xref: /freebsd/contrib/llvm-project/openmp/runtime/src/z_Linux_util.cpp (revision 6be3386466ab79a84b48429ae66244f21526d3df)
1 /*
2  * z_Linux_util.cpp -- platform specific routines.
3  */
4 
5 //===----------------------------------------------------------------------===//
6 //
7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
8 // See https://llvm.org/LICENSE.txt for license information.
9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "kmp.h"
14 #include "kmp_affinity.h"
15 #include "kmp_i18n.h"
16 #include "kmp_io.h"
17 #include "kmp_itt.h"
18 #include "kmp_lock.h"
19 #include "kmp_stats.h"
20 #include "kmp_str.h"
21 #include "kmp_wait_release.h"
22 #include "kmp_wrapper_getpid.h"
23 
24 #if !KMP_OS_DRAGONFLY && !KMP_OS_FREEBSD && !KMP_OS_NETBSD && !KMP_OS_OPENBSD
25 #include <alloca.h>
26 #endif
27 #include <math.h> // HUGE_VAL.
28 #include <sys/resource.h>
29 #include <sys/syscall.h>
30 #include <sys/time.h>
31 #include <sys/times.h>
32 #include <unistd.h>
33 
34 #if KMP_OS_LINUX && !KMP_OS_CNK
35 #include <sys/sysinfo.h>
36 #if KMP_USE_FUTEX
37 // We should really include <futex.h>, but that causes compatibility problems on
38 // different Linux* OS distributions that either require that you include (or
39 // break when you try to include) <pci/types.h>. Since all we need is the two
40 // macros below (which are part of the kernel ABI, so can't change) we just
41 // define the constants here and don't include <futex.h>
42 #ifndef FUTEX_WAIT
43 #define FUTEX_WAIT 0
44 #endif
45 #ifndef FUTEX_WAKE
46 #define FUTEX_WAKE 1
47 #endif
48 #endif
49 #elif KMP_OS_DARWIN
50 #include <mach/mach.h>
51 #include <sys/sysctl.h>
52 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD
53 #include <sys/types.h>
54 #include <sys/sysctl.h>
55 #include <sys/user.h>
56 #include <pthread_np.h>
57 #elif KMP_OS_NETBSD || KMP_OS_OPENBSD
58 #include <sys/types.h>
59 #include <sys/sysctl.h>
60 #endif
61 
62 #include <ctype.h>
63 #include <dirent.h>
64 #include <fcntl.h>
65 
66 #include "tsan_annotations.h"
67 
68 struct kmp_sys_timer {
69   struct timespec start;
70 };
71 
72 // Convert timespec to nanoseconds.
73 #define TS2NS(timespec) (((timespec).tv_sec * 1e9) + (timespec).tv_nsec)
74 
75 static struct kmp_sys_timer __kmp_sys_timer_data;
76 
77 #if KMP_HANDLE_SIGNALS
78 typedef void (*sig_func_t)(int);
79 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG];
80 static sigset_t __kmp_sigset;
81 #endif
82 
83 static int __kmp_init_runtime = FALSE;
84 
85 static int __kmp_fork_count = 0;
86 
87 static pthread_condattr_t __kmp_suspend_cond_attr;
88 static pthread_mutexattr_t __kmp_suspend_mutex_attr;
89 
90 static kmp_cond_align_t __kmp_wait_cv;
91 static kmp_mutex_align_t __kmp_wait_mx;
92 
93 kmp_uint64 __kmp_ticks_per_msec = 1000000;
94 
95 #ifdef DEBUG_SUSPEND
96 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
97   KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))",
98                cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock,
99                cond->c_cond.__c_waiting);
100 }
101 #endif
102 
103 #if ((KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED)
104 
105 /* Affinity support */
106 
107 void __kmp_affinity_bind_thread(int which) {
108   KMP_ASSERT2(KMP_AFFINITY_CAPABLE(),
109               "Illegal set affinity operation when not capable");
110 
111   kmp_affin_mask_t *mask;
112   KMP_CPU_ALLOC_ON_STACK(mask);
113   KMP_CPU_ZERO(mask);
114   KMP_CPU_SET(which, mask);
115   __kmp_set_system_affinity(mask, TRUE);
116   KMP_CPU_FREE_FROM_STACK(mask);
117 }
118 
119 /* Determine if we can access affinity functionality on this version of
120  * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set
121  * __kmp_affin_mask_size to the appropriate value (0 means not capable). */
122 void __kmp_affinity_determine_capable(const char *env_var) {
123 // Check and see if the OS supports thread affinity.
124 
125 #if KMP_OS_LINUX
126 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024)
127 #elif KMP_OS_FREEBSD
128 #define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t))
129 #endif
130 
131 
132 #if KMP_OS_LINUX
133   // If Linux* OS:
134   // If the syscall fails or returns a suggestion for the size,
135   // then we don't have to search for an appropriate size.
136   int gCode;
137   int sCode;
138   unsigned char *buf;
139   buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
140   gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_SIZE_LIMIT, buf);
141   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
142                 "initial getaffinity call returned %d errno = %d\n",
143                 gCode, errno));
144 
145   // if ((gCode < 0) && (errno == ENOSYS))
146   if (gCode < 0) {
147     // System call not supported
148     if (__kmp_affinity_verbose ||
149         (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
150          (__kmp_affinity_type != affinity_default) &&
151          (__kmp_affinity_type != affinity_disabled))) {
152       int error = errno;
153       kmp_msg_t err_code = KMP_ERR(error);
154       __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
155                 err_code, __kmp_msg_null);
156       if (__kmp_generate_warnings == kmp_warnings_off) {
157         __kmp_str_free(&err_code.str);
158       }
159     }
160     KMP_AFFINITY_DISABLE();
161     KMP_INTERNAL_FREE(buf);
162     return;
163   }
164   if (gCode > 0) { // Linux* OS only
165     // The optimal situation: the OS returns the size of the buffer it expects.
166     //
167     // A verification of correct behavior is that setaffinity on a NULL
168     // buffer with the same size fails with errno set to EFAULT.
169     sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
170     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
171                   "setaffinity for mask size %d returned %d errno = %d\n",
172                   gCode, sCode, errno));
173     if (sCode < 0) {
174       if (errno == ENOSYS) {
175         if (__kmp_affinity_verbose ||
176             (__kmp_affinity_warnings &&
177              (__kmp_affinity_type != affinity_none) &&
178              (__kmp_affinity_type != affinity_default) &&
179              (__kmp_affinity_type != affinity_disabled))) {
180           int error = errno;
181           kmp_msg_t err_code = KMP_ERR(error);
182           __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
183                     err_code, __kmp_msg_null);
184           if (__kmp_generate_warnings == kmp_warnings_off) {
185             __kmp_str_free(&err_code.str);
186           }
187         }
188         KMP_AFFINITY_DISABLE();
189         KMP_INTERNAL_FREE(buf);
190       }
191       if (errno == EFAULT) {
192         KMP_AFFINITY_ENABLE(gCode);
193         KA_TRACE(10, ("__kmp_affinity_determine_capable: "
194                       "affinity supported (mask size %d)\n",
195                       (int)__kmp_affin_mask_size));
196         KMP_INTERNAL_FREE(buf);
197         return;
198       }
199     }
200   }
201 
202   // Call the getaffinity system call repeatedly with increasing set sizes
203   // until we succeed, or reach an upper bound on the search.
