallel_init) {
1214     // If ! parallel_init, we do not install handlers, just save original
1215     // handlers. Let us do it even __handle_signals is 0.
1216     sigemptyset(&__kmp_sigset);
1217     __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1218     __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1219     __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1220     __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1221     __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1222     __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1223     __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1224     __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1225 #ifdef SIGSYS
1226     __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1227 #endif // SIGSYS
1228     __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1229 #ifdef SIGPIPE
1230     __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1231 #endif // SIGPIPE
1232   }
1233 } // __kmp_install_signals
1234 
1235 void __kmp_remove_signals(void) {
1236   int sig;
1237   KB_TRACE(10, ("__kmp_remove_signals()\n"));
1238   for (sig = 1; sig < NSIG; ++sig) {
1239     __kmp_remove_one_handler(sig);
1240   }
1241 } // __kmp_remove_signals
1242 
1243 #endif // KMP_HANDLE_SIGNALS
1244 
1245 void __kmp_enable(int new_state) {
1246 #ifdef KMP_CANCEL_THREADS
1247   int status, old_state;
1248   status = pthread_setcancelstate(new_state, &old_state);
1249   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1250   KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1251 #endif
1252 }
1253 
1254 void __kmp_disable(int *old_state) {
1255 #ifdef KMP_CANCEL_THREADS
1256   int status;
1257   status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1258   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1259 #endif
1260 }
1261 
1262 static void __kmp_atfork_prepare(void) {
1263   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1264   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1265 }
1266 
1267 static void __kmp_atfork_parent(void) {
1268   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1269   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1270 }
1271 
1272 /* Reset the library so execution in the child starts "all over again" with
1273    clean data structures in initial states.  Don't worry about freeing memory
1274    allocated by parent, just abandon it to be safe. */
1275 static void __kmp_atfork_child(void) {
1276   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1277   /* TODO make sure this is done right for nested/sibling */
1278   // ATT:  Memory leaks are here? TODO: Check it and fix.
1279   /* KMP_ASSERT( 0 ); */
1280 
1281   ++__kmp_fork_count;
1282 
1283 #if KMP_AFFINITY_SUPPORTED
1284 #if KMP_OS_LINUX
1285   // reset the affinity in the child to the initial thread
1286   // affinity in the parent
1287   kmp_set_thread_affinity_mask_initial();
1288 #endif
1289   // Set default not to bind threads tightly in the child (we’re expecting
1290   // over-subscription after the fork and this can improve things for
1291   // scripting languages that use OpenMP inside process-parallel code).
1292   __kmp_affinity_type = affinity_none;
1293 #if OMP_40_ENABLED
1294   if (__kmp_nested_proc_bind.bind_types != NULL) {
1295     __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1296   }
1297 #endif // OMP_40_ENABLED
1298 #endif // KMP_AFFINITY_SUPPORTED
1299 
1300   __kmp_init_runtime = FALSE;
1301 #if KMP_USE_MONITOR
1302   __kmp_init_monitor = 0;
1303 #endif
1304   __kmp_init_parallel = FALSE;
1305   __kmp_init_middle = FALSE;
1306   __kmp_init_serial = FALSE;
1307   TCW_4(__kmp_init_gtid, FALSE);
1308   __kmp_init_common = FALSE;
1309 
1310   TCW_4(__kmp_init_user_locks, FALSE);
1311 #if !KMP_USE_DYNAMIC_LOCK
1312   __kmp_user_lock_table.used = 1;
1313   __kmp_user_lock_table.allocated = 0;
1314   __kmp_user_lock_table.table = NULL;
1315   __kmp_lock_blocks = NULL;
1316 #endif
1317 
1318   __kmp_all_nth = 0;
1319   TCW_4(__kmp_nth, 0);
1320 
1321   __kmp_thread_pool = NULL;
1322   __kmp_thread_pool_insert_pt = NULL;
1323   __kmp_team_pool = NULL;
1324 
1325   /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1326      here so threadprivate doesn't use stale data */
1327   KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1328                 __kmp_threadpriv_cache_list));
1329 
1330   while (__kmp_threadpriv_cache_list != NULL) {
1331 
1332     if (*__kmp_threadpriv_cache_list->addr != NULL) {
1333       KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1334                     &(*__kmp_threadpriv_cache_list->addr)));
1335 
1336       *__kmp_threadpriv_cache_list->addr = NULL;
1337     }
1338     __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1339   }
1340 
1341   __kmp_init_runtime = FALSE;
1342 
1343   /* reset statically initialized locks */
1344   __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1345   __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1346   __kmp_init_bootstrap_lock(&__kmp_console_lock);
1347   __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1348 
1349 #if USE_ITT_BUILD
1350   __kmp_itt_reset(); // reset ITT's global state
1351 #endif /* USE_ITT_BUILD */
1352 
1353   /* This is necessary to make sure no stale data is left around */
1354   /* AC: customers complain that we use unsafe routines in the atfork
1355      handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1356      in dynamic_link when check the presence of shared tbbmalloc library.
