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