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_FREEBSD && !KMP_OS_NETBSD
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_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_FREEBSD || KMP_OS_NETBSD
448   /* Linux* OS only -- no pthread_getattr_np support on OS X* */
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_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_FREEBSD || KMP_OS_NETBSD */
487   /* Use incremental refinement starting from initial conservative estimate */
488   TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
489   TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
490   TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
491   return FALSE;
492 }
493 
494 static void *__kmp_launch_worker(void *thr) {
495   int status, old_type, old_state;
496 #ifdef KMP_BLOCK_SIGNALS
497   sigset_t new_set, old_set;
498 #endif /* KMP_BLOCK_SIGNALS */
499   void *exit_val;
500 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
501   void *volatile padding = 0;
502 #endif
503   int gtid;
504 
505   gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
506   __kmp_gtid_set_specific(gtid);
507 #ifdef KMP_TDATA_GTID
508   __kmp_gtid = gtid;
509 #endif
510 #if KMP_STATS_ENABLED
511   // set thread local index to point to thread-specific stats
512   __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
513   KMP_START_EXPLICIT_TIMER(OMP_worker_thread_life);
514   KMP_SET_THREAD_STATE(IDLE);
515   KMP_INIT_PARTITIONED_TIMERS(OMP_idle);
516 #endif
517 
518 #if USE_ITT_BUILD
519   __kmp_itt_thread_name(gtid);
520 #endif /* USE_ITT_BUILD */
521 
522 #if KMP_AFFINITY_SUPPORTED
523   __kmp_affinity_set_init_mask(gtid, FALSE);
524 #endif
525 
526 #ifdef KMP_CANCEL_THREADS
527   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
528   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
529   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
530   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
531   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
532 #endif
533 
534 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
535   // Set FP control regs to be a copy of the parallel initialization thread's.
536   __kmp_clear_x87_fpu_status_word();
537   __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
538   __kmp_load_mxcsr(&__kmp_init_mxcsr);
539 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
540 
541 #ifdef KMP_BLOCK_SIGNALS
542   status = sigfillset(&new_set);
543   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
544   status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
545   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
546 #endif /* KMP_BLOCK_SIGNALS */
547 
548 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
549   if (__kmp_stkoffset > 0 && gtid > 0) {
550     padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
551   }
552 #endif
553 
554   KMP_MB();
555   __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
556 
557   __kmp_check_stack_overlap((kmp_info_t *)thr);
558 
559   exit_val = __kmp_launch_thread((kmp_info_t *)thr);
560 
561 #ifdef KMP_BLOCK_SIGNALS
562   status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
563   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
564 #endif /* KMP_BLOCK_SIGNALS */
565 
566   return exit_val;
567 }
568 
569 #if KMP_USE_MONITOR
570 /* The monitor thread controls all of the threads in the complex */
571 
572 static void *__kmp_launch_monitor(void *thr) {
573   int status, old_type, old_state;
574 #ifdef KMP_BLOCK_SIGNALS
575   sigset_t new_set;
576 #endif /* KMP_BLOCK_SIGNALS */
577   struct timespec interval;
578   int yield_count;
579   int yield_cycles = 0;
580 
581   KMP_MB(); /* Flush all pending memory write invalidates.  */
582 
583   KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
584 
585   /* register us as the monitor thread */
586   __kmp_gtid_set_specific(KMP_GTID_MONITOR);
587 #ifdef KMP_TDATA_GTID
588   __kmp_gtid = KMP_GTID_MONITOR;
589 #endif
590 
591   KMP_MB();
592 
593 #if USE_ITT_BUILD
594   // Instruct Intel(R) Threading Tools to ignore monitor thread.
595   __kmp_itt_thread_ignore();
596 #endif /* USE_ITT_BUILD */
597 
598   __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
599                        (kmp_info_t *)thr);
600 
601   __kmp_check_stack_overlap((kmp_info_t *)thr);
602 
603 #ifdef KMP_CANCEL_THREADS
604   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
605   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
606   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
607   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
608   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
609 #endif
610 
611 #if KMP_REAL_TIME_FIX
612   // This is a potential fix which allows application with real-time scheduling
613   // policy work. However, decision about the fix is not made yet, so it is
614   // disabled by default.
615   { // Are program started with real-time scheduling policy?
616     int sched = sched_getscheduler(0);
617     if (sched == SCHED_FIFO || sched == SCHED_RR) {
618       // Yes, we are a part of real-time application. Try to increase the
619       // priority of the monitor.
620       struct sched_param param;
621       int max_priority = sched_get_priority_max(sched);
622       int rc;
623       KMP_WARNING(RealTimeSchedNotSupported);
624       sched_getparam(0, &param);
625       if (param.sched_priority < max_priority) {
626         param.sched_priority += 1;
627         rc = sched_setscheduler(0, sched, &param);
628         if (rc != 0) {
629           int error = errno;
630           kmp_msg_t err_code = KMP_ERR(error);
631           __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
632                     err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
633           if (__kmp_generate_warnings == kmp_warnings_off) {
634             __kmp_str_free(&err_code.str);
635           }
636         }
637       } else {
638         // We cannot abort here, because number of CPUs may be enough for all
639         // the threads, including the monitor thread, so application could
640         // potentially work...
641         __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
642                   KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
643                   __kmp_msg_null);
644       }
645     }
646     // AC: free thread that waits for monitor started
647     TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
648   }
649 #endif // KMP_REAL_TIME_FIX
650 
651   KMP_MB(); /* Flush all pending memory write invalidates.  */
652 
653   if (__kmp_monitor_wakeups == 1) {
654     interval.tv_sec = 1;
655     interval.tv_nsec = 0;
656   } else {
657     interval.tv_sec = 0;
658     interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
659   }
660 
661   KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
662 
663   if (__kmp_yield_cycle) {
664     __kmp_yielding_on = 0; /* Start out with yielding shut off */
665     yield_count = __kmp_yield_off_count;
666   } else {
667     __kmp_yielding_on = 1; /* Yielding is on permanently */
668   }
669 
670   while (!TCR_4(__kmp_global.g.g_done)) {
671     struct timespec now;
672     struct timeval tval;
673 
674     /*  This thread monitors the state of the system */
675 
676     KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
677 
678     status = gettimeofday(&tval, NULL);
679     KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
680     TIMEVAL_TO_TIMESPEC(&tval, &now);
681 
682     now.tv_sec += interval.tv_sec;
683     now.tv_nsec += interval.tv_nsec;
684 
685     if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
686       now.tv_sec += 1;
687       now.tv_nsec -= KMP_NSEC_PER_SEC;
688     }
689 
690     status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
691     KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
692     // AC: the monitor should not fall asleep if g_done has been set
693     if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
694       status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
695                                       &__kmp_wait_mx.m_mutex, &now);
696       if (status != 0) {
697         if (status != ETIMEDOUT && status != EINTR) {
698           KMP_SYSFAIL("pthread_cond_timedwait", status);
699         }
700       }
701     }
702     status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
703     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
704 
705     if (__kmp_yield_cycle) {
706       yield_cycles++;
707       if ((yield_cycles % yield_count) == 0) {
708         if (__kmp_yielding_on) {
709           __kmp_yielding_on = 0; /* Turn it off now */
710           yield_count = __kmp_yield_off_count;
711         } else {
712           __kmp_yielding_on = 1; /* Turn it on now */
713           yield_count = __kmp_yield_on_count;
714         }
715         yield_cycles = 0;
716       }
717     } else {
718       __kmp_yielding_on = 1;
719     }
720 
721     TCW_4(__kmp_global.g.g_time.dt.t_value,
722           TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
723 
724     KMP_MB(); /* Flush all pending memory write invalidates.  */
725   }
726 
727   KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
728 
729 #ifdef KMP_BLOCK_SIGNALS
730   status = sigfillset(&new_set);
731   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
732   status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
733   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
734 #endif /* KMP_BLOCK_SIGNALS */
735 
736   KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
737 
738   if (__kmp_global.g.g_abort != 0) {
739     /* now we need to terminate the worker threads  */
740     /* the value of t_abort is the signal we caught */
741 
742     int gtid;
743 
744     KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
745                   __kmp_global.g.g_abort));
746 
747     /* terminate the OpenMP worker threads */
748     /* TODO this is not valid for sibling threads!!