1357      Suggestion is to make the library initialization lazier, similar
1358      to what done for __kmpc_begin(). */
1359   // TODO: synchronize all static initializations with regular library
1360   //       startup; look at kmp_global.cpp and etc.
1361   //__kmp_internal_begin ();
1362 }
1363 
1364 void __kmp_register_atfork(void) {
1365   if (__kmp_need_register_atfork) {
1366     int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1367                                 __kmp_atfork_child);
1368     KMP_CHECK_SYSFAIL("pthread_atfork", status);
1369     __kmp_need_register_atfork = FALSE;
1370   }
1371 }
1372 
1373 void __kmp_suspend_initialize(void) {
1374   int status;
1375   status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1376   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1377   status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1378   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1379 }
1380 
1381 static void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1382   ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count);
1383   if (th->th.th_suspend_init_count <= __kmp_fork_count) {
1384     /* this means we haven't initialized the suspension pthread objects for this
1385        thread in this instance of the process */
1386     int status;
1387     status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1388                                &__kmp_suspend_cond_attr);
1389     KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1390     status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1391                                 &__kmp_suspend_mutex_attr);
1392     KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1393     *(volatile int *)&th->th.th_suspend_init_count = __kmp_fork_count + 1;
1394     ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count);
1395   }
1396 }
1397 
1398 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1399   if (th->th.th_suspend_init_count > __kmp_fork_count) {
1400     /* this means we have initialize the suspension pthread objects for this
1401        thread in this instance of the process */
1402     int status;
1403 
1404     status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1405     if (status != 0 && status != EBUSY) {
1406       KMP_SYSFAIL("pthread_cond_destroy", status);
1407     }
1408     status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1409     if (status != 0 && status != EBUSY) {
1410       KMP_SYSFAIL("pthread_mutex_destroy", status);
1411     }
1412     --th->th.th_suspend_init_count;
1413     KMP_DEBUG_ASSERT(th->th.th_suspend_init_count == __kmp_fork_count);
1414   }
1415 }
1416 
1417 // return true if lock obtained, false otherwise
1418 int __kmp_try_suspend_mx(kmp_info_t *th) {
1419   return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1420 }
1421 
1422 void __kmp_lock_suspend_mx(kmp_info_t *th) {
1423   int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1424   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1425 }
1426 
1427 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1428   int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1429   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1430 }
1431 
1432 /* This routine puts the calling thread to sleep after setting the
1433    sleep bit for the indicated flag variable to true. */
1434 template <class C>
1435 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1436   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1437   kmp_info_t *th = __kmp_threads[th_gtid];
1438   int status;
1439   typename C::flag_t old_spin;
1440 
1441   KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1442                 flag->get()));
1443 
1444   __kmp_suspend_initialize_thread(th);
1445 
1446   status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1447   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1448 
1449   KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1450                 th_gtid, flag->get()));
1451 
1452   /* TODO: shouldn't this use release semantics to ensure that
1453      __kmp_suspend_initialize_thread gets called first? */
1454   old_spin = flag->set_sleeping();
1455 #if OMP_50_ENABLED
1456   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1457       __kmp_pause_status != kmp_soft_paused) {
1458     flag->unset_sleeping();
1459     status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1460     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1461     return;
1462   }
1463 #endif
1464   KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1465                " was %x\n",
1466                th_gtid, flag->get(), flag->load(), old_spin));
1467 
1468   if (flag->done_check_val(old_spin)) {
1469     old_spin = flag->unset_sleeping();
1470     KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1471                  "for spin(%p)\n",
1472                  th_gtid, flag->get()));
1473   } else {
1474     /* Encapsulate in a loop as the documentation states that this may
1475        "with low probability" return when the condition variable has
1476        not been signaled or broadcast */
1477     int deactivated = FALSE;
1478     TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1479 
1480     while (flag->is_sleeping()) {
1481 #ifdef DEBUG_SUSPEND
1482       char buffer[128];
1483       __kmp_suspend_count++;
1484       __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1485       __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1486                    buffer);
1487 #endif
1488       // Mark the thread as no longer active (only in the first iteration of the
1489       // loop).