204   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
205                 "searching for proper set size\n"));
206   int size;
207   for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) {
208     gCode = syscall(__NR_sched_getaffinity, 0, size, buf);
209     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
210                   "getaffinity for mask size %d returned %d errno = %d\n",
211                   size, gCode, errno));
212 
213     if (gCode < 0) {
214       if (errno == ENOSYS) {
215         // We shouldn't get here
216         KA_TRACE(30, ("__kmp_affinity_determine_capable: "
217                       "inconsistent OS call behavior: errno == ENOSYS for mask "
218                       "size %d\n",
219                       size));
220         if (__kmp_affinity_verbose ||
221             (__kmp_affinity_warnings &&
222              (__kmp_affinity_type != affinity_none) &&
223              (__kmp_affinity_type != affinity_default) &&
224              (__kmp_affinity_type != affinity_disabled))) {
225           int error = errno;
226           kmp_msg_t err_code = KMP_ERR(error);
227           __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var),
228                     err_code, __kmp_msg_null);
229           if (__kmp_generate_warnings == kmp_warnings_off) {
230             __kmp_str_free(&err_code.str);
231           }
232         }
233         KMP_AFFINITY_DISABLE();
234         KMP_INTERNAL_FREE(buf);
235         return;
236       }
237       continue;
238     }
239 
240     sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL);
241     KA_TRACE(30, ("__kmp_affinity_determine_capable: "
242                   "setaffinity for mask size %d returned %d errno = %d\n",
243                   gCode, sCode, errno));
244     if (sCode < 0) {
245       if (errno == ENOSYS) { // Linux* OS only
246         // We shouldn't get here
247         KA_TRACE(30, ("__kmp_affinity_determine_capable: "
248                       "inconsistent OS call behavior: errno == ENOSYS for mask "
249                       "size %d\n",
250                       size));
251         if (__kmp_affinity_verbose ||
252             (__kmp_affinity_warnings &&
253              (__kmp_affinity_type != affinity_none) &&
254              (__kmp_affinity_type != affinity_default) &&
255              (__kmp_affinity_type != affinity_disabled))) {
256           int error = errno;
257           kmp_msg_t err_code = KMP_ERR(error);
258           __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var),
259                     err_code, __kmp_msg_null);
260           if (__kmp_generate_warnings == kmp_warnings_off) {
261             __kmp_str_free(&err_code.str);
262           }
263         }
264         KMP_AFFINITY_DISABLE();
265         KMP_INTERNAL_FREE(buf);
266         return;
267       }
268       if (errno == EFAULT) {
269         KMP_AFFINITY_ENABLE(gCode);
270         KA_TRACE(10, ("__kmp_affinity_determine_capable: "
271                       "affinity supported (mask size %d)\n",
272                       (int)__kmp_affin_mask_size));
273         KMP_INTERNAL_FREE(buf);
274         return;
275       }
276     }
277   }
278 #elif KMP_OS_FREEBSD
279   int gCode;
280   unsigned char *buf;
281   buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
282   gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT, reinterpret_cast<cpuset_t *>(buf));
283   KA_TRACE(30, ("__kmp_affinity_determine_capable: "
284                 "initial getaffinity call returned %d errno = %d\n",
285                 gCode, errno));
286   if (gCode == 0) {
287     KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT);
288     KA_TRACE(10, ("__kmp_affinity_determine_capable: "
289                   "affinity supported (mask size %d)\n",
290 		  (int)__kmp_affin_mask_size));
291     KMP_INTERNAL_FREE(buf);
292     return;
293   }
294 #endif
295   // save uncaught error code
296   // int error = errno;
297   KMP_INTERNAL_FREE(buf);
298   // restore uncaught error code, will be printed at the next KMP_WARNING below
299   // errno = error;
300 
301   // Affinity is not supported
302   KMP_AFFINITY_DISABLE();
303   KA_TRACE(10, ("__kmp_affinity_determine_capable: "
304                 "cannot determine mask size - affinity not supported\n"));
305   if (__kmp_affinity_verbose ||
306       (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) &&
307        (__kmp_affinity_type != affinity_default) &&
308        (__kmp_affinity_type != affinity_disabled))) {
309     KMP_WARNING(AffCantGetMaskSize, env_var);
310   }
311 }
312 
313 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED
314 
315 #if KMP_USE_FUTEX
316 
317 int __kmp_futex_determine_capable() {
318   int loc = 0;
319   int rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0);
320   int retval = (rc == 0) || (errno != ENOSYS);
321 
322   KA_TRACE(10,
323            ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno));
324   KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n",
325                 retval ? "" : " not"));
326 
327   return retval;
328 }
329 
330 #endif // KMP_USE_FUTEX
331 
332 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS)
333 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to
334    use compare_and_store for these routines */
335 
336 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) {
337   kmp_int8 old_value, new_value;
338 
339   old_value = TCR_1(*p);
340   new_value = old_value | d;
341 
342   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
343     KMP_CPU_PAUSE();
344     old_value = TCR_1(*p);
345     new_value = old_value | d;
346   }
347   return old_value;
348 }
349 
350 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) {
351   kmp_int8 old_value, new_value;
352 
353   old_value = TCR_1(*p);
354   new_value = old_value & d;
355 
356   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
357     KMP_CPU_PAUSE();
358     old_value = TCR_1(*p);
359     new_value = old_value & d;
360   }
361   return old_value;
362 }
363 
364 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) {
365   kmp_uint32 old_value, new_value;
366 
367   old_value = TCR_4(*p);
368   new_value = old_value | d;
369 
370   while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
371     KMP_CPU_PAUSE();
372     old_value = TCR_4(*p);
373     new_value = old_value | d;
374   }
375   return old_value;
376 }
377 
378 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) {
379   kmp_uint32 old_value, new_value;
380 
381   old_value = TCR_4(*p);
382   new_value = old_value & d;
383 
384   while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) {
385     KMP_CPU_PAUSE();
386     old_value = TCR_4(*p);
387     new_value = old_value & d;
388   }
389   return old_value;
390 }
391 
392 #if KMP_ARCH_X86
393 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) {
394   kmp_int8 old_value, new_value;
395 
396   old_value = TCR_1(*p);
397   new_value = old_value + d;
398 
399   while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) {
400     KMP_CPU_PAUSE();
401     old_value = TCR_1(*p);
402     new_value = old_value + d;
403   }
404   return old_value;
405 }
406 
407 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) {
408   kmp_int64 old_value, new_value;
409 
410   old_value = TCR_8(*p);
411   new_value = old_value + d;
412 
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 #endif /* KMP_ARCH_X86 */
421 
422 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) {
423   kmp_uint64 old_value, new_value;
424 
425   old_value = TCR_8(*p);
426   new_value = old_value | d;
427   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
428     KMP_CPU_PAUSE();
429     old_value = TCR_8(*p);
430     new_value = old_value | d;
431   }
432   return old_value;
433 }
434 
435 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) {
436   kmp_uint64 old_value, new_value;
437 
438   old_value = TCR_8(*p);
439   new_value = old_value & d;
440   while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) {
441     KMP_CPU_PAUSE();
442     old_value = TCR_8(*p);
443     new_value = old_value & d;
444   }
445   return old_value;
446 }
447 
448 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */
449 
450 void __kmp_terminate_thread(int gtid) {
451   int status;
452   kmp_info_t *th = __kmp_threads[gtid];
453 
454   if (!th)
455     return;
456 
457 #ifdef KMP_CANCEL_THREADS
458   KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
459   status = pthread_cancel(th->th.th_info.ds.ds_thread);
460   if (status != 0 && status != ESRCH) {
461     __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status),
462                 __kmp_msg_null);
463   }
464 #endif
465   KMP_YIELD(TRUE);
466 } //
467 
468 /* Set thread stack info according to values returned by pthread_getattr_np().
469    If values are unreasonable, assume call failed and use incremental stack
470    refinement method instead. Returns TRUE if the stack parameters could be
471    determined exactly, FALSE if incremental refinement is necessary. */
472 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) {
473   int stack_data;
474 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
475         KMP_OS_HURD
476   pthread_attr_t attr;
477   int status;
478   size_t size = 0;
479   void *addr = 0;
480 
481   /* Always do incremental stack refinement for ubermaster threads since the
482      initial thread stack range can be reduced by sibling thread creation so
483      pthread_attr_getstack may cause thread gtid aliasing */
484   if (!KMP_UBER_GTID(gtid)) {
485 
486     /* Fetch the real thread attributes */
487     status = pthread_attr_init(&attr);
488     KMP_CHECK_SYSFAIL("pthread_attr_init", status);
489 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD
490     status = pthread_attr_get_np(pthread_self(), &attr);
491     KMP_CHECK_SYSFAIL("pthread_attr_get_np", status);
492 #else
493     status = pthread_getattr_np(pthread_self(), &attr);
494     KMP_CHECK_SYSFAIL("pthread_getattr_np", status);
495 #endif
496     status = pthread_attr_getstack(&attr, &addr, &size);
497     KMP_CHECK_SYSFAIL("pthread_attr_getstack", status);
498     KA_TRACE(60,
499              ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:"
500               " %lu, low addr: %p\n",
501               gtid, size, addr));
502     status = pthread_attr_destroy(&attr);
503     KMP_CHECK_SYSFAIL("pthread_attr_destroy", status);
504   }
505 
506   if (size != 0 && addr != 0) { // was stack parameter determination successful?
507     /* Store the correct base and size */
508     TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size));
509     TCW_PTR(th->th.th_info.ds.ds_stacksize, size);
510     TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
511     return TRUE;
512   }
513 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||
514               KMP_OS_HURD */
515   /* Use incremental refinement starting from initial conservative estimate */
516   TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
517   TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
518   TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
519   return FALSE;
520 }
521 
522 static void *__kmp_launch_worker(void *thr) {
523   int status, old_type, old_state;
524 #ifdef KMP_BLOCK_SIGNALS
525   sigset_t new_set, old_set;
526 #endif /* KMP_BLOCK_SIGNALS */
527   void *exit_val;
528 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
529         KMP_OS_OPENBSD || KMP_OS_HURD
530   void *volatile padding = 0;
531 #endif
532   int gtid;
533 
534   gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid;
535   __kmp_gtid_set_specific(gtid);
536 #ifdef KMP_TDATA_GTID
537   __kmp_gtid = gtid;
538 #endif
539 #if KMP_STATS_ENABLED
540   // set thread local index to point to thread-specific stats
541   __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats;
542   __kmp_stats_thread_ptr->startLife();
543   KMP_SET_THREAD_STATE(IDLE);
544   KMP_INIT_PARTITIONED_TIMERS(OMP_idle);
545 #endif
546 
547 #if USE_ITT_BUILD
548   __kmp_itt_thread_name(gtid);
549 #endif /* USE_ITT_BUILD */
550 
551 #if KMP_AFFINITY_SUPPORTED
552   __kmp_affinity_set_init_mask(gtid, FALSE);
553 #endif
554 
555 #ifdef KMP_CANCEL_THREADS
556   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
557   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
558   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
559   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
560   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
561 #endif
562 
563 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
564   // Set FP control regs to be a copy of the parallel initialization thread's.
565   __kmp_clear_x87_fpu_status_word();
566   __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
567   __kmp_load_mxcsr(&__kmp_init_mxcsr);
568 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
569 
570 #ifdef KMP_BLOCK_SIGNALS
571   status = sigfillset(&new_set);
572   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
573   status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set);
574   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
575 #endif /* KMP_BLOCK_SIGNALS */
576 
577 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
578         KMP_OS_OPENBSD
579   if (__kmp_stkoffset > 0 && gtid > 0) {
580     padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
581   }
582 #endif
583 
584   KMP_MB();
585   __kmp_set_stack_info(gtid, (kmp_info_t *)thr);
586 
587   __kmp_check_stack_overlap((kmp_info_t *)thr);
588 
589   exit_val = __kmp_launch_thread((kmp_info_t *)thr);
590 
591 #ifdef KMP_BLOCK_SIGNALS
592   status = pthread_sigmask(SIG_SETMASK, &old_set, NULL);
593   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
594 #endif /* KMP_BLOCK_SIGNALS */
595 
596   return exit_val;
597 }
598 
599 #if KMP_USE_MONITOR
600 /* The monitor thread controls all of the threads in the complex */
601 
602 static void *__kmp_launch_monitor(void *thr) {
603   int status, old_type, old_state;
604 #ifdef KMP_BLOCK_SIGNALS
605   sigset_t new_set;
606 #endif /* KMP_BLOCK_SIGNALS */
607   struct timespec interval;
608 
609   KMP_MB(); /* Flush all pending memory write invalidates.  */
610 
611   KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n"));
612 
613   /* register us as the monitor thread */
614   __kmp_gtid_set_specific(KMP_GTID_MONITOR);
615 #ifdef KMP_TDATA_GTID
616   __kmp_gtid = KMP_GTID_MONITOR;
617 #endif
618 
619   KMP_MB();
620 
621 #if USE_ITT_BUILD
622   // Instruct Intel(R) Threading Tools to ignore monitor thread.
623   __kmp_itt_thread_ignore();
624 #endif /* USE_ITT_BUILD */
625 
626   __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid,
627                        (kmp_info_t *)thr);
628 
629   __kmp_check_stack_overlap((kmp_info_t *)thr);
630 
631 #ifdef KMP_CANCEL_THREADS
632   status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type);
633   KMP_CHECK_SYSFAIL("pthread_setcanceltype", status);
634   // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads?