749      * the uber master might not be 0 anymore.. */
750     for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
751       __kmp_terminate_thread(gtid);
752 
753     __kmp_cleanup();
754 
755     KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
756                   __kmp_global.g.g_abort));
757 
758     if (__kmp_global.g.g_abort > 0)
759       raise(__kmp_global.g.g_abort);
760   }
761 
762   KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
763 
764   return thr;
765 }
766 #endif // KMP_USE_MONITOR
767 
768 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
769   pthread_t handle;
770   pthread_attr_t thread_attr;
771   int status;
772 
773   th->th.th_info.ds.ds_gtid = gtid;
774 
775 #if KMP_STATS_ENABLED
776   // sets up worker thread stats
777   __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
778 
779   // th->th.th_stats is used to transfer thread-specific stats-pointer to
780   // __kmp_launch_worker. So when thread is created (goes into
781   // __kmp_launch_worker) it will set its thread local pointer to
782   // th->th.th_stats
783   if (!KMP_UBER_GTID(gtid)) {
784     th->th.th_stats = __kmp_stats_list->push_back(gtid);
785   } else {
786     // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
787     // so set the th->th.th_stats field to it.
788     th->th.th_stats = __kmp_stats_thread_ptr;
789   }
790   __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
791 
792 #endif // KMP_STATS_ENABLED
793 
794   if (KMP_UBER_GTID(gtid)) {
795     KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
796     th->th.th_info.ds.ds_thread = pthread_self();
797     __kmp_set_stack_info(gtid, th);
798     __kmp_check_stack_overlap(th);
799     return;
800   }
801 
802   KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
803 
804   KMP_MB(); /* Flush all pending memory write invalidates.  */
805 
806 #ifdef KMP_THREAD_ATTR
807   status = pthread_attr_init(&thread_attr);
808   if (status != 0) {
809     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
810   }
811   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
812   if (status != 0) {
813     __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
814   }
815 
816   /* Set stack size for this thread now.
817      The multiple of 2 is there because on some machines, requesting an unusual
818      stacksize causes the thread to have an offset before the dummy alloca()
819      takes place to create the offset.  Since we want the user to have a
820      sufficient stacksize AND support a stack offset, we alloca() twice the
821      offset so that the upcoming alloca() does not eliminate any premade offset,
822      and also gives the user the stack space they requested for all threads */
823   stack_size += gtid * __kmp_stkoffset * 2;
824 
825   KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
826                 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
827                 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
828 
829 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
830   status = pthread_attr_setstacksize(&thread_attr, stack_size);
831 #ifdef KMP_BACKUP_STKSIZE
832   if (status != 0) {
833     if (!__kmp_env_stksize) {
834       stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
835       __kmp_stksize = KMP_BACKUP_STKSIZE;
836       KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
837                     "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
838                     "bytes\n",
839                     gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
840       status = pthread_attr_setstacksize(&thread_attr, stack_size);
841     }
842   }
843 #endif /* KMP_BACKUP_STKSIZE */
844   if (status != 0) {
845     __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
846                 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
847   }
848 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
849 
850 #endif /* KMP_THREAD_ATTR */
851 
852   status =
853       pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
854   if (status != 0 || !handle) { // ??? Why do we check handle??
855 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
856     if (status == EINVAL) {
857       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
858                   KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
859     }
860     if (status == ENOMEM) {
861       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
862                   KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
863     }
864 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
865     if (status == EAGAIN) {
866       __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
867                   KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
868     }
869     KMP_SYSFAIL("pthread_create", status);
870   }
871 
872   th->th.th_info.ds.ds_thread = handle;
873 
874 #ifdef KMP_THREAD_ATTR
875   status = pthread_attr_destroy(&thread_attr);
876   if (status) {
877     kmp_msg_t err_code = KMP_ERR(status);
878     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
879               __kmp_msg_null);
880     if (__kmp_generate_warnings == kmp_warnings_off) {
881       __kmp_str_free(&err_code.str);
882     }
883   }
884 #endif /* KMP_THREAD_ATTR */
885 
886   KMP_MB(); /* Flush all pending memory write invalidates.  */
887 
888   KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
889 
890 } // __kmp_create_worker
891 
892 #if KMP_USE_MONITOR
893 void __kmp_create_monitor(kmp_info_t *th) {
894   pthread_t handle;
895   pthread_attr_t thread_attr;
896   size_t size;
897   int status;
898   int auto_adj_size = FALSE;
899 
900   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
901     // We don't need monitor thread in case of MAX_BLOCKTIME
902     KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
903                   "MAX blocktime\n"));
904     th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
905     th->th.th_info.ds.ds_gtid = 0;
906     return;
907   }
908   KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
909 
910   KMP_MB(); /* Flush all pending memory write invalidates.  */
911 
912   th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
913   th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
914 #if KMP_REAL_TIME_FIX
915   TCW_4(__kmp_global.g.g_time.dt.t_value,
916         -1); // Will use it for synchronization a bit later.
917 #else
918   TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
919 #endif // KMP_REAL_TIME_FIX
920 
921 #ifdef KMP_THREAD_ATTR
922   if (__kmp_monitor_stksize == 0) {
923     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
924     auto_adj_size = TRUE;
925   }
926   status = pthread_attr_init(&thread_attr);
927   if (status != 0) {
928     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
929   }
930   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
931   if (status != 0) {
932     __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
933   }
934 
935 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
936   status = pthread_attr_getstacksize(&thread_attr, &size);
937   KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
938 #else
939   size = __kmp_sys_min_stksize;
940 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
941 #endif /* KMP_THREAD_ATTR */
942 
943   if (__kmp_monitor_stksize == 0) {
944     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
945   }
946   if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
947     __kmp_monitor_stksize = __kmp_sys_min_stksize;
948   }
949 
950   KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
951                 "requested stacksize = %lu bytes\n",
952                 size, __kmp_monitor_stksize));
953 
954 retry:
955 
956 /* Set stack size for this thread now. */
957 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
958   KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
959                 __kmp_monitor_stksize));
960   status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
961   if (status != 0) {
962     if (auto_adj_size) {
963       __kmp_monitor_stksize *= 2;
964       goto retry;
965     }
966     kmp_msg_t err_code = KMP_ERR(status);
967     __kmp_msg(kmp_ms_warning, // should this be fatal?  BB
968               KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
969               err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
970     if (__kmp_generate_warnings == kmp_warnings_off) {
971       __kmp_str_free(&err_code.str);
972     }
973   }
974 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
975 
976   status =
977       pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
978 
979   if (status != 0) {
980 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
981     if (status == EINVAL) {
982       if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
983         __kmp_monitor_stksize *= 2;
984         goto retry;
985       }
986       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
987                   KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
988                   __kmp_msg_null);
989     }
990     if (status == ENOMEM) {
991       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
992                   KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
993                   __kmp_msg_null);
994     }
995 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
996     if (status == EAGAIN) {
997       __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
998                   KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
999     }
1000     KMP_SYSFAIL("pthread_create", status);
1001   }
1002 
1003   th->th.th_info.ds.ds_thread = handle;
1004 
1005 #if KMP_REAL_TIME_FIX
1006   // Wait for the monitor thread is really started and set its *priority*.