1490       if (!deactivated) {
1491         th->th.th_active = FALSE;
1492         if (th->th.th_active_in_pool) {
1493           th->th.th_active_in_pool = FALSE;
1494           KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1495           KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1496         }
1497         deactivated = TRUE;
1498       }
1499 
1500 #if USE_SUSPEND_TIMEOUT
1501       struct timespec now;
1502       struct timeval tval;
1503       int msecs;
1504 
1505       status = gettimeofday(&tval, NULL);
1506       KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1507       TIMEVAL_TO_TIMESPEC(&tval, &now);
1508 
1509       msecs = (4 * __kmp_dflt_blocktime) + 200;
1510       now.tv_sec += msecs / 1000;
1511       now.tv_nsec += (msecs % 1000) * 1000;
1512 
1513       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1514                     "pthread_cond_timedwait\n",
1515                     th_gtid));
1516       status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1517                                       &th->th.th_suspend_mx.m_mutex, &now);
1518 #else
1519       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1520                     " pthread_cond_wait\n",
1521                     th_gtid));
1522       status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1523                                  &th->th.th_suspend_mx.m_mutex);
1524 #endif
1525 
1526       if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1527         KMP_SYSFAIL("pthread_cond_wait", status);
1528       }
1529 #ifdef KMP_DEBUG
1530       if (status == ETIMEDOUT) {
1531         if (flag->is_sleeping()) {
1532           KF_TRACE(100,
1533                    ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1534         } else {
1535           KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1536                        "not set!\n",
1537                        th_gtid));
1538         }
1539       } else if (flag->is_sleeping()) {
1540         KF_TRACE(100,
1541                  ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1542       }
1543 #endif
1544     } // while
1545 
1546     // Mark the thread as active again (if it was previous marked as inactive)
1547     if (deactivated) {
1548       th->th.th_active = TRUE;
1549       if (TCR_4(th->th.th_in_pool)) {
1550         KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1551         th->th.th_active_in_pool = TRUE;
1552       }
1553     }
1554   }
1555 #ifdef DEBUG_SUSPEND
1556   {
1557     char buffer[128];
1558     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1559     __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1560                  buffer);
1561   }
1562 #endif
1563 
1564   status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1565   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1566   KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1567 }
1568 
1569 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) {
1570   __kmp_suspend_template(th_gtid, flag);
1571 }
1572 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) {
1573   __kmp_suspend_template(th_gtid, flag);
1574 }
1575 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1576   __kmp_suspend_template(th_gtid, flag);
1577 }
1578 
1579 /* This routine signals the thread specified by target_gtid to wake up
1580    after setting the sleep bit indicated by the flag argument to FALSE.
1581    The target thread must already have called __kmp_suspend_template() */
1582 template <class C>
1583 static inline void __kmp_resume_template(int target_gtid, C *flag) {
1584   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1585   kmp_info_t *th = __kmp_threads[target_gtid];
1586   int status;
1587 
1588 #ifdef KMP_DEBUG
1589   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1590 #endif
1591 
1592   KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1593                 gtid, target_gtid));
1594   KMP_DEBUG_ASSERT(gtid != target_gtid);
1595 
1596   __kmp_suspend_initialize_thread(th);
1597 
1598   status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1599   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1600 
1601   if (!flag) { // coming from __kmp_null_resume_wrapper
1602     flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1603   }
1604 
1605   // First, check if the flag is null or its type has changed. If so, someone
1606   // else woke it up.
1607   if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
1608     // simply shows what
1609     // flag was cast to
1610     KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1611                  "awake: flag(%p)\n",
1612                  gtid, target_gtid, NULL));
1613     status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1614     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1615     return;
1616   } else { // if multiple threads are sleeping, flag should be internally
1617     // referring to a specific thread here
1618     typename C::flag_t old_spin = flag->unset_sleeping();
1619     if (!flag->is_sleeping_val(old_spin)) {
1620       KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1621                    "awake: flag(%p): "
1622                    "%u => %u\n",
1623                    gtid, target_gtid, flag->get(), old_spin, flag->load()));
1624       status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1625       KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1626       return;
1627     }
1628     KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1629                  "sleep bit for flag's loc(%p): "
1630                  "%u => %u\n",
1631                  gtid, target_gtid, flag->get(), old_spin, flag->load()));
1632   }
1633   TCW_PTR(th->th.th_sleep_loc, NULL);
1634 
1635 #ifdef DEBUG_SUSPEND
1636   {
1637     char buffer[128];
1638     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1639     __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1640                  target_gtid, buffer);
1641   }
1642 #endif