635   status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state);
636   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
637 #endif
638 
639 #if KMP_REAL_TIME_FIX
640   // This is a potential fix which allows application with real-time scheduling
641   // policy work. However, decision about the fix is not made yet, so it is
642   // disabled by default.
643   { // Are program started with real-time scheduling policy?
644     int sched = sched_getscheduler(0);
645     if (sched == SCHED_FIFO || sched == SCHED_RR) {
646       // Yes, we are a part of real-time application. Try to increase the
647       // priority of the monitor.
648       struct sched_param param;
649       int max_priority = sched_get_priority_max(sched);
650       int rc;
651       KMP_WARNING(RealTimeSchedNotSupported);
652       sched_getparam(0, &param);
653       if (param.sched_priority < max_priority) {
654         param.sched_priority += 1;
655         rc = sched_setscheduler(0, sched, &param);
656         if (rc != 0) {
657           int error = errno;
658           kmp_msg_t err_code = KMP_ERR(error);
659           __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority),
660                     err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null);
661           if (__kmp_generate_warnings == kmp_warnings_off) {
662             __kmp_str_free(&err_code.str);
663           }
664         }
665       } else {
666         // We cannot abort here, because number of CPUs may be enough for all
667         // the threads, including the monitor thread, so application could
668         // potentially work...
669         __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority),
670                   KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority),
671                   __kmp_msg_null);
672       }
673     }
674     // AC: free thread that waits for monitor started
675     TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
676   }
677 #endif // KMP_REAL_TIME_FIX
678 
679   KMP_MB(); /* Flush all pending memory write invalidates.  */
680 
681   if (__kmp_monitor_wakeups == 1) {
682     interval.tv_sec = 1;
683     interval.tv_nsec = 0;
684   } else {
685     interval.tv_sec = 0;
686     interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups);
687   }
688 
689   KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n"));
690 
691   while (!TCR_4(__kmp_global.g.g_done)) {
692     struct timespec now;
693     struct timeval tval;
694 
695     /*  This thread monitors the state of the system */
696 
697     KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
698 
699     status = gettimeofday(&tval, NULL);
700     KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
701     TIMEVAL_TO_TIMESPEC(&tval, &now);
702 
703     now.tv_sec += interval.tv_sec;
704     now.tv_nsec += interval.tv_nsec;
705 
706     if (now.tv_nsec >= KMP_NSEC_PER_SEC) {
707       now.tv_sec += 1;
708       now.tv_nsec -= KMP_NSEC_PER_SEC;
709     }
710 
711     status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
712     KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
713     // AC: the monitor should not fall asleep if g_done has been set
714     if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex
715       status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond,
716                                       &__kmp_wait_mx.m_mutex, &now);
717       if (status != 0) {
718         if (status != ETIMEDOUT && status != EINTR) {
719           KMP_SYSFAIL("pthread_cond_timedwait", status);
720         }
721       }
722     }
723     status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
724     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
725 
726     TCW_4(__kmp_global.g.g_time.dt.t_value,
727           TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
728 
729     KMP_MB(); /* Flush all pending memory write invalidates.  */
730   }
731 
732   KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n"));
733 
734 #ifdef KMP_BLOCK_SIGNALS
735   status = sigfillset(&new_set);
736   KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status);
737   status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL);
738   KMP_CHECK_SYSFAIL("pthread_sigmask", status);
739 #endif /* KMP_BLOCK_SIGNALS */
740 
741   KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n"));
742 
743   if (__kmp_global.g.g_abort != 0) {
744     /* now we need to terminate the worker threads  */
745     /* the value of t_abort is the signal we caught */
746 
747     int gtid;
748 
749     KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n",
750                   __kmp_global.g.g_abort));
751 
752     /* terminate the OpenMP worker threads */
753     /* TODO this is not valid for sibling threads!!
754      * the uber master might not be 0 anymore.. */
755     for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
756       __kmp_terminate_thread(gtid);
757 
758     __kmp_cleanup();
759 
760     KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n",
761                   __kmp_global.g.g_abort));
762 
763     if (__kmp_global.g.g_abort > 0)
764       raise(__kmp_global.g.g_abort);
765   }
766 
767   KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n"));
768 
769   return thr;
770 }
771 #endif // KMP_USE_MONITOR
772 
773 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
774   pthread_t handle;
775   pthread_attr_t thread_attr;
776   int status;
777 
778   th->th.th_info.ds.ds_gtid = gtid;
779 
780 #if KMP_STATS_ENABLED
781   // sets up worker thread stats
782   __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid);
783 
784   // th->th.th_stats is used to transfer thread-specific stats-pointer to
785   // __kmp_launch_worker. So when thread is created (goes into
786   // __kmp_launch_worker) it will set its thread local pointer to
787   // th->th.th_stats
788   if (!KMP_UBER_GTID(gtid)) {
789     th->th.th_stats = __kmp_stats_list->push_back(gtid);
790   } else {
791     // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(),
792     // so set the th->th.th_stats field to it.
793     th->th.th_stats = __kmp_stats_thread_ptr;
794   }
795   __kmp_release_tas_lock(&__kmp_stats_lock, gtid);
796 
797 #endif // KMP_STATS_ENABLED
798 
799   if (KMP_UBER_GTID(gtid)) {
800     KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid));
801     th->th.th_info.ds.ds_thread = pthread_self();
802     __kmp_set_stack_info(gtid, th);
803     __kmp_check_stack_overlap(th);
804     return;
805   }
806 
807   KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
808 
809   KMP_MB(); /* Flush all pending memory write invalidates.  */
810 
811 #ifdef KMP_THREAD_ATTR
812   status = pthread_attr_init(&thread_attr);
813   if (status != 0) {
814     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
815   }
816   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
817   if (status != 0) {
818     __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null);
819   }
820 
821   /* Set stack size for this thread now.
822      The multiple of 2 is there because on some machines, requesting an unusual
823      stacksize causes the thread to have an offset before the dummy alloca()
824      takes place to create the offset.  Since we want the user to have a
825      sufficient stacksize AND support a stack offset, we alloca() twice the
826      offset so that the upcoming alloca() does not eliminate any premade offset,
827      and also gives the user the stack space they requested for all threads */
828   stack_size += gtid * __kmp_stkoffset * 2;
829 
830 #if defined(__ANDROID__) && __ANDROID_API__ < 19
831     // Round the stack size to a multiple of the page size. Older versions of
832     // Android (until KitKat) would fail pthread_attr_setstacksize with EINVAL
833     // if the stack size was not a multiple of the page size.
834     stack_size = (stack_size + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1);
835 #endif
836 
837   KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
838                 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n",
839                 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
840 
841 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
842   status = pthread_attr_setstacksize(&thread_attr, stack_size);
843 #ifdef KMP_BACKUP_STKSIZE
844   if (status != 0) {
845     if (!__kmp_env_stksize) {
846       stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset;
847       __kmp_stksize = KMP_BACKUP_STKSIZE;
848       KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, "
849                     "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu "
850                     "bytes\n",
851                     gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size));
852       status = pthread_attr_setstacksize(&thread_attr, stack_size);
853     }
854   }
855 #endif /* KMP_BACKUP_STKSIZE */
856   if (status != 0) {
857     __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
858                 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null);
859   }
860 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
861 
862 #endif /* KMP_THREAD_ATTR */
863 
864   status =
865       pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th);
866   if (status != 0 || !handle) { // ??? Why do we check handle??
867 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
868     if (status == EINVAL) {
869       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
870                   KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null);
871     }
872     if (status == ENOMEM) {
873       __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status),
874                   KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null);
875     }
876 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
877     if (status == EAGAIN) {
878       __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status),
879                   KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null);
880     }
881     KMP_SYSFAIL("pthread_create", status);
882   }
883 
884   th->th.th_info.ds.ds_thread = handle;
885 
886 #ifdef KMP_THREAD_ATTR
887   status = pthread_attr_destroy(&thread_attr);
888   if (status) {
889     kmp_msg_t err_code = KMP_ERR(status);
890     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
891               __kmp_msg_null);
892     if (__kmp_generate_warnings == kmp_warnings_off) {
893       __kmp_str_free(&err_code.str);
894     }
895   }
896 #endif /* KMP_THREAD_ATTR */
897 
898   KMP_MB(); /* Flush all pending memory write invalidates.  */
899 
900   KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
901 
902 } // __kmp_create_worker
903 
904 #if KMP_USE_MONITOR
905 void __kmp_create_monitor(kmp_info_t *th) {
906   pthread_t handle;
907   pthread_attr_t thread_attr;
908   size_t size;
909   int status;
910   int auto_adj_size = FALSE;
911 
912   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
913     // We don't need monitor thread in case of MAX_BLOCKTIME
914     KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
915                   "MAX blocktime\n"));
916     th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
917     th->th.th_info.ds.ds_gtid = 0;
918     return;
919   }
920   KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
921 
922   KMP_MB(); /* Flush all pending memory write invalidates.  */
923 
924   th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
925   th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
926 #if KMP_REAL_TIME_FIX
927   TCW_4(__kmp_global.g.g_time.dt.t_value,
928         -1); // Will use it for synchronization a bit later.
929 #else
930   TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
931 #endif // KMP_REAL_TIME_FIX
932 
933 #ifdef KMP_THREAD_ATTR
934   if (__kmp_monitor_stksize == 0) {
935     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
936     auto_adj_size = TRUE;
937   }
938   status = pthread_attr_init(&thread_attr);
939   if (status != 0) {
940     __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null);
941   }
942   status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE);
943   if (status != 0) {
944     __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null);
945   }
946 
947 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
948   status = pthread_attr_getstacksize(&thread_attr, &size);
949   KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status);
950 #else
951   size = __kmp_sys_min_stksize;
952 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
953 #endif /* KMP_THREAD_ATTR */
954 
955   if (__kmp_monitor_stksize == 0) {
956     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
957   }
958   if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
959     __kmp_monitor_stksize = __kmp_sys_min_stksize;
960   }
961 
962   KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes,"
963                 "requested stacksize = %lu bytes\n",
964                 size, __kmp_monitor_stksize));
965 
966 retry:
967 
968 /* Set stack size for this thread now. */
969 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
970   KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,",
971                 __kmp_monitor_stksize));
972   status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize);
973   if (status != 0) {
974     if (auto_adj_size) {
975       __kmp_monitor_stksize *= 2;
976       goto retry;
977     }
978     kmp_msg_t err_code = KMP_ERR(status);
979     __kmp_msg(kmp_ms_warning, // should this be fatal?  BB
980               KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize),
981               err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null);
982     if (__kmp_generate_warnings == kmp_warnings_off) {
983       __kmp_str_free(&err_code.str);
984     }
985   }
986 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
987 
988   status =
989       pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th);
990 
991   if (status != 0) {
992 #ifdef _POSIX_THREAD_ATTR_STACKSIZE
993     if (status == EINVAL) {
994       if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) {
995         __kmp_monitor_stksize *= 2;
996         goto retry;
997       }
998       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
999                   KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize),
1000                   __kmp_msg_null);
1001     }
1002     if (status == ENOMEM) {
1003       __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize),
1004                   KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize),
1005                   __kmp_msg_null);
1006     }
1007 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */
1008     if (status == EAGAIN) {
1009       __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status),
1010                   KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null);
1011     }
1012     KMP_SYSFAIL("pthread_create", status);
1013   }
1014 
1015   th->th.th_info.ds.ds_thread = handle;
1016 
1017 #if KMP_REAL_TIME_FIX
1018   // Wait for the monitor thread is really started and set its *priority*.
