LLVM OpenMP* Runtime Library
z_Windows_NT_util.cpp
1 /*
2  * z_Windows_NT_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_wait_release.h"
19 
20 /* This code is related to NtQuerySystemInformation() function. This function
21  is used in the Load balance algorithm for OMP_DYNAMIC=true to find the
22  number of running threads in the system. */
23 
24 #include <ntsecapi.h> // UNICODE_STRING
25 #include <ntstatus.h>
26 
27 enum SYSTEM_INFORMATION_CLASS {
28  SystemProcessInformation = 5
29 }; // SYSTEM_INFORMATION_CLASS
30 
31 struct CLIENT_ID {
32  HANDLE UniqueProcess;
33  HANDLE UniqueThread;
34 }; // struct CLIENT_ID
35 
36 enum THREAD_STATE {
37  StateInitialized,
38  StateReady,
39  StateRunning,
40  StateStandby,
41  StateTerminated,
42  StateWait,
43  StateTransition,
44  StateUnknown
45 }; // enum THREAD_STATE
46 
47 struct VM_COUNTERS {
48  SIZE_T PeakVirtualSize;
49  SIZE_T VirtualSize;
50  ULONG PageFaultCount;
51  SIZE_T PeakWorkingSetSize;
52  SIZE_T WorkingSetSize;
53  SIZE_T QuotaPeakPagedPoolUsage;
54  SIZE_T QuotaPagedPoolUsage;
55  SIZE_T QuotaPeakNonPagedPoolUsage;
56  SIZE_T QuotaNonPagedPoolUsage;
57  SIZE_T PagefileUsage;
58  SIZE_T PeakPagefileUsage;
59  SIZE_T PrivatePageCount;
60 }; // struct VM_COUNTERS
61 
62 struct SYSTEM_THREAD {
63  LARGE_INTEGER KernelTime;
64  LARGE_INTEGER UserTime;
65  LARGE_INTEGER CreateTime;
66  ULONG WaitTime;
67  LPVOID StartAddress;
68  CLIENT_ID ClientId;
69  DWORD Priority;
70  LONG BasePriority;
71  ULONG ContextSwitchCount;
72  THREAD_STATE State;
73  ULONG WaitReason;
74 }; // SYSTEM_THREAD
75 
76 KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, KernelTime) == 0);
77 #if KMP_ARCH_X86
78 KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, StartAddress) == 28);
79 KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, State) == 52);
80 #else
81 KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, StartAddress) == 32);
82 KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, State) == 68);
83 #endif
84 
85 struct SYSTEM_PROCESS_INFORMATION {
86  ULONG NextEntryOffset;
87  ULONG NumberOfThreads;
88  LARGE_INTEGER Reserved[3];
89  LARGE_INTEGER CreateTime;
90  LARGE_INTEGER UserTime;
91  LARGE_INTEGER KernelTime;
92  UNICODE_STRING ImageName;
93  DWORD BasePriority;
94  HANDLE ProcessId;
95  HANDLE ParentProcessId;
96  ULONG HandleCount;
97  ULONG Reserved2[2];
98  VM_COUNTERS VMCounters;
99  IO_COUNTERS IOCounters;
100  SYSTEM_THREAD Threads[1];
101 }; // SYSTEM_PROCESS_INFORMATION
102 typedef SYSTEM_PROCESS_INFORMATION *PSYSTEM_PROCESS_INFORMATION;
103 
104 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, NextEntryOffset) == 0);
105 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, CreateTime) == 32);
106 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, ImageName) == 56);
107 #if KMP_ARCH_X86
108 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, ProcessId) == 68);
109 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, HandleCount) == 76);
110 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, VMCounters) == 88);
111 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, IOCounters) == 136);
112 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, Threads) == 184);
113 #else
114 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, ProcessId) == 80);
115 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, HandleCount) == 96);
116 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, VMCounters) == 112);
117 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, IOCounters) == 208);
118 KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, Threads) == 256);
119 #endif
120 
121 typedef NTSTATUS(NTAPI *NtQuerySystemInformation_t)(SYSTEM_INFORMATION_CLASS,
122  PVOID, ULONG, PULONG);
123 NtQuerySystemInformation_t NtQuerySystemInformation = NULL;
124 
125 HMODULE ntdll = NULL;
126 
127 /* End of NtQuerySystemInformation()-related code */
128 
129 static HMODULE kernel32 = NULL;
130 
131 #if KMP_HANDLE_SIGNALS
132 typedef void (*sig_func_t)(int);
133 static sig_func_t __kmp_sighldrs[NSIG];
134 static int __kmp_siginstalled[NSIG];
135 #endif
136 
137 #if KMP_USE_MONITOR
138 static HANDLE __kmp_monitor_ev;
139 #endif
140 static kmp_int64 __kmp_win32_time;
141 double __kmp_win32_tick;
142 
143 int __kmp_init_runtime = FALSE;
144 CRITICAL_SECTION __kmp_win32_section;
145 
146 void __kmp_win32_mutex_init(kmp_win32_mutex_t *mx) {
147  InitializeCriticalSection(&mx->cs);
148 #if USE_ITT_BUILD
149  __kmp_itt_system_object_created(&mx->cs, "Critical Section");
150 #endif /* USE_ITT_BUILD */
151 }
152 
153 void __kmp_win32_mutex_destroy(kmp_win32_mutex_t *mx) {
154  DeleteCriticalSection(&mx->cs);
155 }
156 
157 void __kmp_win32_mutex_lock(kmp_win32_mutex_t *mx) {
158  EnterCriticalSection(&mx->cs);
159 }
160 
161 int __kmp_win32_mutex_trylock(kmp_win32_mutex_t *mx) {
162  return TryEnterCriticalSection(&mx->cs);
163 }
164 
165 void __kmp_win32_mutex_unlock(kmp_win32_mutex_t *mx) {
166  LeaveCriticalSection(&mx->cs);
167 }
168 
169 void __kmp_win32_cond_init(kmp_win32_cond_t *cv) {
170  cv->waiters_count_ = 0;
171  cv->wait_generation_count_ = 0;
172  cv->release_count_ = 0;
173 
174  /* Initialize the critical section */
175  __kmp_win32_mutex_init(&cv->waiters_count_lock_);
176 
177  /* Create a manual-reset event. */
178  cv->event_ = CreateEvent(NULL, // no security
179  TRUE, // manual-reset
180  FALSE, // non-signaled initially
181  NULL); // unnamed
182 #if USE_ITT_BUILD
183  __kmp_itt_system_object_created(cv->event_, "Event");
184 #endif /* USE_ITT_BUILD */
185 }
186 
187 void __kmp_win32_cond_destroy(kmp_win32_cond_t *cv) {
188  __kmp_win32_mutex_destroy(&cv->waiters_count_lock_);
189  __kmp_free_handle(cv->event_);
190  memset(cv, '\0', sizeof(*cv));
191 }
192 
193 /* TODO associate cv with a team instead of a thread so as to optimize
194  the case where we wake up a whole team */
195 
196 template <class C>
197 static void __kmp_win32_cond_wait(kmp_win32_cond_t *cv, kmp_win32_mutex_t *mx,
198  kmp_info_t *th, C *flag) {
199  int my_generation;
200  int last_waiter;
201 
202  /* Avoid race conditions */
203  __kmp_win32_mutex_lock(&cv->waiters_count_lock_);
204 
205  /* Increment count of waiters */
206  cv->waiters_count_++;
207 
208  /* Store current generation in our activation record. */
209  my_generation = cv->wait_generation_count_;
210 