1007   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
1008                    sizeof(__kmp_global.g.g_time.dt.t_value));
1009   __kmp_wait_yield_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value,
1010                      -1, &__kmp_neq_4, NULL);
1011 #endif // KMP_REAL_TIME_FIX
1012 
1013 #ifdef KMP_THREAD_ATTR
1014   status = pthread_attr_destroy(&thread_attr);
1015   if (status != 0) {
1016     kmp_msg_t err_code = KMP_ERR(status);
1017     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
1018               __kmp_msg_null);
1019     if (__kmp_generate_warnings == kmp_warnings_off) {
1020       __kmp_str_free(&err_code.str);
1021     }
1022   }
1023 #endif
1024 
1025   KMP_MB(); /* Flush all pending memory write invalidates.  */
1026 
1027   KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
1028                 th->th.th_info.ds.ds_thread));
1029 
1030 } // __kmp_create_monitor
1031 #endif // KMP_USE_MONITOR
1032 
1033 void __kmp_exit_thread(int exit_status) {
1034   pthread_exit((void *)(intptr_t)exit_status);
1035 } // __kmp_exit_thread
1036 
1037 #if KMP_USE_MONITOR
1038 void __kmp_resume_monitor();
1039 
1040 void __kmp_reap_monitor(kmp_info_t *th) {
1041   int status;
1042   void *exit_val;
1043 
1044   KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
1045                 " %#.8lx\n",
1046                 th->th.th_info.ds.ds_thread));
1047 
1048   // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1049   // If both tid and gtid are 0, it means the monitor did not ever start.
1050   // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1051   KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1052   if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1053     KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1054     return;
1055   }
1056 
1057   KMP_MB(); /* Flush all pending memory write invalidates.  */
1058 
1059   /* First, check to see whether the monitor thread exists to wake it up. This
1060      is to avoid performance problem when the monitor sleeps during
1061      blocktime-size interval */
1062 
1063   status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1064   if (status != ESRCH) {
1065     __kmp_resume_monitor(); // Wake up the monitor thread
1066   }
1067   KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1068   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1069   if (exit_val != th) {
1070     __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1071   }
1072 
1073   th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1074   th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1075 
1076   KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1077                 " %#.8lx\n",
1078                 th->th.th_info.ds.ds_thread));
1079 
1080   KMP_MB(); /* Flush all pending memory write invalidates.  */
1081 }
1082 #endif // KMP_USE_MONITOR
1083 
1084 void __kmp_reap_worker(kmp_info_t *th) {
1085   int status;
1086   void *exit_val;
1087 
1088   KMP_MB(); /* Flush all pending memory write invalidates.  */
1089 
1090   KA_TRACE(
1091       10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1092 
1093   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1094 #ifdef KMP_DEBUG
1095   /* Don't expose these to the user until we understand when they trigger */
1096   if (status != 0) {
1097     __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1098   }
1099   if (exit_val != th) {
1100     KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1101                   "exit_val = %p\n",
1102                   th->th.th_info.ds.ds_gtid, exit_val));
1103   }
1104 #endif /* KMP_DEBUG */
1105 
1106   KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1107                 th->th.th_info.ds.ds_gtid));
1108 
1109   KMP_MB(); /* Flush all pending memory write invalidates.  */
1110 }
1111 
1112 #if KMP_HANDLE_SIGNALS
1113 
1114 static void __kmp_null_handler(int signo) {
1115   //  Do nothing, for doing SIG_IGN-type actions.
1116 } // __kmp_null_handler
1117 
1118 static void __kmp_team_handler(int signo) {
1119   if (__kmp_global.g.g_abort == 0) {
1120 /* Stage 1 signal handler, let's shut down all of the threads */
1121 #ifdef KMP_DEBUG
1122     __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1123 #endif
1124     switch (signo) {
1125     case SIGHUP:
1126     case SIGINT:
1127     case SIGQUIT:
1128     case SIGILL:
1129     case SIGABRT:
1130     case SIGFPE:
1131     case SIGBUS:
1132     case SIGSEGV:
1133 #ifdef SIGSYS
1134     case SIGSYS:
1135 #endif
1136     case SIGTERM:
1137       if (__kmp_debug_buf) {
1138         __kmp_dump_debug_buffer();
1139       }
1140       KMP_MB(); // Flush all pending memory write invalidates.
1141       TCW_4(__kmp_global.g.g_abort, signo);
1142       KMP_MB(); // Flush all pending memory write invalidates.
1143       TCW_4(__kmp_global.g.g_done, TRUE);
1144       KMP_MB(); // Flush all pending memory write invalidates.
1145       break;
1146     default:
1147 #ifdef KMP_DEBUG
1148       __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1149 #endif
1150       break;
1151     }
1152   }
1153 } // __kmp_team_handler
1154 
1155 static void __kmp_sigaction(int signum, const struct sigaction *act,
1156                             struct sigaction *oldact) {
1157   int rc = sigaction(signum, act, oldact);
1158   KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1159 }
1160 
1161 static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1162                                       int parallel_init) {
1163   KMP_MB(); // Flush all pending memory write invalidates.
1164   KB_TRACE(60,
1165            ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1166   if (parallel_init) {
1167     struct sigaction new_action;
1168     struct sigaction old_action;
1169     new_action.sa_handler = handler_func;
1170     new_action.sa_flags = 0;
1171     sigfillset(&new_action.sa_mask);
1172     __kmp_sigaction(sig, &new_action, &old_action);
1173     if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1174       sigaddset(&__kmp_sigset, sig);
1175     } else {
1176       // Restore/keep user's handler if one previously installed.
1177       __kmp_sigaction(sig, &old_action, NULL);
1178     }
1179   } else {
1180     // Save initial/system signal handlers to see if user handlers installed.
1181     __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1182   }
1183   KMP_MB(); // Flush all pending memory write invalidates.
1184 } // __kmp_install_one_handler
1185 
1186 static void __kmp_remove_one_handler(int sig) {
1187   KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1188   if (sigismember(&__kmp_sigset, sig)) {
1189     struct sigaction old;
1190     KMP_MB(); // Flush all pending memory write invalidates.
1191     __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1192     if ((old.sa_handler != __kmp_team_handler) &&
1193         (old.sa_handler != __kmp_null_handler)) {
1194       // Restore the users signal handler.
1195       KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1196                     "restoring: sig=%d\n",
1197                     sig));
1198       __kmp_sigaction(sig, &old, NULL);
1199     }
1200     sigdelset(&__kmp_sigset, sig);
1201     KMP_MB(); // Flush all pending memory write invalidates.
1202   }
1203 } // __kmp_remove_one_handler
1204 
1205 void __kmp_install_signals(int parallel_init) {
1206   KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1207   if (__kmp_handle_signals || !parallel_init) {
1208     // If ! parallel_init, we do not install handlers, just save original
1209     // handlers. Let us do it even __handle_signals is 0.