1019   KMP_DEBUG_ASSERT(sizeof(kmp_uint32) ==
1020                    sizeof(__kmp_global.g.g_time.dt.t_value));
1021   __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1,
1022                &__kmp_neq_4, NULL);
1023 #endif // KMP_REAL_TIME_FIX
1024 
1025 #ifdef KMP_THREAD_ATTR
1026   status = pthread_attr_destroy(&thread_attr);
1027   if (status != 0) {
1028     kmp_msg_t err_code = KMP_ERR(status);
1029     __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code,
1030               __kmp_msg_null);
1031     if (__kmp_generate_warnings == kmp_warnings_off) {
1032       __kmp_str_free(&err_code.str);
1033     }
1034   }
1035 #endif
1036 
1037   KMP_MB(); /* Flush all pending memory write invalidates.  */
1038 
1039   KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n",
1040                 th->th.th_info.ds.ds_thread));
1041 
1042 } // __kmp_create_monitor
1043 #endif // KMP_USE_MONITOR
1044 
1045 void __kmp_exit_thread(int exit_status) {
1046   pthread_exit((void *)(intptr_t)exit_status);
1047 } // __kmp_exit_thread
1048 
1049 #if KMP_USE_MONITOR
1050 void __kmp_resume_monitor();
1051 
1052 void __kmp_reap_monitor(kmp_info_t *th) {
1053   int status;
1054   void *exit_val;
1055 
1056   KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle"
1057                 " %#.8lx\n",
1058                 th->th.th_info.ds.ds_thread));
1059 
1060   // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1061   // If both tid and gtid are 0, it means the monitor did not ever start.
1062   // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1063   KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1064   if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1065     KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1066     return;
1067   }
1068 
1069   KMP_MB(); /* Flush all pending memory write invalidates.  */
1070 
1071   /* First, check to see whether the monitor thread exists to wake it up. This
1072      is to avoid performance problem when the monitor sleeps during
1073      blocktime-size interval */
1074 
1075   status = pthread_kill(th->th.th_info.ds.ds_thread, 0);
1076   if (status != ESRCH) {
1077     __kmp_resume_monitor(); // Wake up the monitor thread
1078   }
1079   KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n"));
1080   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1081   if (exit_val != th) {
1082     __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null);
1083   }
1084 
1085   th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1086   th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1087 
1088   KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle"
1089                 " %#.8lx\n",
1090                 th->th.th_info.ds.ds_thread));
1091 
1092   KMP_MB(); /* Flush all pending memory write invalidates.  */
1093 }
1094 #endif // KMP_USE_MONITOR
1095 
1096 void __kmp_reap_worker(kmp_info_t *th) {
1097   int status;
1098   void *exit_val;
1099 
1100   KMP_MB(); /* Flush all pending memory write invalidates.  */
1101 
1102   KA_TRACE(
1103       10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid));
1104 
1105   status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val);
1106 #ifdef KMP_DEBUG
1107   /* Don't expose these to the user until we understand when they trigger */
1108   if (status != 0) {
1109     __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null);
1110   }
1111   if (exit_val != th) {
1112     KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, "
1113                   "exit_val = %p\n",
1114                   th->th.th_info.ds.ds_gtid, exit_val));
1115   }
1116 #endif /* KMP_DEBUG */
1117 
1118   KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n",
1119                 th->th.th_info.ds.ds_gtid));
1120 
1121   KMP_MB(); /* Flush all pending memory write invalidates.  */
1122 }
1123 
1124 #if KMP_HANDLE_SIGNALS
1125 
1126 static void __kmp_null_handler(int signo) {
1127   //  Do nothing, for doing SIG_IGN-type actions.
1128 } // __kmp_null_handler
1129 
1130 static void __kmp_team_handler(int signo) {
1131   if (__kmp_global.g.g_abort == 0) {
1132 /* Stage 1 signal handler, let's shut down all of the threads */
1133 #ifdef KMP_DEBUG
1134     __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo);
1135 #endif
1136     switch (signo) {
1137     case SIGHUP:
1138     case SIGINT:
1139     case SIGQUIT:
1140     case SIGILL:
1141     case SIGABRT:
1142     case SIGFPE:
1143     case SIGBUS:
1144     case SIGSEGV:
1145 #ifdef SIGSYS
1146     case SIGSYS:
1147 #endif
1148     case SIGTERM:
1149       if (__kmp_debug_buf) {
1150         __kmp_dump_debug_buffer();
1151       }
1152       KMP_MB(); // Flush all pending memory write invalidates.
1153       TCW_4(__kmp_global.g.g_abort, signo);
1154       KMP_MB(); // Flush all pending memory write invalidates.
1155       TCW_4(__kmp_global.g.g_done, TRUE);
1156       KMP_MB(); // Flush all pending memory write invalidates.
1157       break;
1158     default:
1159 #ifdef KMP_DEBUG
1160       __kmp_debug_printf("__kmp_team_handler: unknown signal type");
1161 #endif
1162       break;
1163     }
1164   }
1165 } // __kmp_team_handler
1166 
1167 static void __kmp_sigaction(int signum, const struct sigaction *act,
1168                             struct sigaction *oldact) {
1169   int rc = sigaction(signum, act, oldact);
1170   KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc);
1171 }
1172 
1173 static void __kmp_install_one_handler(int sig, sig_func_t handler_func,
1174                                       int parallel_init) {
1175   KMP_MB(); // Flush all pending memory write invalidates.
1176   KB_TRACE(60,
1177            ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init));
1178   if (parallel_init) {
1179     struct sigaction new_action;
1180     struct sigaction old_action;
1181     new_action.sa_handler = handler_func;
1182     new_action.sa_flags = 0;
1183     sigfillset(&new_action.sa_mask);
1184     __kmp_sigaction(sig, &new_action, &old_action);
1185     if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) {
1186       sigaddset(&__kmp_sigset, sig);
1187     } else {
1188       // Restore/keep user's handler if one previously installed.
1189       __kmp_sigaction(sig, &old_action, NULL);
1190     }
1191   } else {
1192     // Save initial/system signal handlers to see if user handlers installed.
1193     __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]);
1194   }
1195   KMP_MB(); // Flush all pending memory write invalidates.
1196 } // __kmp_install_one_handler
1197 
1198 static void __kmp_remove_one_handler(int sig) {
1199   KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig));
1200   if (sigismember(&__kmp_sigset, sig)) {
1201     struct sigaction old;
1202     KMP_MB(); // Flush all pending memory write invalidates.
1203     __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old);
1204     if ((old.sa_handler != __kmp_team_handler) &&
1205         (old.sa_handler != __kmp_null_handler)) {
1206       // Restore the users signal handler.
1207       KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1208                     "restoring: sig=%d\n",
1209                     sig));
1210       __kmp_sigaction(sig, &old, NULL);
1211     }
1212     sigdelset(&__kmp_sigset, sig);
1213     KMP_MB(); // Flush all pending memory write invalidates.
1214   }
1215 } // __kmp_remove_one_handler
1216 
1217 void __kmp_install_signals(int parallel_init) {
1218   KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init));
1219   if (__kmp_handle_signals || !parallel_init) {
1220     // If ! parallel_init, we do not install handlers, just save original
1221     // handlers. Let us do it even __handle_signals is 0.
1222     sigemptyset(&__kmp_sigset);
1223     __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init);
1224     __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1225     __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init);
1226     __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1227     __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1228     __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1229     __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init);
1230     __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1231 #ifdef SIGSYS
1232     __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init);
1233 #endif // SIGSYS
1234     __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1235 #ifdef SIGPIPE
1236     __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init);
1237 #endif // SIGPIPE
1238   }
1239 } // __kmp_install_signals
1240 
1241 void __kmp_remove_signals(void) {
1242   int sig;
1243   KB_TRACE(10, ("__kmp_remove_signals()\n"));
1244   for (sig = 1; sig < NSIG; ++sig) {
1245     __kmp_remove_one_handler(sig);
1246   }
1247 } // __kmp_remove_signals
1248 
1249 #endif // KMP_HANDLE_SIGNALS
1250 
1251 void __kmp_enable(int new_state) {
1252 #ifdef KMP_CANCEL_THREADS
1253   int status, old_state;
1254   status = pthread_setcancelstate(new_state, &old_state);
1255   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1256   KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE);
1257 #endif
1258 }
1259 
1260 void __kmp_disable(int *old_state) {
1261 #ifdef KMP_CANCEL_THREADS
1262   int status;
1263   status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state);
1264   KMP_CHECK_SYSFAIL("pthread_setcancelstate", status);
1265 #endif
1266 }
1267 
1268 static void __kmp_atfork_prepare(void) {
1269   __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);
1270   __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock);
1271 }
1272 
1273 static void __kmp_atfork_parent(void) {
1274   __kmp_release_bootstrap_lock(&__kmp_initz_lock);
1275   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1276 }
1277 
1278 /* Reset the library so execution in the child starts "all over again" with
1279    clean data structures in initial states.  Don't worry about freeing memory
1280    allocated by parent, just abandon it to be safe. */
1281 static void __kmp_atfork_child(void) {
1282   __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock);
1283   /* TODO make sure this is done right for nested/sibling */
1284   // ATT:  Memory leaks are here? TODO: Check it and fix.