211  __kmp_win32_mutex_unlock(&cv->waiters_count_lock_);
212  __kmp_win32_mutex_unlock(mx);
213 
214  for (;;) {
215  int wait_done = 0;
216  DWORD res, timeout = 5000; // just tried to quess an appropriate number
217  /* Wait until the event is signaled */
218  res = WaitForSingleObject(cv->event_, timeout);
219 
220  if (res == WAIT_OBJECT_0) {
221  // event signaled
222  __kmp_win32_mutex_lock(&cv->waiters_count_lock_);
223  /* Exit the loop when the <cv->event_> is signaled and there are still
224  waiting threads from this <wait_generation> that haven't been released
225  from this wait yet. */
226  wait_done = (cv->release_count_ > 0) &&
227  (cv->wait_generation_count_ != my_generation);
228  __kmp_win32_mutex_unlock(&cv->waiters_count_lock_);
229  } else if (res == WAIT_TIMEOUT || res == WAIT_FAILED) {
230  // check if the flag and cv counters are in consistent state
231  // as MS sent us debug dump whith inconsistent state of data
232  __kmp_win32_mutex_lock(mx);
233  typename C::flag_t old_f = flag->set_sleeping();
234  if (!flag->done_check_val(old_f & ~KMP_BARRIER_SLEEP_STATE)) {
235  __kmp_win32_mutex_unlock(mx);
236  continue;
237  }
238  // condition fulfilled, exiting
239  old_f = flag->unset_sleeping();
240  KMP_DEBUG_ASSERT(old_f & KMP_BARRIER_SLEEP_STATE);
241  TCW_PTR(th->th.th_sleep_loc, NULL);
242  KF_TRACE(50, ("__kmp_win32_cond_wait: exiting, condition "
243  "fulfilled: flag's loc(%p): %u => %u\n",
244  flag->get(), old_f, *(flag->get())));
245 
246  __kmp_win32_mutex_lock(&cv->waiters_count_lock_);
247  KMP_DEBUG_ASSERT(cv->waiters_count_ > 0);
248  cv->release_count_ = cv->waiters_count_;
249  cv->wait_generation_count_++;
250  wait_done = 1;
251  __kmp_win32_mutex_unlock(&cv->waiters_count_lock_);
252 
253  __kmp_win32_mutex_unlock(mx);
254  }
255  /* there used to be a semicolon after the if statement, it looked like a
256  bug, so i removed it */
257  if (wait_done)
258  break;
259  }
260 
261  __kmp_win32_mutex_lock(mx);
262  __kmp_win32_mutex_lock(&cv->waiters_count_lock_);
263 
264  cv->waiters_count_--;
265  cv->release_count_--;
266 
267  last_waiter = (cv->release_count_ == 0);
268 
269  __kmp_win32_mutex_unlock(&cv->waiters_count_lock_);
270 
271  if (last_waiter) {
272  /* We're the last waiter to be notified, so reset the manual event. */
273  ResetEvent(cv->event_);
274  }
275 }
276 
277 void __kmp_win32_cond_broadcast(kmp_win32_cond_t *cv) {
278  __kmp_win32_mutex_lock(&cv->waiters_count_lock_);
279 
280  if (cv->waiters_count_ > 0) {
281  SetEvent(cv->event_);
282  /* Release all the threads in this generation. */
283 
284  cv->release_count_ = cv->waiters_count_;
285 
286  /* Start a new generation. */
287  cv->wait_generation_count_++;
288  }
289 
290  __kmp_win32_mutex_unlock(&cv->waiters_count_lock_);
291 }
292 
293 void __kmp_win32_cond_signal(kmp_win32_cond_t *cv) {
294  __kmp_win32_cond_broadcast(cv);
295 }
296 
297 void __kmp_enable(int new_state) {
298  if (__kmp_init_runtime)
299  LeaveCriticalSection(&__kmp_win32_section);
300 }
301 
302 void __kmp_disable(int *old_state) {
303  *old_state = 0;
304 
305  if (__kmp_init_runtime)
306  EnterCriticalSection(&__kmp_win32_section);
307 }
308 
309 void __kmp_suspend_initialize(void) { /* do nothing */
310 }
311 
312 void __kmp_suspend_initialize_thread(kmp_info_t *th) {
313  if (!TCR_4(th->th.th_suspend_init)) {
314  /* this means we haven't initialized the suspension pthread objects for this
315  thread in this instance of the process */
316  __kmp_win32_cond_init(&th->th.th_suspend_cv);
317  __kmp_win32_mutex_init(&th->th.th_suspend_mx);
318  TCW_4(th->th.th_suspend_init, TRUE);
319  }
320 }
321 
322 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
323  if (TCR_4(th->th.th_suspend_init)) {
324  /* this means we have initialize the suspension pthread objects for this
325  thread in this instance of the process */
326  __kmp_win32_cond_destroy(&th->th.th_suspend_cv);
327  __kmp_win32_mutex_destroy(&th->th.th_suspend_mx);
328  TCW_4(th->th.th_suspend_init, FALSE);
329  }
330 }
331 
332 int __kmp_try_suspend_mx(kmp_info_t *th) {
333  return __kmp_win32_mutex_trylock(&th->th.th_suspend_mx);
334 }
335 
336 void __kmp_lock_suspend_mx(kmp_info_t *th) {
337  __kmp_win32_mutex_lock(&th->th.th_suspend_mx);
338 }
339 
340 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
341  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
342 }
343 
344 /* This routine puts the calling thread to sleep after setting the
345  sleep bit for the indicated flag variable to true. */
346 template <class C>
347 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
348  kmp_info_t *th = __kmp_threads[th_gtid];
349  int status;
350  typename C::flag_t old_spin;
351 
352  KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag's loc(%p)\n",
353  th_gtid, flag->get()));
354 
355  __kmp_suspend_initialize_thread(th);
356  __kmp_win32_mutex_lock(&th->th.th_suspend_mx);
357 
358  KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for flag's"
359  " loc(%p)\n",
360  th_gtid, flag->get()));
361 
362  /* TODO: shouldn't this use release semantics to ensure that
363  __kmp_suspend_initialize_thread gets called first? */
364  old_spin = flag->set_sleeping();
365 #if OMP_50_ENABLED
366  if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
367  __kmp_pause_status != kmp_soft_paused) {
368  flag->unset_sleeping();
369  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
370  return;
371  }
372 #endif
373 
374  KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for flag's"
375  " loc(%p)==%d\n",
376  th_gtid, flag->get(), *(flag->get())));
377 
378  if (flag->done_check_val(old_spin)) {
379  old_spin = flag->unset_sleeping();
380  KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
381  "for flag's loc(%p)\n",
382  th_gtid, flag->get()));
383  } else {
384 #ifdef DEBUG_SUSPEND
385  __kmp_suspend_count++;
386 #endif
387  /* Encapsulate in a loop as the documentation states that this may "with
388  low probability" return when the condition variable has not been signaled
389  or broadcast */
390  int deactivated = FALSE;
391  TCW_PTR(th->th.th_sleep_loc, (void *)flag);
392  while (flag->is_sleeping()) {
393  KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
394  "kmp_win32_cond_wait()\n",
395  th_gtid));
396  // Mark the thread as no longer active (only in the first iteration of the
397  // loop).