1643   status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1644   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1645   status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1646   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1647   KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1648                 " for T#%d\n",
1649                 gtid, target_gtid));
1650 }
1651 
1652 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) {
1653   __kmp_resume_template(target_gtid, flag);
1654 }
1655 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) {
1656   __kmp_resume_template(target_gtid, flag);
1657 }
1658 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1659   __kmp_resume_template(target_gtid, flag);
1660 }
1661 
1662 #if KMP_USE_MONITOR
1663 void __kmp_resume_monitor() {
1664   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1665   int status;
1666 #ifdef KMP_DEBUG
1667   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1668   KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1669                 KMP_GTID_MONITOR));
1670   KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1671 #endif
1672   status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1673   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1674 #ifdef DEBUG_SUSPEND
1675   {
1676     char buffer[128];
1677     __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1678     __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1679                  KMP_GTID_MONITOR, buffer);
1680   }
1681 #endif
1682   status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1683   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1684   status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1685   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1686   KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1687                 " for T#%d\n",
1688                 gtid, KMP_GTID_MONITOR));
1689 }
1690 #endif // KMP_USE_MONITOR
1691 
1692 void __kmp_yield(int cond) {
1693   if (!cond)
1694     return;
1695 #if KMP_USE_MONITOR
1696   if (!__kmp_yielding_on)
1697     return;
1698 #else
1699   if (__kmp_yield_cycle && !KMP_YIELD_NOW())
1700     return;
1701 #endif
1702   sched_yield();
1703 }
1704 
1705 void __kmp_gtid_set_specific(int gtid) {
1706   if (__kmp_init_gtid) {
1707     int status;
1708     status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1709                                  (void *)(intptr_t)(gtid + 1));
1710     KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1711   } else {
1712     KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1713   }
1714 }
1715 
1716 int __kmp_gtid_get_specific() {
1717   int gtid;
1718   if (!__kmp_init_gtid) {
1719     KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1720                   "KMP_GTID_SHUTDOWN\n"));
1721     return KMP_GTID_SHUTDOWN;
1722   }
1723   gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1724   if (gtid == 0) {
1725     gtid = KMP_GTID_DNE;
1726   } else {
1727     gtid--;
1728   }
1729   KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1730                 __kmp_gtid_threadprivate_key, gtid));
1731   return gtid;
1732 }
1733 
1734 double __kmp_read_cpu_time(void) {
1735   /*clock_t   t;*/
1736   struct tms buffer;
1737 
1738   /*t =*/times(&buffer);
1739 
1740   return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC;
1741 }
1742 
1743 int __kmp_read_system_info(struct kmp_sys_info *info) {
1744   int status;
1745   struct rusage r_usage;
1746 
1747   memset(info, 0, sizeof(*info));
1748 
1749   status = getrusage(RUSAGE_SELF, &r_usage);
1750   KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1751 
1752   // The maximum resident set size utilized (in kilobytes)
1753   info->maxrss = r_usage.ru_maxrss;
1754   // The number of page faults serviced without any I/O
1755   info->minflt = r_usage.ru_minflt;
1756   // The number of page faults serviced that required I/O
1757   info->majflt = r_usage.ru_majflt;
1758   // The number of times a process was "swapped" out of memory
1759   info->nswap = r_usage.ru_nswap;
1760   // The number of times the file system had to perform input
1761   info->inblock = r_usage.ru_inblock;
1762   // The number of times the file system had to perform output
1763   info->oublock = r_usage.ru_oublock;
1764   // The number of times a context switch was voluntarily
1765   info->nvcsw = r_usage.ru_nvcsw;
1766   // The number of times a context switch was forced
1767   info->nivcsw = r_usage.ru_nivcsw;
1768 
1769   return (status != 0);
1770 }
1771 
1772 void __kmp_read_system_time(double *delta) {
1773   double t_ns;
1774   struct timeval tval;
1775   struct timespec stop;
1776   int status;
1777 
1778   status = gettimeofday(&tval, NULL);
1779   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1780   TIMEVAL_TO_TIMESPEC(&tval, &stop);
1781   t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start);
1782   *delta = (t_ns * 1e-9);
1783 }
1784 
1785 void __kmp_clear_system_time(void) {
1786   struct timeval tval;
1787   int status;
1788   status = gettimeofday(&tval, NULL);
1789   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1790   TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1791 }
1792 
1793 static int __kmp_get_xproc(void) {
1794 
1795   int r = 0;
1796 
1797 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
1798         KMP_OS_OPENBSD || KMP_OS_HURD
1799 
1800   r = sysconf(_SC_NPROCESSORS_ONLN);
1801 
1802 #elif KMP_OS_DARWIN
1803 
1804   // Bug C77011 High "OpenMP Threads and number of active cores".
1805 
1806   // Find the number of available CPUs.
1807   kern_return_t rc;
1808   host_basic_info_data_t info;
1809   mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1810   rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1811   if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1812     // Cannot use KA_TRACE() here because this code works before trace support
1813     // is initialized.
1814     r = info.avail_cpus;
1815   } else {
1816     KMP_WARNING(CantGetNumAvailCPU);
1817     KMP_INFORM(AssumedNumCPU);
1818   }
1819 
1820 #else
1821 
1822 #error "Unknown or unsupported OS."