1285   /* KMP_ASSERT( 0 ); */
1286 
1287   ++__kmp_fork_count;
1288 
1289 #if KMP_AFFINITY_SUPPORTED
1290 #if KMP_OS_LINUX || KMP_OS_FREEBSD
1291   // reset the affinity in the child to the initial thread
1292   // affinity in the parent
1293   kmp_set_thread_affinity_mask_initial();
1294 #endif
1295   // Set default not to bind threads tightly in the child (we’re expecting
1296   // over-subscription after the fork and this can improve things for
1297   // scripting languages that use OpenMP inside process-parallel code).
1298   __kmp_affinity_type = affinity_none;
1299   if (__kmp_nested_proc_bind.bind_types != NULL) {
1300     __kmp_nested_proc_bind.bind_types[0] = proc_bind_false;
1301   }
1302 #endif // KMP_AFFINITY_SUPPORTED
1303 
1304   __kmp_init_runtime = FALSE;
1305 #if KMP_USE_MONITOR
1306   __kmp_init_monitor = 0;
1307 #endif
1308   __kmp_init_parallel = FALSE;
1309   __kmp_init_middle = FALSE;
1310   __kmp_init_serial = FALSE;
1311   TCW_4(__kmp_init_gtid, FALSE);
1312   __kmp_init_common = FALSE;
1313 
1314   TCW_4(__kmp_init_user_locks, FALSE);
1315 #if !KMP_USE_DYNAMIC_LOCK
1316   __kmp_user_lock_table.used = 1;
1317   __kmp_user_lock_table.allocated = 0;
1318   __kmp_user_lock_table.table = NULL;
1319   __kmp_lock_blocks = NULL;
1320 #endif
1321 
1322   __kmp_all_nth = 0;
1323   TCW_4(__kmp_nth, 0);
1324 
1325   __kmp_thread_pool = NULL;
1326   __kmp_thread_pool_insert_pt = NULL;
1327   __kmp_team_pool = NULL;
1328 
1329   /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate
1330      here so threadprivate doesn't use stale data */
1331   KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n",
1332                 __kmp_threadpriv_cache_list));
1333 
1334   while (__kmp_threadpriv_cache_list != NULL) {
1335 
1336     if (*__kmp_threadpriv_cache_list->addr != NULL) {
1337       KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n",
1338                     &(*__kmp_threadpriv_cache_list->addr)));
1339 
1340       *__kmp_threadpriv_cache_list->addr = NULL;
1341     }
1342     __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next;
1343   }
1344 
1345   __kmp_init_runtime = FALSE;
1346 
1347   /* reset statically initialized locks */
1348   __kmp_init_bootstrap_lock(&__kmp_initz_lock);
1349   __kmp_init_bootstrap_lock(&__kmp_stdio_lock);
1350   __kmp_init_bootstrap_lock(&__kmp_console_lock);
1351   __kmp_init_bootstrap_lock(&__kmp_task_team_lock);
1352 
1353 #if USE_ITT_BUILD
1354   __kmp_itt_reset(); // reset ITT's global state
1355 #endif /* USE_ITT_BUILD */
1356 
1357   /* This is necessary to make sure no stale data is left around */
1358   /* AC: customers complain that we use unsafe routines in the atfork
1359      handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen
1360      in dynamic_link when check the presence of shared tbbmalloc library.
1361      Suggestion is to make the library initialization lazier, similar
1362      to what done for __kmpc_begin(). */
1363   // TODO: synchronize all static initializations with regular library
1364   //       startup; look at kmp_global.cpp and etc.
1365   //__kmp_internal_begin ();
1366 }
1367 
1368 void __kmp_register_atfork(void) {
1369   if (__kmp_need_register_atfork) {
1370     int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent,
1371                                 __kmp_atfork_child);
1372     KMP_CHECK_SYSFAIL("pthread_atfork", status);
1373     __kmp_need_register_atfork = FALSE;
1374   }
1375 }
1376 
1377 void __kmp_suspend_initialize(void) {
1378   int status;
1379   status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr);
1380   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1381   status = pthread_condattr_init(&__kmp_suspend_cond_attr);
1382   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1383 }
1384 
1385 void __kmp_suspend_initialize_thread(kmp_info_t *th) {
1386   ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count);
1387   int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count);
1388   int new_value = __kmp_fork_count + 1;
1389   // Return if already initialized
1390   if (old_value == new_value)
1391     return;
1392   // Wait, then return if being initialized
1393   if (old_value == -1 ||
1394       !__kmp_atomic_compare_store(&th->th.th_suspend_init_count, old_value,
1395                                   -1)) {
1396     while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) {
1397       KMP_CPU_PAUSE();
1398     }
1399   } else {
1400     // Claim to be the initializer and do initializations
1401     int status;
1402     status = pthread_cond_init(&th->th.th_suspend_cv.c_cond,
1403                                &__kmp_suspend_cond_attr);
1404     KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1405     status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex,
1406                                 &__kmp_suspend_mutex_attr);
1407     KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1408     KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value);
1409     ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count);
1410   }
1411 }
1412 
1413 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
1414   if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) {
1415     /* this means we have initialize the suspension pthread objects for this
1416        thread in this instance of the process */
1417     int status;
1418 
1419     status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond);
1420     if (status != 0 && status != EBUSY) {
1421       KMP_SYSFAIL("pthread_cond_destroy", status);
1422     }
1423     status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex);
1424     if (status != 0 && status != EBUSY) {
1425       KMP_SYSFAIL("pthread_mutex_destroy", status);
1426     }
1427     --th->th.th_suspend_init_count;
1428     KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) ==
1429                      __kmp_fork_count);
1430   }
1431 }
1432 
1433 // return true if lock obtained, false otherwise
1434 int __kmp_try_suspend_mx(kmp_info_t *th) {
1435   return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0);
1436 }
1437 
1438 void __kmp_lock_suspend_mx(kmp_info_t *th) {
1439   int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1440   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1441 }
1442 
1443 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
1444   int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1445   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1446 }
1447 
1448 /* This routine puts the calling thread to sleep after setting the
1449    sleep bit for the indicated flag variable to true. */
1450 template <class C>
1451 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
1452   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend);
1453   kmp_info_t *th = __kmp_threads[th_gtid];
1454   int status;
1455   typename C::flag_t old_spin;
1456 
1457   KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid,
1458                 flag->get()));
1459 
1460   __kmp_suspend_initialize_thread(th);
1461 
1462   status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1463   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1464 
1465   KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n",
1466                 th_gtid, flag->get()));
1467 
1468   /* TODO: shouldn't this use release semantics to ensure that
1469      __kmp_suspend_initialize_thread gets called first? */
1470   old_spin = flag->set_sleeping();
1471   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
1472       __kmp_pause_status != kmp_soft_paused) {
1473     flag->unset_sleeping();
1474     status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1475     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1476     return;
1477   }
1478   KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x,"
1479                " was %x\n",
1480                th_gtid, flag->get(), flag->load(), old_spin));
1481 
1482   if (flag->done_check_val(old_spin)) {
1483     old_spin = flag->unset_sleeping();
1484     KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
1485                  "for spin(%p)\n",
1486                  th_gtid, flag->get()));
1487   } else {
1488     /* Encapsulate in a loop as the documentation states that this may
1489        "with low probability" return when the condition variable has
1490        not been signaled or broadcast */
1491     int deactivated = FALSE;
1492     TCW_PTR(th->th.th_sleep_loc, (void *)flag);
1493 
1494     while (flag->is_sleeping()) {
1495 #ifdef DEBUG_SUSPEND
1496       char buffer[128];
1497       __kmp_suspend_count++;
1498       __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1499       __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid,
1500                    buffer);
1501 #endif
1502       // Mark the thread as no longer active (only in the first iteration of the
1503       // loop).
1504       if (!deactivated) {
1505         th->th.th_active = FALSE;
1506         if (th->th.th_active_in_pool) {
1507           th->th.th_active_in_pool = FALSE;
1508           KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
1509           KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
1510         }
1511         deactivated = TRUE;
1512       }
1513 
1514 #if USE_SUSPEND_TIMEOUT
1515       struct timespec now;
1516       struct timeval tval;
1517       int msecs;
1518 
1519       status = gettimeofday(&tval, NULL);
1520       KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1521       TIMEVAL_TO_TIMESPEC(&tval, &now);
1522 
1523       msecs = (4 * __kmp_dflt_blocktime) + 200;
1524       now.tv_sec += msecs / 1000;
1525       now.tv_nsec += (msecs % 1000) * 1000;
1526 
1527       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
1528                     "pthread_cond_timedwait\n",
1529                     th_gtid));
1530       status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond,
1531                                       &th->th.th_suspend_mx.m_mutex, &now);
1532 #else
1533       KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform"
1534                     " pthread_cond_wait\n",
1535                     th_gtid));
1536       status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond,
1537                                  &th->th.th_suspend_mx.m_mutex);
1538 #endif
1539 
1540       if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) {
1541         KMP_SYSFAIL("pthread_cond_wait", status);
1542       }
1543 #ifdef KMP_DEBUG
1544       if (status == ETIMEDOUT) {
1545         if (flag->is_sleeping()) {
1546           KF_TRACE(100,
1547                    ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid));
1548         } else {
1549           KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit "
1550                        "not set!\n",
1551                        th_gtid));
1552         }
1553       } else if (flag->is_sleeping()) {
1554         KF_TRACE(100,
1555                  ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
1556       }
1557 #endif
1558     } // while
1559 
1560     // Mark the thread as active again (if it was previous marked as inactive)
1561     if (deactivated) {
1562       th->th.th_active = TRUE;
1563       if (TCR_4(th->th.th_in_pool)) {
1564         KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
1565         th->th.th_active_in_pool = TRUE;
1566       }
1567     }
1568   }
1569 #ifdef DEBUG_SUSPEND
1570   {
1571     char buffer[128];
1572     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1573     __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid,
1574                  buffer);
1575   }
1576 #endif
1577 
1578   status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1579   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1580   KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
1581 }
1582 
1583 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) {
1584   __kmp_suspend_template(th_gtid, flag);
1585 }
1586 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) {
1587   __kmp_suspend_template(th_gtid, flag);
1588 }
1589 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
1590   __kmp_suspend_template(th_gtid, flag);
1591 }
1592 
1593 /* This routine signals the thread specified by target_gtid to wake up
1594    after setting the sleep bit indicated by the flag argument to FALSE.