398  if (!deactivated) {
399  th->th.th_active = FALSE;
400  if (th->th.th_active_in_pool) {
401  th->th.th_active_in_pool = FALSE;
402  KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
403  KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
404  }
405  deactivated = TRUE;
406  __kmp_win32_cond_wait(&th->th.th_suspend_cv, &th->th.th_suspend_mx, th,
407  flag);
408  } else {
409  __kmp_win32_cond_wait(&th->th.th_suspend_cv, &th->th.th_suspend_mx, th,
410  flag);
411  }
412 
413 #ifdef KMP_DEBUG
414  if (flag->is_sleeping()) {
415  KF_TRACE(100,
416  ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
417  }
418 #endif /* KMP_DEBUG */
419 
420  } // while
421 
422  // Mark the thread as active again (if it was previous marked as inactive)
423  if (deactivated) {
424  th->th.th_active = TRUE;
425  if (TCR_4(th->th.th_in_pool)) {
426  KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
427  th->th.th_active_in_pool = TRUE;
428  }
429  }
430  }
431 
432  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
433 
434  KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
435 }
436 
437 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) {
438  __kmp_suspend_template(th_gtid, flag);
439 }
440 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) {
441  __kmp_suspend_template(th_gtid, flag);
442 }
443 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
444  __kmp_suspend_template(th_gtid, flag);
445 }
446 
447 /* This routine signals the thread specified by target_gtid to wake up
448  after setting the sleep bit indicated by the flag argument to FALSE */
449 template <class C>
450 static inline void __kmp_resume_template(int target_gtid, C *flag) {
451  kmp_info_t *th = __kmp_threads[target_gtid];
452  int status;
453 
454 #ifdef KMP_DEBUG
455  int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
456 #endif
457 
458  KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
459  gtid, target_gtid));
460 
461  __kmp_suspend_initialize_thread(th);
462  __kmp_win32_mutex_lock(&th->th.th_suspend_mx);
463 
464  if (!flag) { // coming from __kmp_null_resume_wrapper
465  flag = (C *)th->th.th_sleep_loc;
466  }
467 
468  // First, check if the flag is null or its type has changed. If so, someone
469  // else woke it up.
470  if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
471  // simply shows what
472  // flag was cast to
473  KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
474  "awake: flag's loc(%p)\n",
475  gtid, target_gtid, NULL));
476  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
477  return;
478  } else {
479  typename C::flag_t old_spin = flag->unset_sleeping();
480  if (!flag->is_sleeping_val(old_spin)) {
481  KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
482  "awake: flag's loc(%p): %u => %u\n",
483  gtid, target_gtid, flag->get(), old_spin, *(flag->get())));
484  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
485  return;
486  }
487  }
488  TCW_PTR(th->th.th_sleep_loc, NULL);
489  KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset sleep "
490  "bit for flag's loc(%p)\n",
491  gtid, target_gtid, flag->get()));
492 
493  __kmp_win32_cond_signal(&th->th.th_suspend_cv);
494  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
495 
496  KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
497  " for T#%d\n",
498  gtid, target_gtid));
499 }
500 
501 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) {
502  __kmp_resume_template(target_gtid, flag);
503 }
504 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) {
505  __kmp_resume_template(target_gtid, flag);
506 }
507 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
508  __kmp_resume_template(target_gtid, flag);
509 }
510 
511 void __kmp_yield() { Sleep(0); }
512 
513 void __kmp_gtid_set_specific(int gtid) {
514  if (__kmp_init_gtid) {
515  KA_TRACE(50, ("__kmp_gtid_set_specific: T#%d key:%d\n", gtid,
516  __kmp_gtid_threadprivate_key));
517  if (!TlsSetValue(__kmp_gtid_threadprivate_key, (LPVOID)(gtid + 1)))
518  KMP_FATAL(TLSSetValueFailed);
519  } else {
520  KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
521  }
522 }
523 
524 int __kmp_gtid_get_specific() {
525  int gtid;
526  if (!__kmp_init_gtid) {
527  KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
528  "KMP_GTID_SHUTDOWN\n"));
529  return KMP_GTID_SHUTDOWN;
530  }
531  gtid = (int)(kmp_intptr_t)TlsGetValue(__kmp_gtid_threadprivate_key);
532  if (gtid == 0) {
533  gtid = KMP_GTID_DNE;
534  } else {
535  gtid--;
536  }
537  KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
538  __kmp_gtid_threadprivate_key, gtid));
539  return gtid;
540 }
541 
542 void __kmp_affinity_bind_thread(int proc) {
543  if (__kmp_num_proc_groups > 1) {
544  // Form the GROUP_AFFINITY struct directly, rather than filling
545  // out a bit vector and calling __kmp_set_system_affinity().