1210     sigemptyset(&__kmp_sigset);
1211     __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1212     __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1213     __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1214     __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1215     __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1216     __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1217     __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1218     __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1219 #ifdef SIGSYS
1220     __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1221 #endif // SIGSYS
1222     __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1223 #ifdef SIGPIPE
1224     __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1225 #endif // SIGPIPE
1226   }
1227 } // __kmp_install_signals
1228 
1229 void __kmp_remove_signals(void) {
1230   int sig;
1231   KB_TRACE(10, ("__kmp_remove_signals()\n"));
1232   for (sig = 1; sig < NSIG; ++sig) {
1233     __kmp_remove_one_handler(sig);
1234   }
1235 } // __kmp_remove_signals
1236 
1237 #endif // KMP_HANDLE_SIGNALS
1238 
1239 void __kmp_enable(int new_state) {
1240 #ifdef KMP_CANCEL_THREADS
1241   int status, old_state;
1242   status = pthread_setcancelstate(new_state, &old_state);
1243   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1244   KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1245 #endif
1246 }
1247 
1248 void __kmp_disable(int *old_state) {
1249 #ifdef KMP_CANCEL_THREADS
1250   int status;
1251   status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1252   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1253 #endif
1254 }
1255 
1256 static void __kmp_atfork_prepare(void) {
1257   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1258   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1259 }
1260 
1261 static void __kmp_atfork_parent(void) {
1262   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1263   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1264 }
1265 
1266 /* Reset the library so execution in the child starts "all over again" with
1267    clean data structures in initial states.  Don't worry about freeing memory
1268    allocated by parent, just abandon it to be safe. */
1269 static void __kmp_atfork_child(void) {
1270   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1271   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1272   /* TODO make sure this is done right for nested/sibling */
1273   // ATT:  Memory leaks are here? TODO: Check it and fix.
1274   /* KMP_ASSERT( 0 ); */
1275 
1276   ++__kmp_fork_count;
1277 
1278 #if KMP_AFFINITY_SUPPORTED
1279 #if KMP_OS_LINUX
1280   // reset the affinity in the child to the initial thread
1281   // affinity in the parent
1282   kmp_set_thread_affinity_mask_initial();
1283 #endif
1284   // Set default not to bind threads tightly in the child (we’re expecting
1285   // over-subscription after the fork and this can improve things for
1286   // scripting languages that use OpenMP inside process-parallel code).
1287   __kmp_affinity_type = affinity_none;
1288 #if OMP_40_ENABLED
1289   if (__kmp_nested_proc_bind.bind_types != NULL) {
1290     __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1291   }
1292 #endif // OMP_40_ENABLED
1293 #endif // KMP_AFFINITY_SUPPORTED
1294 
1295   __kmp_init_runtime = FALSE;
1296 #if KMP_USE_MONITOR
1297   __kmp_init_monitor = 0;
1298 #endif
1299   __kmp_init_parallel = FALSE;
1300   __kmp_init_middle = FALSE;
1301   __kmp_init_serial = FALSE;
1302   TCW_4(__kmp_init_gtid, FALSE);
1303   __kmp_init_common = FALSE;
1304 
1305   TCW_4(__kmp_init_user_locks, FALSE);
1306 #if !KMP_USE_DYNAMIC_LOCK
1307   __kmp_user_lock_table.used = 1;
1308   __kmp_user_lock_table.allocated = 0;
1309   __kmp_user_lock_table.table = NULL;
1310   __kmp_lock_blocks = NULL;
1311 #endif
1312 
1313   __kmp_all_nth = 0;
1314   TCW_4(__kmp_nth, 0);
1315 
1316   __kmp_thread_pool = NULL;
1317   __kmp_thread_pool_insert_pt = NULL;
1318   __kmp_team_pool = NULL;
1319 
1320   /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1321      here so threadprivate doesn't use stale data */
1322   KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1323                 __kmp_threadpriv_cache_list));
1324 
1325   while (__kmp_threadpriv_cache_list != NULL) {
1326 
1327     if (*__kmp_threadpriv_cache_list->addr != NULL) {
1328       KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1329                     &(*__kmp_threadpriv_cache_list->addr)));
1330 
1331       *__kmp_threadpriv_cache_list->addr = NULL;
1332     }
1333     __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1334   }
1335 
1336   __kmp_init_runtime = FALSE;
1337 
1338   /* reset statically initialized locks */
1339   __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1340   __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1341   __kmp_init_bootstrap_lock(&__kmp_console_lock);
1342   __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1343 
1344 #if USE_ITT_BUILD
1345   __kmp_itt_reset(); // reset ITT's global state
1346 #endif /* USE_ITT_BUILD */
1347 
1348   /* This is necessary to make sure no stale data is left around */
1349   /* AC: customers complain that we use unsafe routines in the atfork
1350      handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1351      in dynamic_link when check the presence of shared tbbmalloc library.
1352      Suggestion is to make the library initialization lazier, similar
1353      to what done for __kmpc_begin(). */
1354   // TODO: synchronize all static initializations with regular library
1355   //       startup; look at kmp_global.cpp and etc.
1356   //__kmp_internal_begin ();
1357 }
1358 
1359 void __kmp_register_atfork(void) {
1360   if (__kmp_need_register_atfork) {
1361     int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1362                                 __kmp_atfork_child);
1363     KMP_CHECK_SYSFAIL("pthread_atfork", status);
1364     __kmp_need_register_atfork = FALSE;
1365   }
1366 }
1367 
1368 void __kmp_suspend_initialize(void) {
1369   int status;
1370   status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1371   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1372   status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1373   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1374 }
1375 
1376 static void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1377   ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count);
1378   if (th->th.th_suspend_init_count <= __kmp_fork_count) {
1379     /* this means we haven't initialized the suspension pthread objects for this
1380        thread in this instance of the process */
1381     int status;
1382     status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1383                                &__kmp_suspend_cond_attr);
1384     KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1385     status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1386                                 &__kmp_suspend_mutex_attr);
1387     KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1388     *(volatile int *)&th->th.th_suspend_init_count = __kmp_fork_count + 1;
1389     ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count);
1390   }
1391 }
1392 
1393 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1394   if (th->th.th_suspend_init_count > __kmp_fork_count) {
1395     /* this means we have initialize the suspension pthread objects for this
1396        thread in this instance of the process */
1397     int status;
1398 
1399     status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1400     if (status != 0 && status != EBUSY) {
1401       KMP_SYSFAIL("pthread_cond_destroy", status);
1402     }
1403     status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1404     if (status != 0 && status != EBUSY) {
1405       KMP_SYSFAIL("pthread_mutex_destroy", status);
1406     }
1407     --th->th.th_suspend_init_count;
1408     KMP_DEBUG_ASSERT(th->th.th_suspend_init_count == __kmp_fork_count);
1409   }
1410 }
1411 
1412 /* This routine puts the calling thread to sleep after setting the
1413    sleep bit for the indicated flag variable to true. */
1414 template <class C>
1415 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1416   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1417   kmp_info_t *th = __kmp_threads[th_gtid];
1418   int status;
1419   typename C::flag_t old_spin;
1420 
1421   KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1422                 flag->get()));
1423 
1424   __kmp_suspend_initialize_thread(th);
1425 
1426   status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1427   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1428 
1429   KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1430                 th_gtid, flag->get()));
1431 
1432   /* TODO: shouldn't this use release semantics to ensure that
1433      __kmp_suspend_initialize_thread gets called first? */
1434   old_spin = flag->set_sleeping();
1435 
1436   KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1437                " was %x\n",
1438                th_gtid, flag->get(), *(flag->get()), old_spin));
1439 
1440   if (flag->done_check_val(old_spin)) {
1441     old_spin = flag->unset_sleeping();
1442     KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1443                  "for spin(%p)\n",
1444                  th_gtid, flag->get()));
1445   } else {
1446     /* Encapsulate in a loop as the documentation states that this may
1447        "with low probability" return when the condition variable has
1448        not been signaled or broadcast */
1449     int deactivated = FALSE;
1450     TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1451 
1452     while (flag->is_sleeping()) {
1453 #ifdef DEBUG_SUSPEND
1454       char buffer[128];
1455       __kmp_suspend_count++;
1456       __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1457       __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1458                    buffer);
1459 #endif
1460       // Mark the thread as no longer active (only in the first iteration of the
1461       // loop).