1595    The target thread must already have called __kmp_suspend_template() */
1596 template <class C>
1597 static inline void __kmp_resume_template(int target_gtid, C *flag) {
1598   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1599   kmp_info_t *th = __kmp_threads[target_gtid];
1600   int status;
1601 
1602 #ifdef KMP_DEBUG
1603   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1604 #endif
1605 
1606   KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
1607                 gtid, target_gtid));
1608   KMP_DEBUG_ASSERT(gtid != target_gtid);
1609 
1610   __kmp_suspend_initialize_thread(th);
1611 
1612   status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex);
1613   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1614 
1615   if (!flag) { // coming from __kmp_null_resume_wrapper
1616     flag = (C *)CCAST(void *, th->th.th_sleep_loc);
1617   }
1618 
1619   // First, check if the flag is null or its type has changed. If so, someone
1620   // else woke it up.
1621   if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
1622     // simply shows what
1623     // flag was cast to
1624     KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1625                  "awake: flag(%p)\n",
1626                  gtid, target_gtid, NULL));
1627     status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1628     KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1629     return;
1630   } else { // if multiple threads are sleeping, flag should be internally
1631     // referring to a specific thread here
1632     typename C::flag_t old_spin = flag->unset_sleeping();
1633     if (!flag->is_sleeping_val(old_spin)) {
1634       KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
1635                    "awake: flag(%p): "
1636                    "%u => %u\n",
1637                    gtid, target_gtid, flag->get(), old_spin, flag->load()));
1638       status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1639       KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1640       return;
1641     }
1642     KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset "
1643                  "sleep bit for flag's loc(%p): "
1644                  "%u => %u\n",
1645                  gtid, target_gtid, flag->get(), old_spin, flag->load()));
1646   }
1647   TCW_PTR(th->th.th_sleep_loc, NULL);
1648 
1649 #ifdef DEBUG_SUSPEND
1650   {
1651     char buffer[128];
1652     __kmp_print_cond(buffer, &th->th.th_suspend_cv);
1653     __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid,
1654                  target_gtid, buffer);
1655   }
1656 #endif
1657   status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond);
1658   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1659   status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex);
1660   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1661   KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
1662                 " for T#%d\n",
1663                 gtid, target_gtid));
1664 }
1665 
1666 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) {
1667   __kmp_resume_template(target_gtid, flag);
1668 }
1669 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) {
1670   __kmp_resume_template(target_gtid, flag);
1671 }
1672 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
1673   __kmp_resume_template(target_gtid, flag);
1674 }
1675 
1676 #if KMP_USE_MONITOR
1677 void __kmp_resume_monitor() {
1678   KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume);
1679   int status;
1680 #ifdef KMP_DEBUG
1681   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
1682   KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid,
1683                 KMP_GTID_MONITOR));
1684   KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR);
1685 #endif
1686   status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex);
1687   KMP_CHECK_SYSFAIL("pthread_mutex_lock", status);
1688 #ifdef DEBUG_SUSPEND
1689   {
1690     char buffer[128];
1691     __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond);
1692     __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid,
1693                  KMP_GTID_MONITOR, buffer);
1694   }
1695 #endif
1696   status = pthread_cond_signal(&__kmp_wait_cv.c_cond);
1697   KMP_CHECK_SYSFAIL("pthread_cond_signal", status);
1698   status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex);
1699   KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status);
1700   KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up"
1701                 " for T#%d\n",
1702                 gtid, KMP_GTID_MONITOR));
1703 }
1704 #endif // KMP_USE_MONITOR
1705 
1706 void __kmp_yield() { sched_yield(); }
1707 
1708 void __kmp_gtid_set_specific(int gtid) {
1709   if (__kmp_init_gtid) {
1710     int status;
1711     status = pthread_setspecific(__kmp_gtid_threadprivate_key,
1712                                  (void *)(intptr_t)(gtid + 1));
1713     KMP_CHECK_SYSFAIL("pthread_setspecific", status);
1714   } else {
1715     KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
1716   }
1717 }
1718 
1719 int __kmp_gtid_get_specific() {
1720   int gtid;
1721   if (!__kmp_init_gtid) {
1722     KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
1723                   "KMP_GTID_SHUTDOWN\n"));
1724     return KMP_GTID_SHUTDOWN;
1725   }
1726   gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key);
1727   if (gtid == 0) {
1728     gtid = KMP_GTID_DNE;
1729   } else {
1730     gtid--;
1731   }
1732   KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
1733                 __kmp_gtid_threadprivate_key, gtid));
1734   return gtid;
1735 }
1736 
1737 double __kmp_read_cpu_time(void) {
1738   /*clock_t   t;*/
1739   struct tms buffer;
1740 
1741   /*t =*/times(&buffer);
1742 
1743   return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC;
1744 }
1745 
1746 int __kmp_read_system_info(struct kmp_sys_info *info) {
1747   int status;
1748   struct rusage r_usage;
1749 
1750   memset(info, 0, sizeof(*info));
1751 
1752   status = getrusage(RUSAGE_SELF, &r_usage);
1753   KMP_CHECK_SYSFAIL_ERRNO("getrusage", status);
1754 
1755   // The maximum resident set size utilized (in kilobytes)
1756   info->maxrss = r_usage.ru_maxrss;
1757   // The number of page faults serviced without any I/O
1758   info->minflt = r_usage.ru_minflt;
1759   // The number of page faults serviced that required I/O
1760   info->majflt = r_usage.ru_majflt;
1761   // The number of times a process was "swapped" out of memory
1762   info->nswap = r_usage.ru_nswap;
1763   // The number of times the file system had to perform input
1764   info->inblock = r_usage.ru_inblock;
1765   // The number of times the file system had to perform output
1766   info->oublock = r_usage.ru_oublock;
1767   // The number of times a context switch was voluntarily
1768   info->nvcsw = r_usage.ru_nvcsw;
1769   // The number of times a context switch was forced
1770   info->nivcsw = r_usage.ru_nivcsw;
1771 
1772   return (status != 0);
1773 }
1774 
1775 void __kmp_read_system_time(double *delta) {
1776   double t_ns;
1777   struct timeval tval;
1778   struct timespec stop;
1779   int status;
1780 
1781   status = gettimeofday(&tval, NULL);
1782   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1783   TIMEVAL_TO_TIMESPEC(&tval, &stop);
1784   t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start);
1785   *delta = (t_ns * 1e-9);
1786 }
1787 
1788 void __kmp_clear_system_time(void) {
1789   struct timeval tval;
1790   int status;
1791   status = gettimeofday(&tval, NULL);
1792   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1793   TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start);
1794 }
1795 
1796 static int __kmp_get_xproc(void) {
1797 
1798   int r = 0;
1799 
1800 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD ||     \
1801         KMP_OS_OPENBSD || KMP_OS_HURD
1802 
1803   r = sysconf(_SC_NPROCESSORS_ONLN);
1804 
1805 #elif KMP_OS_DARWIN
1806 
1807   // Bug C77011 High "OpenMP Threads and number of active cores".
1808 
1809   // Find the number of available CPUs.
1810   kern_return_t rc;
1811   host_basic_info_data_t info;
1812   mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT;
1813   rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num);
1814   if (rc == 0 && num == HOST_BASIC_INFO_COUNT) {
1815     // Cannot use KA_TRACE() here because this code works before trace support
1816     // is initialized.
1817     r = info.avail_cpus;
1818   } else {
1819     KMP_WARNING(CantGetNumAvailCPU);
1820     KMP_INFORM(AssumedNumCPU);
1821   }
1822 
1823 #else
1824 
1825 #error "Unknown or unsupported OS."
1826 
1827 #endif
1828 
1829   return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */
1830 
1831 } // __kmp_get_xproc
1832 
1833 int __kmp_read_from_file(char const *path, char const *format, ...) {
1834   int result;
1835   va_list args;
1836 
1837   va_start(args, format);
1838   FILE *f = fopen(path, "rb");
1839   if (f == NULL)
1840     return 0;
1841   result = vfscanf(f, format, args);
1842   fclose(f);
1843 
1844   return result;
1845 }
1846 
1847 void __kmp_runtime_initialize(void) {
1848   int status;
1849   pthread_mutexattr_t mutex_attr;
1850   pthread_condattr_t cond_attr;
1851 
1852   if (__kmp_init_runtime) {
1853     return;
1854   }
1855 
1856 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1857   if (!__kmp_cpuinfo.initialized) {
1858     __kmp_query_cpuid(&__kmp_cpuinfo);
1859   }
1860 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
1861 
1862   __kmp_xproc = __kmp_get_xproc();
1863 
1864 #if ! KMP_32_BIT_ARCH
1865   struct rlimit rlim;
1866   // read stack size of calling thread, save it as default for worker threads;
1867   // this should be done before reading environment variables
1868   status = getrlimit(RLIMIT_STACK, &rlim);
1869   if (status == 0) { // success?
1870     __kmp_stksize = rlim.rlim_cur;
1871     __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed
1872   }
1873 #endif /* KMP_32_BIT_ARCH */
1874 
1875   if (sysconf(_SC_THREADS)) {
1876 
1877     /* Query the maximum number of threads */
1878     __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX);
1879     if (__kmp_sys_max_nth == -1) {
1880       /* Unlimited threads for NPTL */
1881       __kmp_sys_max_nth = INT_MAX;
1882     } else if (__kmp_sys_max_nth <= 1) {
1883       /* Can't tell, just use PTHREAD_THREADS_MAX */
1884       __kmp_sys_max_nth = KMP_MAX_NTH;
1885     }
1886 
1887     /* Query the minimum stack size */
1888     __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN);
1889     if (__kmp_sys_min_stksize <= 1) {
1890       __kmp_sys_min_stksize = KMP_MIN_STKSIZE;
1891     }
1892   }
1893 
1894   /* Set up minimum number of threads to switch to TLS gtid */
1895   __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
1896 
1897   status = pthread_key_create(&__kmp_gtid_threadprivate_key,
1898                               __kmp_internal_end_dest);
1899   KMP_CHECK_SYSFAIL("pthread_key_create", status);
1900   status = pthread_mutexattr_init(&mutex_attr);
1901   KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status);
1902   status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr);
1903   KMP_CHECK_SYSFAIL("pthread_mutex_init", status);
1904   status = pthread_condattr_init(&cond_attr);
1905   KMP_CHECK_SYSFAIL("pthread_condattr_init", status);
1906   status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr);
1907   KMP_CHECK_SYSFAIL("pthread_cond_init", status);
1908 #if USE_ITT_BUILD
1909   __kmp_itt_initialize();
1910 #endif /* USE_ITT_BUILD */
1911 
1912   __kmp_init_runtime = TRUE;
1913 }
1914 
1915 void __kmp_runtime_destroy(void) {
1916   int status;
1917 
1918   if (!__kmp_init_runtime) {
1919     return; // Nothing to do.