1823 
1824 #endif
1825 
1826   return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1827 
1828 } // __kmp_get_xproc
1829 
1830 int __kmp_read_from_file(char const *path, char const *format, ...) {
1831   int result;
1832   va_list args;
1833 
1834   va_start(args, format);
1835   FILE *f = fopen(path, "rb");
1836   if (f == NULL)
1837     return 0;
1838   result = vfscanf(f, format, args);
1839   fclose(f);
1840 
1841   return result;
1842 }
1843 
1844 void __kmp_runtime_initialize(void) {
1845   int status;
1846   pthread_mutexattr_t mutex_attr;
1847   pthread_condattr_t cond_attr;
1848 
1849   if (__kmp_init_runtime) {
1850     return;
1851   }
1852 
1853 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1854   if (!__kmp_cpuinfo.initialized) {
1855     __kmp_query_cpuid(&__kmp_cpuinfo);
1856   }
1857 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1858 
1859   __kmp_xproc = __kmp_get_xproc();
1860 
1861   if (sysconf(_SC_THREADS)) {
1862 
1863     /* Query the maximum number of threads */
1864     __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX);
1865     if (__kmp_sys_max_nth == -1) {
1866       /* Unlimited threads for NPTL */
1867       __kmp_sys_max_nth = INT_MAX;
1868     } else if (__kmp_sys_max_nth <= 1) {
1869       /* Can't tell, just use PTHREAD_THREADS_MAX */
1870       __kmp_sys_max_nth = KMP_MAX_NTH;
1871     }
1872 
1873     /* Query the minimum stack size */
1874     __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1875     if (__kmp_sys_min_stksize <= 1) {
1876       __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1877     }
1878   }
1879 
1880   /* Set up minimum number of threads to switch to TLS gtid */
1881   __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1882 
1883   status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1884                               __kmp_internal_end_dest);
1885   KMP_CHECK_SYSFAIL("pthread_key_create", status);
1886   status = pthread_mutexattr_init(&mutex_attr);
1887   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1888   status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1889   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1890   status = pthread_condattr_init(&cond_attr);
1891   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1892   status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1893   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1894 #if USE_ITT_BUILD
1895   __kmp_itt_initialize();
1896 #endif /* USE_ITT_BUILD */
1897 
1898   __kmp_init_runtime = TRUE;
1899 }
1900 
1901 void __kmp_runtime_destroy(void) {
1902   int status;
1903 
1904   if (!__kmp_init_runtime) {
1905     return; // Nothing to do.
1906   }
1907 
1908 #if USE_ITT_BUILD
1909   __kmp_itt_destroy();
1910 #endif /* USE_ITT_BUILD */
1911 
1912   status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1913   KMP_CHECK_SYSFAIL("pthread_key_delete", status);
1914 
1915   status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1916   if (status != 0 && status != EBUSY) {
1917     KMP_SYSFAIL("pthread_mutex_destroy", status);
1918   }
1919   status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1920   if (status != 0 && status != EBUSY) {
1921     KMP_SYSFAIL("pthread_cond_destroy", status);
1922   }
1923 #if KMP_AFFINITY_SUPPORTED
1924   __kmp_affinity_uninitialize();
1925 #endif
1926 
1927   __kmp_init_runtime = FALSE;
1928 }
1929 
1930 /* Put the thread to sleep for a time period */
1931 /* NOTE: not currently used anywhere */
1932 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
1933 
1934 /* Calculate the elapsed wall clock time for the user */
1935 void __kmp_elapsed(double *t) {
1936   int status;
1937 #ifdef FIX_SGI_CLOCK
1938   struct timespec ts;
1939 
1940   status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1941   KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
1942   *t =
1943       (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
1944 #else
1945   struct timeval tv;
1946 
1947   status = gettimeofday(&tv, NULL);
1948   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1949   *t =
1950       (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
1951 #endif
1952 }
1953 
1954 /* Calculate the elapsed wall clock tick for the user */
1955 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1956 
1957 /* Return the current time stamp in nsec */
1958 kmp_uint64 __kmp_now_nsec() {
1959   struct timeval t;
1960   gettimeofday(&t, NULL);
1961   kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
1962                     (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
1963   return nsec;
1964 }
1965 
1966 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1967 /* Measure clock ticks per millisecond */
1968 void __kmp_initialize_system_tick() {
1969   kmp_uint64 now, nsec2, diff;
1970   kmp_uint64 delay = 100000; // 50~100 usec on most machines.
1971   kmp_uint64 nsec = __kmp_now_nsec();
1972   kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
1973   while ((now = __kmp_hardware_timestamp()) < goal)
1974     ;
1975   nsec2 = __kmp_now_nsec();
1976   diff = nsec2 - nsec;
1977   if (diff > 0) {
1978     kmp_uint64 tpms = (kmp_uint64)(1e6 * (delay + (now - goal)) / diff);
1979     if (tpms > 0)
1980       __kmp_ticks_per_msec = tpms;
1981   }
1982 }
1983 #endif
1984 
1985 /* Determine whether the given address is mapped into the current address
1986    space. */
1987 
1988 int __kmp_is_address_mapped(void *addr) {
1989 
1990   int found = 0;
1991   int rc;
1992 
1993 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_HURD
1994 
1995   /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the address
1996      ranges mapped into the address space. */
1997 
1998   char *name = __kmp_str_format("/proc/%d/maps", getpid());
1999   FILE *file = NULL;
2000 
2001   file = fopen(name, "r");
2002   KMP_ASSERT(file != NULL);
2003 
2004   for (;;) {
2005 
2006     void *beginning = NULL;
2007     void *ending = NULL;
2008     char perms[5];
2009 
2010     rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
2011     if (rc == EOF) {
2012       break;
2013     }
2014     KMP_ASSERT(rc == 3 &&
2015                KMP_STRLEN(perms) == 4); // Make sure all fields are read.
2016 
2017     // Ending address is not included in the region, but beginning is.