1462       if (!deactivated) {
1463         th->th.th_active = FALSE;
1464         if (th->th.th_active_in_pool) {
1465           th->th.th_active_in_pool = FALSE;
1466           KMP_TEST_THEN_DEC32(&__kmp_thread_pool_active_nth);
1467           KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1468         }
1469         deactivated = TRUE;
1470       }
1471 
1472 #if USE_SUSPEND_TIMEOUT
1473       struct timespec now;
1474       struct timeval tval;
1475       int msecs;
1476 
1477       status = gettimeofday(&tval, NULL);
1478       KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1479       TIMEVAL_TO_TIMESPEC(&tval, &now);
1480 
1481       msecs = (4 * __kmp_dflt_blocktime) + 200;
1482       now.tv_sec += msecs / 1000;
1483       now.tv_nsec += (msecs % 1000) * 1000;
1484 
1485       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1486                     "pthread_cond_timedwait\n",
1487                     th_gtid));
1488       status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1489                                       &th->th.th_suspend_mx.m_mutex, &now);
1490 #else
1491       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1492                     " pthread_cond_wait\n",
1493                     th_gtid));
1494       status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1495                                  &th->th.th_suspend_mx.m_mutex);
1496 #endif
1497 
1498       if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1499         KMP_SYSFAIL("pthread_cond_wait", status);
1500       }
1501 #ifdef KMP_DEBUG
1502       if (status == ETIMEDOUT) {
1503         if (flag->is_sleeping()) {
1504           KF_TRACE(100,
1505                    ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1506         } else {
1507           KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1508                        "not set!\n",
1509                        th_gtid));
1510         }
1511       } else if (flag->is_sleeping()) {
1512         KF_TRACE(100,
1513                  ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1514       }
1515 #endif
1516     } // while
1517 
1518     // Mark the thread as active again (if it was previous marked as inactive)
1519     if (deactivated) {
1520       th->th.th_active = TRUE;
1521       if (TCR_4(th->th.th_in_pool)) {
1522         KMP_TEST_THEN_INC32(&__kmp_thread_pool_active_nth);
1523         th->th.th_active_in_pool = TRUE;
1524       }
1525     }
1526   }
1527 #ifdef DEBUG_SUSPEND
1528   {
1529     char buffer[128];
1530     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1531     __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1532                  buffer);
1533   }
1534 #endif
1535 
1536   status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1537   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1538   KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1539 }
1540 
1541 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) {
1542   __kmp_suspend_template(th_gtid, flag);
1543 }
1544 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) {
1545   __kmp_suspend_template(th_gtid, flag);
1546 }
1547 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1548   __kmp_suspend_template(th_gtid, flag);
1549 }
1550 
1551 /* This routine signals the thread specified by target_gtid to wake up
1552    after setting the sleep bit indicated by the flag argument to FALSE.
1553    The target thread must already have called __kmp_suspend_template() */
1554 template <class C>
1555 static inline void __kmp_resume_template(int target_gtid, C *flag) {
1556   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1557   kmp_info_t *th = __kmp_threads[target_gtid];
1558   int status;
1559 
1560 #ifdef KMP_DEBUG
1561   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1562 #endif
1563 
1564   KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1565                 gtid, target_gtid));
1566   KMP_DEBUG_ASSERT(gtid != target_gtid);
1567 
1568   __kmp_suspend_initialize_thread(th);
1569 
1570   status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1571   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1572 
1573   if (!flag) { // coming from __kmp_null_resume_wrapper
1574     flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1575   }
1576 
1577   // First, check if the flag is null or its type has changed. If so, someone
1578   // else woke it up.
1579   if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
1580     // simply shows what
1581     // flag was cast to
1582     KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1583                  "awake: flag(%p)\n",
1584                  gtid, target_gtid, NULL));
1585     status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1586     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1587     return;
1588   } else { // if multiple threads are sleeping, flag should be internally
1589     // referring to a specific thread here
1590     typename C::flag_t old_spin = flag->unset_sleeping();
1591     if (!flag->is_sleeping_val(old_spin)) {
1592       KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1593                    "awake: flag(%p): "
1594                    "%u => %u\n",
1595                    gtid, target_gtid, flag->get(), old_spin, *flag->get()));
1596       status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1597       KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1598       return;
1599     }
1600     KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1601                  "sleep bit for flag's loc(%p): "
1602                  "%u => %u\n",
1603                  gtid, target_gtid, flag->get(), old_spin, *flag->get()));
1604   }
1605   TCW_PTR(th->th.th_sleep_loc, NULL);
1606 
1607 #ifdef DEBUG_SUSPEND
1608   {
1609     char buffer[128];
1610     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1611     __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1612                  target_gtid, buffer);
1613   }
1614 #endif
1615   status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1616   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1617   status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1618   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1619   KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1620                 " for T#%d\n",
1621                 gtid, target_gtid));
1622 }
1623 
1624 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) {
1625   __kmp_resume_template(target_gtid, flag);
1626 }
1627 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) {
1628   __kmp_resume_template(target_gtid, flag);
1629 }
1630 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1631   __kmp_resume_template(target_gtid, flag);
1632 }
1633 
1634 #if KMP_USE_MONITOR
1635 void __kmp_resume_monitor() {
1636   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1637   int status;
1638 #ifdef KMP_DEBUG
1639   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1640   KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1641                 KMP_GTID_MONITOR));
1642   KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1643 #endif
1644   status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1645   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1646 #ifdef DEBUG_SUSPEND
1647   {
1648     char buffer[128];
1649     __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1650     __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1651                  KMP_GTID_MONITOR, buffer);
1652   }
1653 #endif
1654   status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1655   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1656   status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1657   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1658   KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1659                 " for T#%d\n",
1660                 gtid, KMP_GTID_MONITOR));
1661 }
1662 #endif // KMP_USE_MONITOR
1663 
1664 void __kmp_yield(int cond) {
1665   if (!cond)
1666     return;
1667 #if KMP_USE_MONITOR
1668   if (!__kmp_yielding_on)
1669     return;
1670 #else
1671   if (__kmp_yield_cycle && !KMP_YIELD_NOW())
1672     return;
1673 #endif
1674   sched_yield();
1675 }
1676 
1677 void __kmp_gtid_set_specific(int gtid) {
1678   if (__kmp_init_gtid) {
1679     int status;
1680     status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1681                                  (void *)(intptr_t)(gtid + 1));