1920   }
1921 
1922 #if USE_ITT_BUILD
1923   __kmp_itt_destroy();
1924 #endif /* USE_ITT_BUILD */
1925 
1926   status = pthread_key_delete(__kmp_gtid_threadprivate_key);
1927   KMP_CHECK_SYSFAIL("pthread_key_delete", status);
1928 
1929   status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex);
1930   if (status != 0 && status != EBUSY) {
1931     KMP_SYSFAIL("pthread_mutex_destroy", status);
1932   }
1933   status = pthread_cond_destroy(&__kmp_wait_cv.c_cond);
1934   if (status != 0 && status != EBUSY) {
1935     KMP_SYSFAIL("pthread_cond_destroy", status);
1936   }
1937 #if KMP_AFFINITY_SUPPORTED
1938   __kmp_affinity_uninitialize();
1939 #endif
1940 
1941   __kmp_init_runtime = FALSE;
1942 }
1943 
1944 /* Put the thread to sleep for a time period */
1945 /* NOTE: not currently used anywhere */
1946 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); }
1947 
1948 /* Calculate the elapsed wall clock time for the user */
1949 void __kmp_elapsed(double *t) {
1950   int status;
1951 #ifdef FIX_SGI_CLOCK
1952   struct timespec ts;
1953 
1954   status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts);
1955   KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status);
1956   *t =
1957       (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec;
1958 #else
1959   struct timeval tv;
1960 
1961   status = gettimeofday(&tv, NULL);
1962   KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status);
1963   *t =
1964       (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec;
1965 #endif
1966 }
1967 
1968 /* Calculate the elapsed wall clock tick for the user */
1969 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; }
1970 
1971 /* Return the current time stamp in nsec */
1972 kmp_uint64 __kmp_now_nsec() {
1973   struct timeval t;
1974   gettimeofday(&t, NULL);
1975   kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec +
1976                     (kmp_uint64)1000 * (kmp_uint64)t.tv_usec;
1977   return nsec;
1978 }
1979 
1980 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1981 /* Measure clock ticks per millisecond */
1982 void __kmp_initialize_system_tick() {
1983   kmp_uint64 now, nsec2, diff;
1984   kmp_uint64 delay = 100000; // 50~100 usec on most machines.
1985   kmp_uint64 nsec = __kmp_now_nsec();
1986   kmp_uint64 goal = __kmp_hardware_timestamp() + delay;
1987   while ((now = __kmp_hardware_timestamp()) < goal)
1988     ;
1989   nsec2 = __kmp_now_nsec();
1990   diff = nsec2 - nsec;
1991   if (diff > 0) {
1992     kmp_uint64 tpms = (kmp_uint64)(1e6 * (delay + (now - goal)) / diff);
1993     if (tpms > 0)
1994       __kmp_ticks_per_msec = tpms;
1995   }
1996 }
1997 #endif
1998 
1999 /* Determine whether the given address is mapped into the current address
2000    space. */
2001 
2002 int __kmp_is_address_mapped(void *addr) {
2003 
2004   int found = 0;
2005   int rc;
2006 
2007 #if KMP_OS_LINUX || KMP_OS_HURD
2008 
2009   /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the address
2010      ranges mapped into the address space. */
2011 
2012   char *name = __kmp_str_format("/proc/%d/maps", getpid());
2013   FILE *file = NULL;
2014 
2015   file = fopen(name, "r");
2016   KMP_ASSERT(file != NULL);
2017 
2018   for (;;) {
2019 
2020     void *beginning = NULL;
2021     void *ending = NULL;
2022     char perms[5];
2023 
2024     rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms);
2025     if (rc == EOF) {
2026       break;
2027     }
2028     KMP_ASSERT(rc == 3 &&
2029                KMP_STRLEN(perms) == 4); // Make sure all fields are read.
2030 
2031     // Ending address is not included in the region, but beginning is.
2032     if ((addr >= beginning) && (addr < ending)) {
2033       perms[2] = 0; // 3th and 4th character does not matter.
2034       if (strcmp(perms, "rw") == 0) {
2035         // Memory we are looking for should be readable and writable.
2036         found = 1;
2037       }
2038       break;
2039     }
2040   }
2041 
2042   // Free resources.
2043   fclose(file);
2044   KMP_INTERNAL_FREE(name);
2045 #elif KMP_OS_FREEBSD
2046   char *buf;
2047   size_t lstsz;
2048   int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()};
2049   rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0);
2050   if (rc < 0)
2051      return 0;
2052   // We pass from number of vm entry's semantic
2053   // to size of whole entry map list.
2054   lstsz = lstsz * 4 / 3;
2055   buf = reinterpret_cast<char *>(kmpc_malloc(lstsz));
2056   rc = sysctl(mib, 4, buf, &lstsz, NULL, 0);
2057   if (rc < 0) {
2058      kmpc_free(buf);
2059      return 0;
2060   }
2061 
2062   char *lw = buf;
2063   char *up = buf + lstsz;
2064 
2065   while (lw < up) {
2066       struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw);
2067       size_t cursz = cur->kve_structsize;
2068       if (cursz == 0)
2069           break;
2070       void *start = reinterpret_cast<void *>(cur->kve_start);
2071       void *end = reinterpret_cast<void *>(cur->kve_end);
2072       // Readable/Writable addresses within current map entry
2073       if ((addr >= start) && (addr < end)) {
2074           if ((cur->kve_protection & KVME_PROT_READ) != 0 &&
2075               (cur->kve_protection & KVME_PROT_WRITE) != 0) {
2076               found = 1;
2077               break;
2078           }
2079       }
2080       lw += cursz;
2081   }
2082   kmpc_free(buf);
2083 
2084 #elif KMP_OS_DARWIN
2085 
2086   /* On OS X*, /proc pseudo filesystem is not available. Try to read memory
2087      using vm interface. */
2088 
2089   int buffer;
2090   vm_size_t count;
2091   rc = vm_read_overwrite(
2092       mach_task_self(), // Task to read memory of.
2093       (vm_address_t)(addr), // Address to read from.
2094       1, // Number of bytes to be read.
2095       (vm_address_t)(&buffer), // Address of buffer to save read bytes in.
2096       &count // Address of var to save number of read bytes in.
2097       );
2098   if (rc == 0) {
2099     // Memory successfully read.
2100     found = 1;
2101   }
2102 
2103 #elif KMP_OS_NETBSD
2104 
2105   int mib[5];
2106   mib[0] = CTL_VM;
2107   mib[1] = VM_PROC;
2108   mib[2] = VM_PROC_MAP;
2109   mib[3] = getpid();
2110   mib[4] = sizeof(struct kinfo_vmentry);
2111 
2112   size_t size;
2113   rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0);
2114   KMP_ASSERT(!rc);
2115   KMP_ASSERT(size);
2116 
2117   size = size * 4 / 3;
2118   struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size);
2119   KMP_ASSERT(kiv);
2120 
2121   rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0);
2122   KMP_ASSERT(!rc);
2123   KMP_ASSERT(size);
2124 
2125   for (size_t i = 0; i < size; i++) {
2126     if (kiv[i].kve_start >= (uint64_t)addr &&
2127         kiv[i].kve_end <= (uint64_t)addr) {
2128       found = 1;
2129       break;
2130     }
2131   }
2132   KMP_INTERNAL_FREE(kiv);
2133 #elif KMP_OS_OPENBSD
2134 
2135   int mib[3];
2136   mib[0] = CTL_KERN;
2137   mib[1] = KERN_PROC_VMMAP;
2138   mib[2] = getpid();
2139 
2140   size_t size;
2141   uint64_t end;
2142   rc = sysctl(mib, 3, NULL, &size, NULL, 0);
2143   KMP_ASSERT(!rc);
2144   KMP_ASSERT(size);
2145   end = size;
2146 
2147   struct kinfo_vmentry kiv = {.kve_start = 0};
2148 
2149   while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) {
2150     KMP_ASSERT(size);
2151     if (kiv.kve_end == end)
2152       break;
2153 
2154     if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) {
2155       found = 1;
2156       break;
2157     }
2158     kiv.kve_start += 1;
2159   }
2160 #elif KMP_OS_DRAGONFLY
2161 
2162   // FIXME(DragonFly): Implement this
2163   found = 1;
2164 
2165 #else
2166 
2167 #error "Unknown or unsupported OS"
2168 
2169 #endif
2170 
2171   return found;
2172 
2173 } // __kmp_is_address_mapped
2174 
2175 #ifdef USE_LOAD_BALANCE
2176 
2177 #if KMP_OS_DARWIN || KMP_OS_NETBSD
2178 
2179 // The function returns the rounded value of the system load average
2180 // during given time interval which depends on the value of
2181 // __kmp_load_balance_interval variable (default is 60 sec, other values
2182 // may be 300 sec or 900 sec).
2183 // It returns -1 in case of error.
2184 int __kmp_get_load_balance(int max) {
2185   double averages[3];
2186   int ret_avg = 0;
2187 
2188   int res = getloadavg(averages, 3);
2189 
2190   // Check __kmp_load_balance_interval to determine which of averages to use.
2191   // getloadavg() may return the number of samples less than requested that is
2192   // less than 3.
2193   if (__kmp_load_balance_interval < 180 && (res >= 1)) {
2194     ret_avg = averages[0]; // 1 min
2195   } else if ((__kmp_load_balance_interval >= 180 &&
2196               __kmp_load_balance_interval < 600) &&
2197              (res >= 2)) {
2198     ret_avg = averages[1]; // 5 min
2199   } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) {
2200     ret_avg = averages[2]; // 15 min
2201   } else { // Error occurred
2202     return -1;
2203   }
2204 
2205   return ret_avg;
2206 }
2207 
2208 #else // Linux* OS
2209 
2210 // The function returns number of running (not sleeping) threads, or -1 in case
2211 // of error. Error could be reported if Linux* OS kernel too old (without
2212 // "/proc" support). Counting running threads stops if max running threads
2213 // encountered.