546  GROUP_AFFINITY ga;
547  KMP_DEBUG_ASSERT((proc >= 0) && (proc < (__kmp_num_proc_groups * CHAR_BIT *
548  sizeof(DWORD_PTR))));
549  ga.Group = proc / (CHAR_BIT * sizeof(DWORD_PTR));
550  ga.Mask = (unsigned long long)1 << (proc % (CHAR_BIT * sizeof(DWORD_PTR)));
551  ga.Reserved[0] = ga.Reserved[1] = ga.Reserved[2] = 0;
552 
553  KMP_DEBUG_ASSERT(__kmp_SetThreadGroupAffinity != NULL);
554  if (__kmp_SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL) == 0) {
555  DWORD error = GetLastError();
556  if (__kmp_affinity_verbose) { // AC: continue silently if not verbose
557  kmp_msg_t err_code = KMP_ERR(error);
558  __kmp_msg(kmp_ms_warning, KMP_MSG(CantSetThreadAffMask), err_code,
559  __kmp_msg_null);
560  if (__kmp_generate_warnings == kmp_warnings_off) {
561  __kmp_str_free(&err_code.str);
562  }
563  }
564  }
565  } else {
566  kmp_affin_mask_t *mask;
567  KMP_CPU_ALLOC_ON_STACK(mask);
568  KMP_CPU_ZERO(mask);
569  KMP_CPU_SET(proc, mask);
570  __kmp_set_system_affinity(mask, TRUE);
571  KMP_CPU_FREE_FROM_STACK(mask);
572  }
573 }
574 
575 void __kmp_affinity_determine_capable(const char *env_var) {
576 // All versions of Windows* OS (since Win '95) support SetThreadAffinityMask().
577 
578 #if KMP_GROUP_AFFINITY
579  KMP_AFFINITY_ENABLE(__kmp_num_proc_groups * sizeof(DWORD_PTR));
580 #else
581  KMP_AFFINITY_ENABLE(sizeof(DWORD_PTR));
582 #endif
583 
584  KA_TRACE(10, ("__kmp_affinity_determine_capable: "
585  "Windows* OS affinity interface functional (mask size = "
586  "%" KMP_SIZE_T_SPEC ").\n",
587  __kmp_affin_mask_size));
588 }
589 
590 double __kmp_read_cpu_time(void) {
591  FILETIME CreationTime, ExitTime, KernelTime, UserTime;
592  int status;
593  double cpu_time;
594 
595  cpu_time = 0;
596 
597  status = GetProcessTimes(GetCurrentProcess(), &CreationTime, &ExitTime,
598  &KernelTime, &UserTime);
599 
600  if (status) {
601  double sec = 0;
602 
603  sec += KernelTime.dwHighDateTime;
604  sec += UserTime.dwHighDateTime;
605 
606  /* Shift left by 32 bits */
607  sec *= (double)(1 << 16) * (double)(1 << 16);
608 
609  sec += KernelTime.dwLowDateTime;
610  sec += UserTime.dwLowDateTime;
611 
612  cpu_time += (sec * 100.0) / KMP_NSEC_PER_SEC;
613  }
614 
615  return cpu_time;
616 }
617 
618 int __kmp_read_system_info(struct kmp_sys_info *info) {
619  info->maxrss = 0; /* the maximum resident set size utilized (in kilobytes) */
620  info->minflt = 0; /* the number of page faults serviced without any I/O */
621  info->majflt = 0; /* the number of page faults serviced that required I/O */
622  info->nswap = 0; // the number of times a process was "swapped" out of memory
623  info->inblock = 0; // the number of times the file system had to perform input
624  info->oublock = 0; // number of times the file system had to perform output
625  info->nvcsw = 0; /* the number of times a context switch was voluntarily */
626  info->nivcsw = 0; /* the number of times a context switch was forced */
627 
628  return 1;
629 }
630 
631 void __kmp_runtime_initialize(void) {
632  SYSTEM_INFO info;
633  kmp_str_buf_t path;
634  UINT path_size;
635 
636  if (__kmp_init_runtime) {
637  return;
638  }
639 
640 #if KMP_DYNAMIC_LIB
641  /* Pin dynamic library for the lifetime of application */
642  {
643  // First, turn off error message boxes
644  UINT err_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
645  HMODULE h;
646  BOOL ret = GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS |
647  GET_MODULE_HANDLE_EX_FLAG_PIN,
648  (LPCTSTR)&__kmp_serial_initialize, &h);
649  KMP_DEBUG_ASSERT2(h && ret, "OpenMP RTL cannot find itself loaded");
650  SetErrorMode(err_mode); // Restore error mode
651  KA_TRACE(10, ("__kmp_runtime_initialize: dynamic library pinned\n"));
652  }
653 #endif
654 
655  InitializeCriticalSection(&__kmp_win32_section);
656 #if USE_ITT_BUILD
657  __kmp_itt_system_object_created(&__kmp_win32_section, "Critical Section");
658 #endif /* USE_ITT_BUILD */
659  __kmp_initialize_system_tick();
660 
661 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
662  if (!__kmp_cpuinfo.initialized) {
663  __kmp_query_cpuid(&__kmp_cpuinfo);
664  }
665 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
666 
667 /* Set up minimum number of threads to switch to TLS gtid */
668 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
669  // Windows* OS, static library.
670  /* New thread may use stack space previously used by another thread,
671  currently terminated. On Windows* OS, in case of static linking, we do not
672  know the moment of thread termination, and our structures (__kmp_threads
673  and __kmp_root arrays) are still keep info about dead threads. This leads
674  to problem in __kmp_get_global_thread_id() function: it wrongly finds gtid
675  (by searching through stack addresses of all known threads) for
676  unregistered foreign tread.
677 
678  Setting __kmp_tls_gtid_min to 0 workarounds this problem:
679  __kmp_get_global_thread_id() does not search through stacks, but get gtid
680  from TLS immediately.
681  --ln
682  */
683  __kmp_tls_gtid_min = 0;
684 #else
685  __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
686 #endif
687 
688  /* for the static library */
689  if (!__kmp_gtid_threadprivate_key) {
690  __kmp_gtid_threadprivate_key = TlsAlloc();
691  if (__kmp_gtid_threadprivate_key == TLS_OUT_OF_INDEXES) {
692  KMP_FATAL(TLSOutOfIndexes);
693  }
694  }
695 
696  // Load ntdll.dll.
697  /* Simple GetModuleHandle( "ntdll.dl" ) is not suitable due to security issue
698  (see http://www.microsoft.com/technet/security/advisory/2269637.mspx). We
699  have to specify full path to the library. */
700  __kmp_str_buf_init(&path);
701  path_size = GetSystemDirectory(path.str, path.size);
702  KMP_DEBUG_ASSERT(path_size > 0);
703  if (path_size >= path.size) {
704  // Buffer is too short. Expand the buffer and try again.
705  __kmp_str_buf_reserve(&path, path_size);
706  path_size = GetSystemDirectory(path.str, path.size);
707  KMP_DEBUG_ASSERT(path_size > 0);
708  }
709  if (path_size > 0 && path_size < path.size) {
710  // Now we have system directory name in the buffer.