1682     KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1683   } else {
1684     KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1685   }
1686 }
1687 
1688 int __kmp_gtid_get_specific() {
1689   int gtid;
1690   if (!__kmp_init_gtid) {
1691     KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1692                   "KMP_GTID_SHUTDOWN\n"));
1693     return KMP_GTID_SHUTDOWN;
1694   }
1695   gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1696   if (gtid == 0) {
1697     gtid = KMP_GTID_DNE;
1698   } else {
1699     gtid--;
1700   }
1701   KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1702                 __kmp_gtid_threadprivate_key, gtid));
1703   return gtid;
1704 }
1705 
1706 double __kmp_read_cpu_time(void) {
1707   /*clock_t   t;*/
1708   struct tms buffer;
1709 
1710   /*t =*/times(&buffer);
1711 
1712   return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC;
1713 }
1714 
1715 int __kmp_read_system_info(struct kmp_sys_info *info) {
1716   int status;
1717   struct rusage r_usage;
1718 
1719   memset(info, 0, sizeof(*info));
1720 
1721   status = getrusage(RUSAGE_SELF, &r_usage);
1722   KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1723 
1724   // The maximum resident set size utilized (in kilobytes)
1725   info->maxrss = r_usage.ru_maxrss;
1726   // The number of page faults serviced without any I/O
1727   info->minflt = r_usage.ru_minflt;
1728   // The number of page faults serviced that required I/O
1729   info->majflt = r_usage.ru_majflt;
1730   // The number of times a process was "swapped" out of memory
1731   info->nswap = r_usage.ru_nswap;
1732   // The number of times the file system had to perform input
1733   info->inblock = r_usage.ru_inblock;
1734   // The number of times the file system had to perform output
1735   info->oublock = r_usage.ru_oublock;
1736   // The number of times a context switch was voluntarily
1737   info->nvcsw = r_usage.ru_nvcsw;
1738   // The number of times a context switch was forced
1739   info->nivcsw = r_usage.ru_nivcsw;
1740 
1741   return (status != 0);
1742 }
1743 
1744 void __kmp_read_system_time(double *delta) {
1745   double t_ns;
1746   struct timeval tval;
1747   struct timespec stop;
1748   int status;
1749 
1750   status = gettimeofday(&tval, NULL);
1751   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1752   TIMEVAL_TO_TIMESPEC(&tval, &stop);
1753   t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start);
1754   *delta = (t_ns * 1e-9);
1755 }
1756 
1757 void __kmp_clear_system_time(void) {
1758   struct timeval tval;
1759   int status;
1760   status = gettimeofday(&tval, NULL);
1761   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1762   TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1763 }
1764 
1765 static int __kmp_get_xproc(void) {
1766 
1767   int r = 0;
1768 
1769 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
1770 
1771   r = sysconf(_SC_NPROCESSORS_ONLN);
1772 
1773 #elif KMP_OS_DARWIN
1774 
1775   // Bug C77011 High "OpenMP Threads and number of active cores".
1776 
1777   // Find the number of available CPUs.
1778   kern_return_t rc;
1779   host_basic_info_data_t info;
1780   mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1781   rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1782   if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1783     // Cannot use KA_TRACE() here because this code works before trace support
1784     // is initialized.
1785     r = info.avail_cpus;
1786   } else {
1787     KMP_WARNING(CantGetNumAvailCPU);
1788     KMP_INFORM(AssumedNumCPU);
1789   }
1790 
1791 #else
1792 
1793 #error "Unknown or unsupported OS."
1794 
1795 #endif
1796 
1797   return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1798 
1799 } // __kmp_get_xproc
1800 
1801 int __kmp_read_from_file(char const *path, char const *format, ...) {
1802   int result;
1803   va_list args;
1804 
1805   va_start(args, format);
1806   FILE *f = fopen(path, "rb");
1807   if (f == NULL)
1808     return 0;
1809   result = vfscanf(f, format, args);
1810   fclose(f);
1811 
1812   return result;
1813 }
1814 
1815 void __kmp_runtime_initialize(void) {
1816   int status;
1817   pthread_mutexattr_t mutex_attr;
1818   pthread_condattr_t cond_attr;
1819 
1820   if (__kmp_init_runtime) {
1821     return;
1822   }
1823 
1824 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1825   if (!__kmp_cpuinfo.initialized) {
1826     __kmp_query_cpuid(&__kmp_cpuinfo);
1827   }
1828 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1829 
1830   __kmp_xproc = __kmp_get_xproc();
1831 
1832   if (sysconf(_SC_THREADS)) {
1833 
1834     /* Query the maximum number of threads */
1835     __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX);
1836     if (__kmp_sys_max_nth == -1) {
1837       /* Unlimited threads for NPTL */
1838       __kmp_sys_max_nth = INT_MAX;
1839     } else if (__kmp_sys_max_nth <= 1) {
1840       /* Can't tell, just use PTHREAD_THREADS_MAX */
1841       __kmp_sys_max_nth = KMP_MAX_NTH;
1842     }
1843 
1844     /* Query the minimum stack size */
1845     __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1846     if (__kmp_sys_min_stksize <= 1) {
1847       __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1848     }
1849   }
1850 
1851   /* Set up minimum number of threads to switch to TLS gtid */
1852   __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1853 
1854   status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1855                               __kmp_internal_end_dest);
1856   KMP_CHECK_SYSFAIL("pthread_key_create", status);
1857   status = pthread_mutexattr_init(&mutex_attr);
1858   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1859   status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1860   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1861   status = pthread_condattr_init(&cond_attr);
1862   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1863   status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1864   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1865 #if USE_ITT_BUILD
1866   __kmp_itt_initialize();
1867 #endif /* USE_ITT_BUILD */
1868 
1869   __kmp_init_runtime = TRUE;
1870 }
1871 
1872 void __kmp_runtime_destroy(void) {
1873   int status;
1874 
1875   if (!__kmp_init_runtime) {
1876     return; // Nothing to do.
1877   }
1878 
1879 #if USE_ITT_BUILD
1880   __kmp_itt_destroy();
1881 #endif /* USE_ITT_BUILD */
1882 
1883   status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1884   KMP_CHECK_SYSFAIL("pthread_key_delete", status);
1885 
1886   status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1887   if (status != 0 && status != EBUSY) {
1888     KMP_SYSFAIL("pthread_mutex_destroy", status);
1889   }
1890   status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1891   if (status != 0 && status != EBUSY) {
1892     KMP_SYSFAIL("pthread_cond_destroy", status);
1893   }
1894 #if KMP_AFFINITY_SUPPORTED
1895   __kmp_affinity_uninitialize();
1896 #endif
1897 
1898   __kmp_init_runtime = FALSE;
1899 }
1900 
1901 /* Put the thread to sleep for a time period */
1902 /* NOTE: not currently used anywhere */
1903 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
1904 
1905 /* Calculate the elapsed wall clock time for the user */
1906 void __kmp_elapsed(double *t) {
1907   int status;
1908 #ifdef FIX_SGI_CLOCK
1909   struct timespec ts;
1910 
1911   status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1912   KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
1913   *t =
1914       (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
1915 #else
1916   struct timeval tv;
1917 
1918   status = gettimeofday(&tv, NULL);
1919   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1920   *t =
1921       (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
1922 #endif
1923 }
1924 
1925 /* Calculate the elapsed wall clock tick for the user */
1926 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1927 
1928 /* Return the current time stamp in nsec */
1929 kmp_uint64 __kmp_now_nsec() {
1930   struct timeval t;
1931   gettimeofday(&t, NULL);
1932   return KMP_NSEC_PER_SEC * t.tv_sec + 1000 * t.tv_usec;
1933 }
1934 
1935 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1936 /* Measure clock ticks per millisecond */
1937 void __kmp_initialize_system_tick() {
1938   kmp_uint64 delay = 100000; // 50~100 usec on most machines.