2214 int __kmp_get_load_balance(int max) {
2215   static int permanent_error = 0;
2216   static int glb_running_threads = 0; // Saved count of the running threads for
2217   // the thread balance algorithm
2218   static double glb_call_time = 0; /* Thread balance algorithm call time */
2219 
2220   int running_threads = 0; // Number of running threads in the system.
2221 
2222   DIR *proc_dir = NULL; // Handle of "/proc/" directory.
2223   struct dirent *proc_entry = NULL;
2224 
2225   kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path.
2226   DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory.
2227   struct dirent *task_entry = NULL;
2228   int task_path_fixed_len;
2229 
2230   kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path.
2231   int stat_file = -1;
2232   int stat_path_fixed_len;
2233 
2234   int total_processes = 0; // Total number of processes in system.
2235   int total_threads = 0; // Total number of threads in system.
2236 
2237   double call_time = 0.0;
2238 
2239   __kmp_str_buf_init(&task_path);
2240   __kmp_str_buf_init(&stat_path);
2241 
2242   __kmp_elapsed(&call_time);
2243 
2244   if (glb_call_time &&
2245       (call_time - glb_call_time < __kmp_load_balance_interval)) {
2246     running_threads = glb_running_threads;
2247     goto finish;
2248   }
2249 
2250   glb_call_time = call_time;
2251 
2252   // Do not spend time on scanning "/proc/" if we have a permanent error.
2253   if (permanent_error) {
2254     running_threads = -1;
2255     goto finish;
2256   }
2257 
2258   if (max <= 0) {
2259     max = INT_MAX;
2260   }
2261 
2262   // Open "/proc/" directory.
2263   proc_dir = opendir("/proc");
2264   if (proc_dir == NULL) {
2265     // Cannot open "/prroc/". Probably the kernel does not support it. Return an
2266     // error now and in subsequent calls.
2267     running_threads = -1;
2268     permanent_error = 1;
2269     goto finish;
2270   }
2271 
2272   // Initialize fixed part of task_path. This part will not change.
2273   __kmp_str_buf_cat(&task_path, "/proc/", 6);
2274   task_path_fixed_len = task_path.used; // Remember number of used characters.
2275 
2276   proc_entry = readdir(proc_dir);
2277   while (proc_entry != NULL) {
2278     // Proc entry is a directory and name starts with a digit. Assume it is a
2279     // process' directory.
2280     if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) {
2281 
2282       ++total_processes;
2283       // Make sure init process is the very first in "/proc", so we can replace
2284       // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes ==
2285       // 1. We are going to check that total_processes == 1 => d_name == "1" is
2286       // true (where "=>" is implication). Since C++ does not have => operator,
2287       // let us replace it with its equivalent: a => b == ! a || b.
2288       KMP_DEBUG_ASSERT(total_processes != 1 ||
2289                        strcmp(proc_entry->d_name, "1") == 0);
2290 
2291       // Construct task_path.
2292       task_path.used = task_path_fixed_len; // Reset task_path to "/proc/".
2293       __kmp_str_buf_cat(&task_path, proc_entry->d_name,
2294                         KMP_STRLEN(proc_entry->d_name));
2295       __kmp_str_buf_cat(&task_path, "/task", 5);
2296 
2297       task_dir = opendir(task_path.str);
2298       if (task_dir == NULL) {
2299         // Process can finish between reading "/proc/" directory entry and
2300         // opening process' "task/" directory. So, in general case we should not
2301         // complain, but have to skip this process and read the next one. But on
2302         // systems with no "task/" support we will spend lot of time to scan
2303         // "/proc/" tree again and again without any benefit. "init" process
2304         // (its pid is 1) should exist always, so, if we cannot open
2305         // "/proc/1/task/" directory, it means "task/" is not supported by
2306         // kernel. Report an error now and in the future.
2307         if (strcmp(proc_entry->d_name, "1") == 0) {
2308           running_threads = -1;
2309           permanent_error = 1;
2310           goto finish;
2311         }
2312       } else {
2313         // Construct fixed part of stat file path.
2314         __kmp_str_buf_clear(&stat_path);
2315         __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used);
2316         __kmp_str_buf_cat(&stat_path, "/", 1);
2317         stat_path_fixed_len = stat_path.used;
2318 
2319         task_entry = readdir(task_dir);
2320         while (task_entry != NULL) {
2321           // It is a directory and name starts with a digit.
2322           if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) {
2323             ++total_threads;
2324 
2325             // Construct complete stat file path. Easiest way would be:
2326             //  __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str,
2327             //  task_entry->d_name );
2328             // but seriae of __kmp_str_buf_cat works a bit faster.
2329             stat_path.used =
2330                 stat_path_fixed_len; // Reset stat path to its fixed part.
2331             __kmp_str_buf_cat(&stat_path, task_entry->d_name,
2332                               KMP_STRLEN(task_entry->d_name));
2333             __kmp_str_buf_cat(&stat_path, "/stat", 5);
2334 
2335             // Note: Low-level API (open/read/close) is used. High-level API
2336             // (fopen/fclose)  works ~ 30 % slower.
2337             stat_file = open(stat_path.str, O_RDONLY);
2338             if (stat_file == -1) {
2339               // We cannot report an error because task (thread) can terminate
2340               // just before reading this file.
2341             } else {
2342               /* Content of "stat" file looks like:
2343                  24285 (program) S ...
2344 
2345                  It is a single line (if program name does not include funny
2346                  symbols). First number is a thread id, then name of executable
2347                  file name in paretheses, then state of the thread. We need just
2348                  thread state.
2349 
2350                  Good news: Length of program name is 15 characters max. Longer
2351                  names are truncated.
2352 
2353                  Thus, we need rather short buffer: 15 chars for program name +
2354                  2 parenthesis, + 3 spaces + ~7 digits of pid = 37.
2355 
2356                  Bad news: Program name may contain special symbols like space,
2357                  closing parenthesis, or even new line. This makes parsing
2358                  "stat" file not 100 % reliable. In case of fanny program names
2359                  parsing may fail (report incorrect thread state).
2360 
2361                  Parsing "status" file looks more promissing (due to different
2362                  file structure and escaping special symbols) but reading and
2363                  parsing of "status" file works slower.
2364                   -- ln
2365               */
2366               char buffer[65];
2367               int len;
2368               len = read(stat_file, buffer, sizeof(buffer) - 1);
2369               if (len >= 0) {
2370                 buffer[len] = 0;
2371                 // Using scanf:
2372                 //     sscanf( buffer, "%*d (%*s) %c ", & state );
2373                 // looks very nice, but searching for a closing parenthesis
2374                 // works a bit faster.
2375                 char *close_parent = strstr(buffer, ") ");
2376                 if (close_parent != NULL) {
2377                   char state = *(close_parent + 2);
2378                   if (state == 'R') {
2379                     ++running_threads;
2380                     if (running_threads >= max) {
2381                       goto finish;
2382                     }
2383                   }
2384                 }
2385               }
2386               close(stat_file);
2387               stat_file = -1;
2388             }
2389           }
2390           task_entry = readdir(task_dir);
2391         }
2392         closedir(task_dir);
2393         task_dir = NULL;
2394       }
2395     }
2396     proc_entry = readdir(proc_dir);
2397   }
2398 
2399   // There _might_ be a timing hole where the thread executing this
2400   // code get skipped in the load balance, and running_threads is 0.
2401   // Assert in the debug builds only!!!
2402   KMP_DEBUG_ASSERT(running_threads > 0);
2403   if (running_threads <= 0) {
2404     running_threads = 1;
2405   }
2406 
2407 finish: // Clean up and exit.
2408   if (proc_dir != NULL) {
2409     closedir(proc_dir);
2410   }
2411   __kmp_str_buf_free(&task_path);
2412   if (task_dir != NULL) {
2413     closedir(task_dir);
2414   }
2415   __kmp_str_buf_free(&stat_path);
2416   if (stat_file != -1) {
2417     close(stat_file);
2418   }
2419 
2420   glb_running_threads = running_threads;
2421 
2422   return running_threads;
2423 
2424 } // __kmp_get_load_balance
2425 
2426 #endif // KMP_OS_DARWIN
2427 
2428 #endif // USE_LOAD_BALANCE
2429 
2430 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC ||                            \
2431       ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) ||                 \
2432       KMP_ARCH_PPC64 || KMP_ARCH_RISCV64)
2433 
2434 // we really only need the case with 1 argument, because CLANG always build
2435 // a struct of pointers to shared variables referenced in the outlined function
2436 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc,
2437                            void *p_argv[]
2438 #if OMPT_SUPPORT
2439                            ,
2440                            void **exit_frame_ptr
2441 #endif
2442                            ) {
2443 #if OMPT_SUPPORT
2444   *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0);
2445 #endif
2446 
2447   switch (argc) {
2448   default:
2449     fprintf(stderr, "Too many args to microtask: %d!\n", argc);
2450     fflush(stderr);
2451     exit(-1);
2452   case 0:
2453     (*pkfn)(&gtid, &tid);
2454     break;
2455   case 1:
2456     (*pkfn)(&gtid, &tid, p_argv[0]);
2457     break;
2458   case 2:
2459     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1]);
2460     break;
2461   case 3:
2462     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2]);
2463     break;
2464   case 4:
2465     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]);
2466     break;
2467   case 5:
2468     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]);
2469     break;
2470   case 6:
2471     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2472             p_argv[5]);
2473     break;
2474   case 7:
2475     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2476             p_argv[5], p_argv[6]);
2477     break;
2478   case 8:
2479     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2480             p_argv[5], p_argv[6], p_argv[7]);
2481     break;
2482   case 9:
2483     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2484             p_argv[5], p_argv[6], p_argv[7], p_argv[8]);
2485     break;
2486   case 10:
2487     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2488             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]);
2489     break;
2490   case 11:
2491     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2492             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]);
2493     break;
2494   case 12:
2495     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2496             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2497             p_argv[11]);
2498     break;
2499   case 13:
2500     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2501             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2502             p_argv[11], p_argv[12]);
2503     break;
2504   case 14:
2505     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2506             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2507             p_argv[11], p_argv[12], p_argv[13]);
2508     break;
2509   case 15:
2510     (*pkfn)(&gtid, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4],
2511             p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10],
2512             p_argv[11], p_argv[12], p_argv[13], p_argv[14]);
2513     break;
2514   }
2515 
2516   return 1;
2517 }
2518 
2519 #endif
2520 
2521 // end of file //
2522