711  // Append backslash and name of dll to form full path,
712  path.used = path_size;
713  __kmp_str_buf_print(&path, "\\%s", "ntdll.dll");
714 
715  // Now load ntdll using full path.
716  ntdll = GetModuleHandle(path.str);
717  }
718 
719  KMP_DEBUG_ASSERT(ntdll != NULL);
720  if (ntdll != NULL) {
721  NtQuerySystemInformation = (NtQuerySystemInformation_t)GetProcAddress(
722  ntdll, "NtQuerySystemInformation");
723  }
724  KMP_DEBUG_ASSERT(NtQuerySystemInformation != NULL);
725 
726 #if KMP_GROUP_AFFINITY
727  // Load kernel32.dll.
728  // Same caveat - must use full system path name.
729  if (path_size > 0 && path_size < path.size) {
730  // Truncate the buffer back to just the system path length,
731  // discarding "\\ntdll.dll", and replacing it with "kernel32.dll".
732  path.used = path_size;
733  __kmp_str_buf_print(&path, "\\%s", "kernel32.dll");
734 
735  // Load kernel32.dll using full path.
736  kernel32 = GetModuleHandle(path.str);
737  KA_TRACE(10, ("__kmp_runtime_initialize: kernel32.dll = %s\n", path.str));
738 
739  // Load the function pointers to kernel32.dll routines
740  // that may or may not exist on this system.
741  if (kernel32 != NULL) {
742  __kmp_GetActiveProcessorCount =
743  (kmp_GetActiveProcessorCount_t)GetProcAddress(
744  kernel32, "GetActiveProcessorCount");
745  __kmp_GetActiveProcessorGroupCount =
746  (kmp_GetActiveProcessorGroupCount_t)GetProcAddress(
747  kernel32, "GetActiveProcessorGroupCount");
748  __kmp_GetThreadGroupAffinity =
749  (kmp_GetThreadGroupAffinity_t)GetProcAddress(
750  kernel32, "GetThreadGroupAffinity");
751  __kmp_SetThreadGroupAffinity =
752  (kmp_SetThreadGroupAffinity_t)GetProcAddress(
753  kernel32, "SetThreadGroupAffinity");
754 
755  KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_GetActiveProcessorCount"
756  " = %p\n",
757  __kmp_GetActiveProcessorCount));
758  KA_TRACE(10, ("__kmp_runtime_initialize: "
759  "__kmp_GetActiveProcessorGroupCount = %p\n",
760  __kmp_GetActiveProcessorGroupCount));
761  KA_TRACE(10, ("__kmp_runtime_initialize:__kmp_GetThreadGroupAffinity"
762  " = %p\n",
763  __kmp_GetThreadGroupAffinity));
764  KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_SetThreadGroupAffinity"
765  " = %p\n",
766  __kmp_SetThreadGroupAffinity));
767  KA_TRACE(10, ("__kmp_runtime_initialize: sizeof(kmp_affin_mask_t) = %d\n",
768  sizeof(kmp_affin_mask_t)));
769 
770  // See if group affinity is supported on this system.
771  // If so, calculate the #groups and #procs.
772  //
773  // Group affinity was introduced with Windows* 7 OS and
774  // Windows* Server 2008 R2 OS.
775  if ((__kmp_GetActiveProcessorCount != NULL) &&
776  (__kmp_GetActiveProcessorGroupCount != NULL) &&
777  (__kmp_GetThreadGroupAffinity != NULL) &&
778  (__kmp_SetThreadGroupAffinity != NULL) &&
779  ((__kmp_num_proc_groups = __kmp_GetActiveProcessorGroupCount()) >
780  1)) {
781  // Calculate the total number of active OS procs.
782  int i;
783 
784  KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups"
785  " detected\n",
786  __kmp_num_proc_groups));
787 
788  __kmp_xproc = 0;
789 
790  for (i = 0; i < __kmp_num_proc_groups; i++) {
791  DWORD size = __kmp_GetActiveProcessorCount(i);
792  __kmp_xproc += size;
793  KA_TRACE(10, ("__kmp_runtime_initialize: proc group %d size = %d\n",
794  i, size));
795  }
796  } else {
797  KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups"
798  " detected\n",
799  __kmp_num_proc_groups));
800  }
801  }
802  }
803  if (__kmp_num_proc_groups <= 1) {
804  GetSystemInfo(&info);
805  __kmp_xproc = info.dwNumberOfProcessors;
806  }
807 #else
808  GetSystemInfo(&info);
809  __kmp_xproc = info.dwNumberOfProcessors;
810 #endif /* KMP_GROUP_AFFINITY */
811 
812  // If the OS said there were 0 procs, take a guess and use a value of 2.
813  // This is done for Linux* OS, also. Do we need error / warning?