1939   kmp_uint64 nsec = __kmp_now_nsec();
1940   kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
1941   kmp_uint64 now;
1942   while ((now = __kmp_hardware_timestamp()) < goal)
1943     ;
1944   __kmp_ticks_per_msec =
1945       (kmp_uint64)(1e6 * (delay + (now - goal)) / (__kmp_now_nsec() - nsec));
1946 }
1947 #endif
1948 
1949 /* Determine whether the given address is mapped into the current address
1950    space. */
1951 
1952 int __kmp_is_address_mapped(void *addr) {
1953 
1954   int found = 0;
1955   int rc;
1956 
1957 #if KMP_OS_LINUX || KMP_OS_FREEBSD
1958 
1959   /* On Linux* OS, read the /proc/<pid>/maps pseudo-file to get all the address
1960      ranges mapped into the address space. */
1961 
1962   char *name = __kmp_str_format("/proc/%d/maps", getpid());
1963   FILE *file = NULL;
1964 
1965   file = fopen(name, "r");
1966   KMP_ASSERT(file != NULL);
1967 
1968   for (;;) {
1969 
1970     void *beginning = NULL;
1971     void *ending = NULL;
1972     char perms[5];
1973 
1974     rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
1975     if (rc == EOF) {
1976       break;
1977     }
1978     KMP_ASSERT(rc == 3 &&
1979                KMP_STRLEN(perms) == 4); // Make sure all fields are read.
1980 
1981     // Ending address is not included in the region, but beginning is.
1982     if ((addr >= beginning) && (addr < ending)) {
1983       perms[2] = 0; // 3th and 4th character does not matter.
1984       if (strcmp(perms, "rw") == 0) {
1985         // Memory we are looking for should be readable and writable.
1986         found = 1;
1987       }
1988       break;
1989     }
1990   }
1991 
1992   // Free resources.
1993   fclose(file);
1994   KMP_INTERNAL_FREE(name);
1995 
1996 #elif KMP_OS_DARWIN
1997 
1998   /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
1999      using vm interface. */
2000 
2001   int buffer;
2002   vm_size_t count;
2003   rc = vm_read_overwrite(
2004       mach_task_self(), // Task to read memory of.
2005       (vm_address_t)(addr), // Address to read from.
2006       1, // Number of bytes to be read.
2007       (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2008       &count // Address of var to save number of read bytes in.
2009       );
2010   if (rc == 0) {
2011     // Memory successfully read.
2012     found = 1;
2013   }
2014 
2015 #elif KMP_OS_FREEBSD || KMP_OS_NETBSD
2016 
2017   // FIXME(FreeBSD, NetBSD): Implement this
2018   found = 1;
2019 
2020 #else
2021 
2022 #error "Unknown or unsupported OS"
2023 
2024 #endif
2025 
2026   return found;
2027 
2028 } // __kmp_is_address_mapped
2029 
2030 #ifdef USE_LOAD_BALANCE
2031 
2032 #if KMP_OS_DARWIN
2033 
2034 // The function returns the rounded value of the system load average
2035 // during given time interval which depends on the value of
2036 // __kmp_load_balance_interval variable (default is 60 sec, other values
2037 // may be 300 sec or 900 sec).
2038 // It returns -1 in case of error.
2039 int __kmp_get_load_balance(int max) {
2040   double averages[3];
2041   int ret_avg = 0;
2042 
2043   int res = getloadavg(averages, 3);
2044 
2045   // Check __kmp_load_balance_interval to determine which of averages to use.
2046   // getloadavg() may return the number of samples less than requested that is
2047   // less than 3.
2048   if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2049     ret_avg = averages[0]; // 1 min
2050   } else if ((__kmp_load_balance_interval >= 180 &&
2051               __kmp_load_balance_interval < 600) &&
2052              (res >= 2)) {
2053     ret_avg = averages[1]; // 5 min
2054   } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2055     ret_avg = averages[2]; // 15 min
2056   } else { // Error occurred
2057     return -1;
2058   }
2059 
2060   return ret_avg;
2061 }
2062 
2063 #else // Linux* OS
2064 
2065 // The fuction returns number of running (not sleeping) threads, or -1 in case
2066 // of error. Error could be reported if Linux* OS kernel too old (without
2067 // "/proc" support). Counting running threads stops if max running threads
2068 // encountered.
2069 int __kmp_get_load_balance(int max) {
2070   static int permanent_error = 0;
2071   static int glb_running_threads = 0; // Saved count of the running threads for
2072   // the thread balance algortihm
2073   static double glb_call_time = 0; /* Thread balance algorithm call time */
2074 
2075   int running_threads = 0; // Number of running threads in the system.
2076 
2077   DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2078   struct dirent *proc_entry = NULL;
2079 
2080   kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2081   DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2082   struct dirent *task_entry = NULL;
2083   int task_path_fixed_len;
2084 
2085   kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2086   int stat_file = -1;
2087   int stat_path_fixed_len;
2088 
2089   int total_processes = 0; // Total number of processes in system.
2090   int total_threads = 0; // Total number of threads in system.
2091 
2092   double call_time = 0.0;
2093 
2094   __kmp_str_buf_init(&task_path);
2095   __kmp_str_buf_init(&stat_path);
2096 
2097   __kmp_elapsed(&call_time);
2098 
2099   if (glb_call_time &&
2100       (call_time - glb_call_time < __kmp_load_balance_interval)) {
2101     running_threads = glb_running_threads;
2102     goto finish;
2103   }
2104 
2105   glb_call_time = call_time;
2106 
2107   // Do not spend time on scanning "/proc/" if we have a permanent error.
2108   if (permanent_error) {
2109     running_threads = -1;
2110     goto finish;
2111   }
2112 
2113   if (max <= 0) {
2114     max = INT_MAX;
2115   }
2116 
2117   // Open "/proc/" directory.
2118   proc_dir = opendir("/proc");
2119   if (proc_dir == NULL) {
2120     // Cannot open "/prroc/". Probably the kernel does not support it. Return an
2121     // error now and in subsequent calls.
2122     running_threads = -1;
2123     permanent_error = 1;
2124     goto finish;
2125   }
2126 
2127   // Initialize fixed part of task_path. This part will not change.
2128   __kmp_str_buf_cat(&task_path, "/proc/", 6);
2129   task_path_fixed_len = task_path.used; // Remember number of used characters.
2130 
2131   proc_entry = readdir(proc_dir);
2132   while (proc_entry != NULL) {
2133     // Proc entry is a directory and name starts with a digit. Assume it is a
2134     // process' directory.
2135     if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2136 
2137       ++total_processes;
2138       // Make sure init process is the very first in "/proc", so we can replace
2139       // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2140       // 1. We are going to check that total_processes == 1 => d_name == "1" is
2141       // true (where "=>" is implication). Since C++ does not have => operator,
2142       // let us replace it with its equivalent: a => b == ! a || b.