2018     if ((addr >= beginning) && (addr < ending)) {
2019       perms[2] = 0; // 3th and 4th character does not matter.
2020       if (strcmp(perms, "rw") == 0) {
2021         // Memory we are looking for should be readable and writable.
2022         found = 1;
2023       }
2024       break;
2025     }
2026   }
2027 
2028   // Free resources.
2029   fclose(file);
2030   KMP_INTERNAL_FREE(name);
2031 
2032 #elif KMP_OS_DARWIN
2033 
2034   /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2035      using vm interface. */
2036 
2037   int buffer;
2038   vm_size_t count;
2039   rc = vm_read_overwrite(
2040       mach_task_self(), // Task to read memory of.
2041       (vm_address_t)(addr), // Address to read from.
2042       1, // Number of bytes to be read.
2043       (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2044       &count // Address of var to save number of read bytes in.
2045       );
2046   if (rc == 0) {
2047     // Memory successfully read.
2048     found = 1;
2049   }
2050 
2051 #elif KMP_OS_NETBSD
2052 
2053   int mib[5];
2054   mib[0] = CTL_VM;
2055   mib[1] = VM_PROC;
2056   mib[2] = VM_PROC_MAP;
2057   mib[3] = getpid();
2058   mib[4] = sizeof(struct kinfo_vmentry);
2059 
2060   size_t size;
2061   rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2062   KMP_ASSERT(!rc);
2063   KMP_ASSERT(size);
2064 
2065   size = size * 4 / 3;
2066   struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2067   KMP_ASSERT(kiv);
2068 
2069   rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2070   KMP_ASSERT(!rc);
2071   KMP_ASSERT(size);
2072 
2073   for (size_t i = 0; i < size; i++) {
2074     if (kiv[i].kve_start >= (uint64_t)addr &&
2075         kiv[i].kve_end <= (uint64_t)addr) {
2076       found = 1;
2077       break;
2078     }
2079   }
2080   KMP_INTERNAL_FREE(kiv);
2081 #elif KMP_OS_DRAGONFLY || KMP_OS_OPENBSD
2082 
2083   // FIXME(DragonFly, OpenBSD): Implement this
2084   found = 1;
2085 
2086 #else
2087 
2088 #error "Unknown or unsupported OS"
2089 
2090 #endif
2091 
2092   return found;
2093 
2094 } // __kmp_is_address_mapped
2095 
2096 #ifdef USE_LOAD_BALANCE
2097 
2098 #if KMP_OS_DARWIN || KMP_OS_NETBSD
2099 
2100 // The function returns the rounded value of the system load average
2101 // during given time interval which depends on the value of
2102 // __kmp_load_balance_interval variable (default is 60 sec, other values
2103 // may be 300 sec or 900 sec).
2104 // It returns -1 in case of error.
2105 int __kmp_get_load_balance(int max) {
2106   double averages[3];
2107   int ret_avg = 0;
2108 
2109   int res = getloadavg(averages, 3);
2110 
2111   // Check __kmp_load_balance_interval to determine which of averages to use.
2112   // getloadavg() may return the number of samples less than requested that is
2113   // less than 3.
2114   if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2115     ret_avg = averages[0]; // 1 min
2116   } else if ((__kmp_load_balance_interval >= 180 &&
2117               __kmp_load_balance_interval < 600) &&
2118              (res >= 2)) {
2119     ret_avg = averages[1]; // 5 min
2120   } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2121     ret_avg = averages[2]; // 15 min
2122   } else { // Error occurred
2123     return -1;
2124   }
2125 
2126   return ret_avg;
2127 }
2128 
2129 #else // Linux* OS
2130 
2131 // The fuction returns number of running (not sleeping) threads, or -1 in case
2132 // of error. Error could be reported if Linux* OS kernel too old (without
2133 // "/proc" support). Counting running threads stops if max running threads
2134 // encountered.
2135 int __kmp_get_load_balance(int max) {
2136   static int permanent_error = 0;
2137   static int glb_running_threads = 0; // Saved count of the running threads for
2138   // the thread balance algortihm
2139   static double glb_call_time = 0; /* Thread balance algorithm call time */
2140 
2141   int running_threads = 0; // Number of running threads in the system.