814  if (__kmp_xproc <= 0) {
815  __kmp_xproc = 2;
816  }
817 
818  KA_TRACE(5,
819  ("__kmp_runtime_initialize: total processors = %d\n", __kmp_xproc));
820 
821  __kmp_str_buf_free(&path);
822 
823 #if USE_ITT_BUILD
824  __kmp_itt_initialize();
825 #endif /* USE_ITT_BUILD */
826 
827  __kmp_init_runtime = TRUE;
828 } // __kmp_runtime_initialize
829 
830 void __kmp_runtime_destroy(void) {
831  if (!__kmp_init_runtime) {
832  return;
833  }
834 
835 #if USE_ITT_BUILD
836  __kmp_itt_destroy();
837 #endif /* USE_ITT_BUILD */
838 
839  /* we can't DeleteCriticalsection( & __kmp_win32_section ); */
840  /* due to the KX_TRACE() commands */
841  KA_TRACE(40, ("__kmp_runtime_destroy\n"));
842 
843  if (__kmp_gtid_threadprivate_key) {
844  TlsFree(__kmp_gtid_threadprivate_key);
845  __kmp_gtid_threadprivate_key = 0;
846  }
847 
848  __kmp_affinity_uninitialize();
849  DeleteCriticalSection(&__kmp_win32_section);
850 
851  ntdll = NULL;
852  NtQuerySystemInformation = NULL;
853 
854 #if KMP_ARCH_X86_64
855  kernel32 = NULL;
856  __kmp_GetActiveProcessorCount = NULL;
857  __kmp_GetActiveProcessorGroupCount = NULL;
858  __kmp_GetThreadGroupAffinity = NULL;
859  __kmp_SetThreadGroupAffinity = NULL;
860 #endif // KMP_ARCH_X86_64
861 
862  __kmp_init_runtime = FALSE;
863 }
864 
865 void __kmp_terminate_thread(int gtid) {
866  kmp_info_t *th = __kmp_threads[gtid];
867 
868  if (!th)
869  return;
870 
871  KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
872 
873  if (TerminateThread(th->th.th_info.ds.ds_thread, (DWORD)-1) == FALSE) {
874  /* It's OK, the thread may have exited already */
875  }
876  __kmp_free_handle(th->th.th_info.ds.ds_thread);
877 }
878 
879 void __kmp_clear_system_time(void) {
880  BOOL status;
881  LARGE_INTEGER time;
882  status = QueryPerformanceCounter(&time);
883  __kmp_win32_time = (kmp_int64)time.QuadPart;
884 }
885 
886 void __kmp_initialize_system_tick(void) {
887  {
888  BOOL status;
889  LARGE_INTEGER freq;
890 
891  status = QueryPerformanceFrequency(&freq);
892  if (!status) {
893  DWORD error = GetLastError();
894  __kmp_fatal(KMP_MSG(FunctionError, "QueryPerformanceFrequency()"),
895  KMP_ERR(error), __kmp_msg_null);
896 
897  } else {
898  __kmp_win32_tick = ((double)1.0) / (double)freq.QuadPart;
899  }
900  }
901 }
902 
903 /* Calculate the elapsed wall clock time for the user */
904 
905 void __kmp_elapsed(double *t) {
906  BOOL status;
907  LARGE_INTEGER now;
908  status = QueryPerformanceCounter(&now);
909  *t = ((double)now.QuadPart) * __kmp_win32_tick;
910 }
911 
912 /* Calculate the elapsed wall clock tick for the user */
913 
914 void __kmp_elapsed_tick(double *t) { *t = __kmp_win32_tick; }
915 
916 void __kmp_read_system_time(double *delta) {
917  if (delta != NULL) {
918  BOOL status;
919  LARGE_INTEGER now;
920 
921  status = QueryPerformanceCounter(&now);
922 
923  *delta = ((double)(((kmp_int64)now.QuadPart) - __kmp_win32_time)) *
924  __kmp_win32_tick;
925  }
926 }
927 
928 /* Return the current time stamp in nsec */
929 kmp_uint64 __kmp_now_nsec() {
930  LARGE_INTEGER now;
931  QueryPerformanceCounter(&now);
932  return 1e9 * __kmp_win32_tick * now.QuadPart;
933 }
934 
935 extern "C"
936 void *__stdcall __kmp_launch_worker(void *arg) {
937  volatile void *stack_data;
938  void *exit_val;
939  void *padding = 0;
940  kmp_info_t *this_thr = (kmp_info_t *)arg;
941  int gtid;
942 
943  gtid = this_thr->th.th_info.ds.ds_gtid;
944  __kmp_gtid_set_specific(gtid);
945 #ifdef KMP_TDATA_GTID
946 #error "This define causes problems with LoadLibrary() + declspec(thread) " \
947  "on Windows* OS. See CQ50564, tests kmp_load_library*.c and this MSDN " \
948  "reference: http://support.microsoft.com/kb/118816"
949 //__kmp_gtid = gtid;
950 #endif
951 
952 #if USE_ITT_BUILD
953  __kmp_itt_thread_name(gtid);
954 #endif /* USE_ITT_BUILD */
955 
956  __kmp_affinity_set_init_mask(gtid, FALSE);
957 
958 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
959  // Set FP control regs to be a copy of the parallel initialization thread's.
960  __kmp_clear_x87_fpu_status_word();
961  __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
962  __kmp_load_mxcsr(&__kmp_init_mxcsr);
963 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
964 
965  if (__kmp_stkoffset > 0 && gtid > 0) {
966  padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
967  }
968 
969  KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive);
970  this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
971  TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE);
972 
973  if (TCR_4(__kmp_gtid_mode) <
974  2) { // check stack only if it is used to get gtid
975  TCW_PTR(this_thr->th.th_info.ds.ds_stackbase, &stack_data);
976  KMP_ASSERT(this_thr->th.th_info.ds.ds_stackgrow == FALSE);
977  __kmp_check_stack_overlap(this_thr);
978  }
979  KMP_MB();
980  exit_val = __kmp_launch_thread(this_thr);
981  KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive);
982  TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE);
983  KMP_MB();
984  return exit_val;
985 }
986 
987 #if KMP_USE_MONITOR
988 /* The monitor thread controls all of the threads in the complex */
989 
990 void *__stdcall __kmp_launch_monitor(void *arg) {
991  DWORD wait_status;
992  kmp_thread_t monitor;
993  int status;
994  int interval;
995  kmp_info_t *this_thr = (kmp_info_t *)arg;
996 
997  KMP_DEBUG_ASSERT(__kmp_init_monitor);
998  TCW_4(__kmp_init_monitor, 2); // AC: Signal library that monitor has started
999  // TODO: hide "2" in enum (like {true,false,started})
1000  this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
1001  TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE);
1002 
1003  KMP_MB(); /* Flush all pending memory write invalidates. */
1004  KA_TRACE(10, ("__kmp_launch_monitor: launched\n"));
1005 
1006  monitor = GetCurrentThread();
1007 
1008  /* set thread priority */
1009  status = SetThreadPriority(monitor, THREAD_PRIORITY_HIGHEST);
1010  if (!status) {
1011  DWORD error = GetLastError();
1012  __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null);
1013  }
1014 
1015  /* register us as monitor */
1016  __kmp_gtid_set_specific(KMP_GTID_MONITOR);
1017 #ifdef KMP_TDATA_GTID
1018 #error "This define causes problems with LoadLibrary() + declspec(thread) " \
1019  "on Windows* OS. See CQ50564, tests kmp_load_library*.c and this MSDN " \
1020  "reference: http://support.microsoft.com/kb/118816"
1021 //__kmp_gtid = KMP_GTID_MONITOR;
1022 #endif
1023 
1024 #if USE_ITT_BUILD
1025  __kmp_itt_thread_ignore(); // Instruct Intel(R) Threading Tools to ignore
1026 // monitor thread.