2143       KMP_DEBUG_ASSERT(total_processes != 1 ||
2144                        strcmp(proc_entry->d_name, "1") == 0);
2145 
2146       // Construct task_path.
2147       task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2148       __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2149                         KMP_STRLEN(proc_entry->d_name));
2150       __kmp_str_buf_cat(&task_path, "/task", 5);
2151 
2152       task_dir = opendir(task_path.str);
2153       if (task_dir == NULL) {
2154         // Process can finish between reading "/proc/" directory entry and
2155         // opening process' "task/" directory. So, in general case we should not
2156         // complain, but have to skip this process and read the next one. But on
2157         // systems with no "task/" support we will spend lot of time to scan
2158         // "/proc/" tree again and again without any benefit. "init" process
2159         // (its pid is 1) should exist always, so, if we cannot open
2160         // "/proc/1/task/" directory, it means "task/" is not supported by
2161         // kernel. Report an error now and in the future.
2162         if (strcmp(proc_entry->d_name, "1") == 0) {
2163           running_threads = -1;
2164           permanent_error = 1;
2165           goto finish;
2166         }
2167       } else {
2168         // Construct fixed part of stat file path.
2169         __kmp_str_buf_clear(&stat_path);
2170         __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2171         __kmp_str_buf_cat(&stat_path, "/", 1);
2172         stat_path_fixed_len = stat_path.used;
2173 
2174         task_entry = readdir(task_dir);
2175         while (task_entry != NULL) {
2176           // It is a directory and name starts with a digit.
2177           if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2178             ++total_threads;
2179 
2180             // Consruct complete stat file path. Easiest way would be:
2181             //  __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2182             //  task_entry->d_name );
2183             // but seriae of __kmp_str_buf_cat works a bit faster.
2184             stat_path.used =
2185                 stat_path_fixed_len; // Reset stat path to its fixed part.
2186             __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2187                               KMP_STRLEN(task_entry->d_name));
2188             __kmp_str_buf_cat(&stat_path, "/stat", 5);
2189 
2190             // Note: Low-level API (open/read/close) is used. High-level API
2191             // (fopen/fclose)  works ~ 30 % slower.
2192             stat_file = open(stat_path.str, O_RDONLY);
2193             if (stat_file == -1) {
2194               // We cannot report an error because task (thread) can terminate
2195               // just before reading this file.
2196             } else {
2197               /* Content of "stat" file looks like:
2198                  24285 (program) S ...
2199 
2200                  It is a single line (if program name does not include funny
2201                  symbols). First number is a thread id, then name of executable
2202                  file name in paretheses, then state of the thread. We need just
2203                  thread state.
2204 
2205                  Good news: Length of program name is 15 characters max. Longer
2206                  names are truncated.
2207 
2208                  Thus, we need rather short buffer: 15 chars for program name +
2209                  2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2210 
2211                  Bad news: Program name may contain special symbols like space,
2212                  closing parenthesis, or even new line. This makes parsing
2213                  "stat" file not 100 % reliable. In case of fanny program names
2214                  parsing may fail (report incorrect thread state).
2215 
2216                  Parsing "status" file looks more promissing (due to different
2217                  file structure and escaping special symbols) but reading and
2218                  parsing of "status" file works slower.
2219                   -- ln
2220               */
2221               char buffer[65];
2222               int len;
2223               len = read(stat_file, buffer, sizeof(buffer) - 1);
2224               if (len >= 0) {
2225                 buffer[len] = 0;
2226                 // Using scanf:
2227                 //     sscanf( buffer, "%*d (%*s) %c ", & state );
2228                 // looks very nice, but searching for a closing parenthesis
2229                 // works a bit faster.
2230                 char *close_parent = strstr(buffer, ") ");
2231                 if (close_parent != NULL) {
2232                   char state = *(close_parent + 2);
2233                   if (state == 'R') {
2234                     ++running_threads;
2235                     if (running_threads >= max) {
2236                       goto finish;
2237                     }
2238                   }
2239                 }
2240               }
2241               close(stat_file);
2242               stat_file = -1;
2243             }
2244           }
2245           task_entry = readdir(task_dir);
2246         }
2247         closedir(task_dir);
2248         task_dir = NULL;
2249       }
2250     }
2251     proc_entry = readdir(proc_dir);
2252   }
2253 
2254   // There _might_ be a timing hole where the thread executing this
2255   // code get skipped in the load balance, and running_threads is 0.
2256   // Assert in the debug builds only!!!
2257   KMP_DEBUG_ASSERT(running_threads > 0);
2258   if (running_threads <= 0) {
2259     running_threads = 1;
2260   }
2261 
2262 finish: // Clean up and exit.
2263   if (proc_dir != NULL) {
2264     closedir(proc_dir);
2265   }
2266   __kmp_str_buf_free(&task_path);
2267   if (task_dir != NULL) {
2268     closedir(task_dir);
2269   }
2270   __kmp_str_buf_free(&stat_path);
2271   if (stat_file != -1) {
2272     close(stat_file);
2273   }
2274 
2275   glb_running_threads = running_threads;
2276 
2277   return running_threads;
2278 
2279 } // __kmp_get_load_balance
2280 
2281 #endif // KMP_OS_DARWIN
2282 
2283 #endif // USE_LOAD_BALANCE
2284 
2285 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC ||                            \
2286       ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || KMP_ARCH_PPC64)
2287 
2288 // we really only need the case with 1 argument, because CLANG always build
2289 // a struct of pointers to shared variables referenced in the outlined function
2290 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2291                            void *p_argv[]
2292 #if OMPT_SUPPORT
2293                            ,
2294                            void **exit_frame_ptr
2295 #endif
2296                            ) {
2297 #if OMPT_SUPPORT
2298   *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2299 #endif
2300 
2301   switch (argc) {
2302   default:
2303     fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2304     fflush(stderr);
2305     exit(-1);
2306   case 0:
2307     (*pkfn)(&gtid, &tid);
2308     break;
2309   case 1:
2310     (*pkfn)(&gtid, &tid, p_argv[0]);
2311     break;
2312   case 2:
2313     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1]);
2314     break;
2315   case 3:
2316     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2]);
2317     break;
2318   case 4:
2319     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2320     break;
2321   case 5:
2322     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2323     break;
2324   case 6:
2325     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2326             p_argv[5]);
2327     break;
2328   case 7:
2329     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2330             p_argv[5], p_argv[6]);
2331     break;
2332   case 8:
2333     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2334             p_argv[5], p_argv[6], p_argv[7]);
2335     break;
2336   case 9:
2337     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2338             p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2339     break;
2340   case 10:
2341     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2342             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2343     break;
2344   case 11:
2345     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2346             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2347     break;
2348   case 12:
2349     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2350             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2351             p_argv[11]);
2352     break;
2353   case 13:
2354     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2355             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2356             p_argv[11], p_argv[12]);
2357     break;
2358   case 14:
2359     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2360             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2361             p_argv[11], p_argv[12], p_argv[13]);
2362     break;
2363   case 15:
2364     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2365             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2366             p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2367     break;
2368   }
2369 
2370 #if OMPT_SUPPORT
2371   *exit_frame_ptr = 0;
2372 #endif
2373 
2374   return 1;
2375 }
2376 
2377 #endif
2378 
2379 // end of file //
2380