2142 
2143   DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2144   struct dirent *proc_entry = NULL;
2145 
2146   kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2147   DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2148   struct dirent *task_entry = NULL;
2149   int task_path_fixed_len;
2150 
2151   kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2152   int stat_file = -1;
2153   int stat_path_fixed_len;
2154 
2155   int total_processes = 0; // Total number of processes in system.
2156   int total_threads = 0; // Total number of threads in system.
2157 
2158   double call_time = 0.0;
2159 
2160   __kmp_str_buf_init(&task_path);
2161   __kmp_str_buf_init(&stat_path);
2162 
2163   __kmp_elapsed(&call_time);
2164 
2165   if (glb_call_time &&
2166       (call_time - glb_call_time < __kmp_load_balance_interval)) {
2167     running_threads = glb_running_threads;
2168     goto finish;
2169   }
2170 
2171   glb_call_time = call_time;
2172 
2173   // Do not spend time on scanning "/proc/" if we have a permanent error.
2174   if (permanent_error) {
2175     running_threads = -1;
2176     goto finish;
2177   }
2178 
2179   if (max <= 0) {
2180     max = INT_MAX;
2181   }
2182 
2183   // Open "/proc/" directory.
2184   proc_dir = opendir("/proc");
2185   if (proc_dir == NULL) {
2186     // Cannot open "/prroc/". Probably the kernel does not support it. Return an
2187     // error now and in subsequent calls.
2188     running_threads = -1;
2189     permanent_error = 1;
2190     goto finish;
2191   }
2192 
2193   // Initialize fixed part of task_path. This part will not change.
2194   __kmp_str_buf_cat(&task_path, "/proc/", 6);
2195   task_path_fixed_len = task_path.used; // Remember number of used characters.
2196 
2197   proc_entry = readdir(proc_dir);
2198   while (proc_entry != NULL) {
2199     // Proc entry is a directory and name starts with a digit. Assume it is a
2200     // process' directory.
2201     if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2202 
2203       ++total_processes;
2204       // Make sure init process is the very first in "/proc", so we can replace
2205       // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2206       // 1. We are going to check that total_processes == 1 => d_name == "1" is
2207       // true (where "=>" is implication). Since C++ does not have => operator,
2208       // let us replace it with its equivalent: a => b == ! a || b.
2209       KMP_DEBUG_ASSERT(total_processes != 1 ||
2210                        strcmp(proc_entry->d_name, "1") == 0);
2211 
2212       // Construct task_path.
2213       task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2214       __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2215                         KMP_STRLEN(proc_entry->d_name));
2216       __kmp_str_buf_cat(&task_path, "/task", 5);
2217 
2218       task_dir = opendir(task_path.str);
2219       if (task_dir == NULL) {
2220         // Process can finish between reading "/proc/" directory entry and
2221         // opening process' "task/" directory. So, in general case we should not
2222         // complain, but have to skip this process and read the next one. But on
2223         // systems with no "task/" support we will spend lot of time to scan
2224         // "/proc/" tree again and again without any benefit. "init" process
2225         // (its pid is 1) should exist always, so, if we cannot open
2226         // "/proc/1/task/" directory, it means "task/" is not supported by
2227         // kernel. Report an error now and in the future.
2228         if (strcmp(proc_entry->d_name, "1") == 0) {
2229           running_threads = -1;
2230           permanent_error = 1;
2231           goto finish;
2232         }
2233       } else {
2234         // Construct fixed part of stat file path.
2235         __kmp_str_buf_clear(&stat_path);
2236         __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2237         __kmp_str_buf_cat(&stat_path, "/", 1);
2238         stat_path_fixed_len = stat_path.used;
2239 
2240         task_entry = readdir(task_dir);
2241         while (task_entry != NULL) {
2242           // It is a directory and name starts with a digit.
2243           if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2244             ++total_threads;
2245 
2246             // Consruct complete stat file path. Easiest way would be:
2247             //  __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2248             //  task_entry->d_name );
2249             // but seriae of __kmp_str_buf_cat works a bit faster.
2250             stat_path.used =
2251                 stat_path_fixed_len; // Reset stat path to its fixed part.
2252             __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2253                               KMP_STRLEN(task_entry->d_name));
2254             __kmp_str_buf_cat(&stat_path, "/stat", 5);
2255 
2256             // Note: Low-level API (open/read/close) is used. High-level API
2257             // (fopen/fclose)  works ~ 30 % slower.
2258             stat_file = open(stat_path.str, O_RDONLY);
2259             if (stat_file == -1) {
2260               // We cannot report an error because task (thread) can terminate
2261               // just before reading this file.
2262             } else {
2263               /* Content of "stat" file looks like:
2264                  24285 (program) S ...