1027 #endif /* USE_ITT_BUILD */
1028 
1029  KMP_MB(); /* Flush all pending memory write invalidates. */
1030 
1031  interval = (1000 / __kmp_monitor_wakeups); /* in milliseconds */
1032 
1033  while (!TCR_4(__kmp_global.g.g_done)) {
1034  /* This thread monitors the state of the system */
1035 
1036  KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
1037 
1038  wait_status = WaitForSingleObject(__kmp_monitor_ev, interval);
1039 
1040  if (wait_status == WAIT_TIMEOUT) {
1041  TCW_4(__kmp_global.g.g_time.dt.t_value,
1042  TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
1043  }
1044 
1045  KMP_MB(); /* Flush all pending memory write invalidates. */
1046  }
1047 
1048  KA_TRACE(10, ("__kmp_launch_monitor: finished\n"));
1049 
1050  status = SetThreadPriority(monitor, THREAD_PRIORITY_NORMAL);
1051  if (!status) {
1052  DWORD error = GetLastError();
1053  __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null);
1054  }
1055 
1056  if (__kmp_global.g.g_abort != 0) {
1057  /* now we need to terminate the worker threads */
1058  /* the value of t_abort is the signal we caught */
1059  int gtid;
1060 
1061  KA_TRACE(10, ("__kmp_launch_monitor: terminate sig=%d\n",
1062  (__kmp_global.g.g_abort)));
1063 
1064  /* terminate the OpenMP worker threads */
1065  /* TODO this is not valid for sibling threads!!
1066  * the uber master might not be 0 anymore.. */
1067  for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
1068  __kmp_terminate_thread(gtid);
1069 
1070  __kmp_cleanup();
1071 
1072  Sleep(0);
1073 
1074  KA_TRACE(10,
1075  ("__kmp_launch_monitor: raise sig=%d\n", __kmp_global.g.g_abort));
1076 
1077  if (__kmp_global.g.g_abort > 0) {
1078  raise(__kmp_global.g.g_abort);
1079  }
1080  }
1081 
1082  TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE);
1083 
1084  KMP_MB();
1085  return arg;
1086 }
1087 #endif
1088 
1089 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
1090  kmp_thread_t handle;
1091  DWORD idThread;
1092 
1093  KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
1094 
1095  th->th.th_info.ds.ds_gtid = gtid;
1096 
1097  if (KMP_UBER_GTID(gtid)) {
1098  int stack_data;
1099 
1100  /* TODO: GetCurrentThread() returns a pseudo-handle that is unsuitable for
1101  other threads to use. Is it appropriate to just use GetCurrentThread?
1102  When should we close this handle? When unregistering the root? */
1103  {
1104  BOOL rc;
1105  rc = DuplicateHandle(GetCurrentProcess(), GetCurrentThread(),
1106  GetCurrentProcess(), &th->th.th_info.ds.ds_thread, 0,
1107  FALSE, DUPLICATE_SAME_ACCESS);
1108  KMP_ASSERT(rc);
1109  KA_TRACE(10, (" __kmp_create_worker: ROOT Handle duplicated, th = %p, "
1110  "handle = %" KMP_UINTPTR_SPEC "\n",
1111  (LPVOID)th, th->th.th_info.ds.ds_thread));
1112  th->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
1113  }
1114  if (TCR_4(__kmp_gtid_mode) < 2) { // check stack only if used to get gtid
1115  /* we will dynamically update the stack range if gtid_mode == 1 */
1116  TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
1117  TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
1118  TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
1119  __kmp_check_stack_overlap(th);
1120  }
1121  } else {
1122  KMP_MB(); /* Flush all pending memory write invalidates. */
1123 
1124  /* Set stack size for this thread now. */
1125  KA_TRACE(10,
1126  ("__kmp_create_worker: stack_size = %" KMP_SIZE_T_SPEC " bytes\n",
1127  stack_size));
1128 
1129  stack_size += gtid * __kmp_stkoffset;
1130 
1131  TCW_PTR(th->th.th_info.ds.ds_stacksize, stack_size);
1132  TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
1133 
1134  KA_TRACE(10,
1135  ("__kmp_create_worker: (before) stack_size = %" KMP_SIZE_T_SPEC
1136  " bytes, &__kmp_launch_worker = %p, th = %p, &idThread = %p\n",
1137  (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker,
1138  (LPVOID)th, &idThread));
1139 
1140  handle = CreateThread(
1141  NULL, (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)__kmp_launch_worker,
1142  (LPVOID)th, STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread);
1143 
1144  KA_TRACE(10,
1145  ("__kmp_create_worker: (after) stack_size = %" KMP_SIZE_T_SPEC
1146  " bytes, &__kmp_launch_worker = %p, th = %p, "
1147  "idThread = %u, handle = %" KMP_UINTPTR_SPEC "\n",
1148  (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker,
1149  (LPVOID)th, idThread, handle));
1150 
1151  if (handle == 0) {
1152  DWORD error = GetLastError();
1153  __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null);
1154  } else {
1155  th->th.th_info.ds.ds_thread = handle;
1156  }
1157 
1158  KMP_MB(); /* Flush all pending memory write invalidates. */
1159  }
1160 
1161  KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
1162 }
1163 
1164 int __kmp_still_running(kmp_info_t *th) {
1165  return (WAIT_TIMEOUT == WaitForSingleObject(th->th.th_info.ds.ds_thread, 0));
1166 }
1167 
1168 #if KMP_USE_MONITOR
1169 void __kmp_create_monitor(kmp_info_t *th) {
1170  kmp_thread_t handle;
1171  DWORD idThread;
1172  int ideal, new_ideal;
1173 
1174  if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
1175  // We don't need monitor thread in case of MAX_BLOCKTIME
1176  KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
1177  "MAX blocktime\n"));
1178  th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
1179  th->th.th_info.ds.ds_gtid = 0;
1180  TCW_4(__kmp_init_monitor, 2); // Signal to stop waiting for monitor creation
1181  return;
1182  }
1183  KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
1184 
1185  KMP_MB(); /* Flush all pending memory write invalidates. */
1186 
1187  __kmp_monitor_ev = CreateEvent(NULL, TRUE, FALSE, NULL);
1188  if (__kmp_monitor_ev == NULL) {
1189  DWORD error = GetLastError();
1190  __kmp_fatal(KMP_MSG(CantCreateEvent), KMP_ERR(error), __kmp_msg_null);
1191  }
1192 #if USE_ITT_BUILD
1193  __kmp_itt_system_object_created(__kmp_monitor_ev, "Event");
1194 #endif /* USE_ITT_BUILD */
1195 
1196  th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
1197  th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
1198 
1199  // FIXME - on Windows* OS, if __kmp_monitor_stksize = 0, figure out how
1200  // to automatically expand stacksize based on CreateThread error code.