2265 
2266                  It is a single line (if program name does not include funny
2267                  symbols). First number is a thread id, then name of executable
2268                  file name in paretheses, then state of the thread. We need just
2269                  thread state.
2270 
2271                  Good news: Length of program name is 15 characters max. Longer
2272                  names are truncated.
2273 
2274                  Thus, we need rather short buffer: 15 chars for program name +
2275                  2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2276 
2277                  Bad news: Program name may contain special symbols like space,
2278                  closing parenthesis, or even new line. This makes parsing
2279                  "stat" file not 100 % reliable. In case of fanny program names
2280                  parsing may fail (report incorrect thread state).
2281 
2282                  Parsing "status" file looks more promissing (due to different
2283                  file structure and escaping special symbols) but reading and
2284                  parsing of "status" file works slower.
2285                   -- ln
2286               */
2287               char buffer[65];
2288               int len;
2289               len = read(stat_file, buffer, sizeof(buffer) - 1);
2290               if (len >= 0) {
2291                 buffer[len] = 0;
2292                 // Using scanf:
2293                 //     sscanf( buffer, "%*d (%*s) %c ", & state );
2294                 // looks very nice, but searching for a closing parenthesis
2295                 // works a bit faster.
2296                 char *close_parent = strstr(buffer, ") ");
2297                 if (close_parent != NULL) {
2298                   char state = *(close_parent + 2);
2299                   if (state == 'R') {
2300                     ++running_threads;
2301                     if (running_threads >= max) {
2302                       goto finish;
2303                     }
2304                   }
2305                 }
2306               }
2307               close(stat_file);
2308               stat_file = -1;
2309             }
2310           }
2311           task_entry = readdir(task_dir);
2312         }
2313         closedir(task_dir);
2314         task_dir = NULL;
2315       }
2316     }
2317     proc_entry = readdir(proc_dir);
2318   }
2319 
2320   // There _might_ be a timing hole where the thread executing this
2321   // code get skipped in the load balance, and running_threads is 0.
2322   // Assert in the debug builds only!!!
2323   KMP_DEBUG_ASSERT(running_threads > 0);
2324   if (running_threads <= 0) {
2325     running_threads = 1;
2326   }
2327 
2328 finish: // Clean up and exit.
2329   if (proc_dir != NULL) {
2330     closedir(proc_dir);
2331   }
2332   __kmp_str_buf_free(&task_path);
2333   if (task_dir != NULL) {
2334     closedir(task_dir);
2335   }
2336   __kmp_str_buf_free(&stat_path);
2337   if (stat_file != -1) {
2338     close(stat_file);
2339   }
2340 
2341   glb_running_threads = running_threads;
2342 
2343   return running_threads;
2344 
2345 } // __kmp_get_load_balance
2346 
2347 #endif // KMP_OS_DARWIN
2348 
2349 #endif // USE_LOAD_BALANCE
2350 
2351 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC ||                            \
2352       ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || KMP_ARCH_PPC64)
2353 
2354 // we really only need the case with 1 argument, because CLANG always build
2355 // a struct of pointers to shared variables referenced in the outlined function
2356 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2357                            void *p_argv[]
2358 #if OMPT_SUPPORT
2359                            ,
2360                            void **exit_frame_ptr
2361 #endif
2362                            ) {
2363 #if OMPT_SUPPORT
2364   *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2365 #endif
2366 
2367   switch (argc) {
2368   default:
2369     fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2370     fflush(stderr);
2371     exit(-1);
2372   case 0:
2373     (*pkfn)(&gtid, &tid);
2374     break;
2375   case 1:
2376     (*pkfn)(&gtid, &tid, p_argv[0]);
2377     break;
2378   case 2:
2379     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1]);
2380     break;
2381   case 3:
2382     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2]);
2383     break;
2384   case 4:
2385     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2386     break;
2387   case 5:
2388     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2389     break;
2390   case 6:
2391     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2392             p_argv[5]);
2393     break;
2394   case 7:
2395     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2396             p_argv[5], p_argv[6]);
2397     break;
2398   case 8:
2399     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2400             p_argv[5], p_argv[6], p_argv[7]);
2401     break;
2402   case 9:
2403     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2404             p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2405     break;
2406   case 10:
2407     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2408             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2409     break;
2410   case 11:
2411     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2412             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2413     break;
2414   case 12:
2415     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2416             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2417             p_argv[11]);
2418     break;
2419   case 13:
2420     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2421             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2422             p_argv[11], p_argv[12]);
2423     break;
2424   case 14:
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], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2427             p_argv[11], p_argv[12], p_argv[13]);
2428     break;
2429   case 15:
2430     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2431             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2432             p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2433     break;
2434   }
2435 
2436 #if OMPT_SUPPORT
2437   *exit_frame_ptr = 0;
2438 #endif
2439 
2440   return 1;
2441 }
2442 
2443 #endif
2444 
2445 // end of file //
2446