1201  if (__kmp_monitor_stksize == 0) {
1202  __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
1203  }
1204  if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
1205  __kmp_monitor_stksize = __kmp_sys_min_stksize;
1206  }
1207 
1208  KA_TRACE(10, ("__kmp_create_monitor: requested stacksize = %d bytes\n",
1209  (int)__kmp_monitor_stksize));
1210 
1211  TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
1212 
1213  handle =
1214  CreateThread(NULL, (SIZE_T)__kmp_monitor_stksize,
1215  (LPTHREAD_START_ROUTINE)__kmp_launch_monitor, (LPVOID)th,
1216  STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread);
1217  if (handle == 0) {
1218  DWORD error = GetLastError();
1219  __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null);
1220  } else
1221  th->th.th_info.ds.ds_thread = handle;
1222 
1223  KMP_MB(); /* Flush all pending memory write invalidates. */
1224 
1225  KA_TRACE(10, ("__kmp_create_monitor: monitor created %p\n",
1226  (void *)th->th.th_info.ds.ds_thread));
1227 }
1228 #endif
1229 
1230 /* Check to see if thread is still alive.
1231  NOTE: The ExitProcess(code) system call causes all threads to Terminate
1232  with a exit_val = code. Because of this we can not rely on exit_val having
1233  any particular value. So this routine may return STILL_ALIVE in exit_val
1234  even after the thread is dead. */
1235 
1236 int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val) {
1237  DWORD rc;
1238  rc = GetExitCodeThread(th->th.th_info.ds.ds_thread, exit_val);
1239  if (rc == 0) {
1240  DWORD error = GetLastError();
1241  __kmp_fatal(KMP_MSG(FunctionError, "GetExitCodeThread()"), KMP_ERR(error),
1242  __kmp_msg_null);
1243  }
1244  return (*exit_val == STILL_ACTIVE);
1245 }
1246 
1247 void __kmp_exit_thread(int exit_status) {
1248  ExitThread(exit_status);
1249 } // __kmp_exit_thread
1250 
1251 // This is a common part for both __kmp_reap_worker() and __kmp_reap_monitor().
1252 static void __kmp_reap_common(kmp_info_t *th) {
1253  DWORD exit_val;
1254 
1255  KMP_MB(); /* Flush all pending memory write invalidates. */
1256 
1257  KA_TRACE(
1258  10, ("__kmp_reap_common: try to reap (%d)\n", th->th.th_info.ds.ds_gtid));
1259 
1260  /* 2006-10-19:
1261  There are two opposite situations:
1262  1. Windows* OS keep thread alive after it resets ds_alive flag and
1263  exits from thread function. (For example, see C70770/Q394281 "unloading of
1264  dll based on OMP is very slow".)
1265  2. Windows* OS may kill thread before it resets ds_alive flag.
1266 
1267  Right solution seems to be waiting for *either* thread termination *or*
1268  ds_alive resetting. */
1269  {
1270  // TODO: This code is very similar to KMP_WAIT. Need to generalize
1271  // KMP_WAIT to cover this usage also.
1272  void *obj = NULL;
1273  kmp_uint32 spins;
1274 #if USE_ITT_BUILD
1275  KMP_FSYNC_SPIN_INIT(obj, (void *)&th->th.th_info.ds.ds_alive);
1276 #endif /* USE_ITT_BUILD */
1277  KMP_INIT_YIELD(spins);
1278  do {
1279 #if USE_ITT_BUILD
1280  KMP_FSYNC_SPIN_PREPARE(obj);
1281 #endif /* USE_ITT_BUILD */
1282  __kmp_is_thread_alive(th, &exit_val);
1283  KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
1284  } while (exit_val == STILL_ACTIVE && TCR_4(th->th.th_info.ds.ds_alive));
1285 #if USE_ITT_BUILD
1286  if (exit_val == STILL_ACTIVE) {
1287  KMP_FSYNC_CANCEL(obj);
1288  } else {
1289  KMP_FSYNC_SPIN_ACQUIRED(obj);
1290  }
1291 #endif /* USE_ITT_BUILD */
1292  }
1293 
1294  __kmp_free_handle(th->th.th_info.ds.ds_thread);
1295 
1296  /* NOTE: The ExitProcess(code) system call causes all threads to Terminate
1297  with a exit_val = code. Because of this we can not rely on exit_val having
1298  any particular value. */
1299  if (exit_val == STILL_ACTIVE) {
1300  KA_TRACE(1, ("__kmp_reap_common: thread still active.\n"));
1301  } else if ((void *)exit_val != (void *)th) {
1302  KA_TRACE(1, ("__kmp_reap_common: ExitProcess / TerminateThread used?\n"));
1303  }
1304 
1305  KA_TRACE(10,
1306  ("__kmp_reap_common: done reaping (%d), handle = %" KMP_UINTPTR_SPEC
1307  "\n",
1308  th->th.th_info.ds.ds_gtid, th->th.th_info.ds.ds_thread));
1309 
1310  th->th.th_info.ds.ds_thread = 0;
1311  th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1312  th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1313  th->th.th_info.ds.ds_thread_id = 0;
1314 
1315  KMP_MB(); /* Flush all pending memory write invalidates. */
1316 }
1317 
1318 #if KMP_USE_MONITOR
1319 void __kmp_reap_monitor(kmp_info_t *th) {
1320  int status;
1321 
1322  KA_TRACE(10, ("__kmp_reap_monitor: try to reap %p\n",
1323  (void *)th->th.th_info.ds.ds_thread));
1324 
1325  // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1326  // If both tid and gtid are 0, it means the monitor did not ever start.
1327  // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1328  KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1329  if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1330  KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1331  return;
1332  }
1333 
1334  KMP_MB(); /* Flush all pending memory write invalidates. */
1335 
1336  status = SetEvent(__kmp_monitor_ev);
1337  if (status == FALSE) {
1338  DWORD error = GetLastError();
1339  __kmp_fatal(KMP_MSG(CantSetEvent), KMP_ERR(error), __kmp_msg_null);
1340  }
1341  KA_TRACE(10, ("__kmp_reap_monitor: reaping thread (%d)\n",
1342  th->th.th_info.ds.ds_gtid));
1343  __kmp_reap_common(th);
1344 
1345  __kmp_free_handle(__kmp_monitor_ev);
1346 
1347  KMP_MB(); /* Flush all pending memory write invalidates. */
1348 }
1349 #endif
1350 
1351 void __kmp_reap_worker(kmp_info_t *th) {
1352  KA_TRACE(10, ("__kmp_reap_worker: reaping thread (%d)\n",
1353  th->th.th_info.ds.ds_gtid));
1354  __kmp_reap_common(th);
1355 }
1356 
1357 #if KMP_HANDLE_SIGNALS
1358 
1359 static void __kmp_team_handler(int signo) {
1360  if (__kmp_global.g.g_abort == 0) {
1361  // Stage 1 signal handler, let's shut down all of the threads.
1362  if (__kmp_debug_buf) {
1363  __kmp_dump_debug_buffer();
1364  }
1365  KMP_MB(); // Flush all pending memory write invalidates.
1366  TCW_4(__kmp_global.g.g_abort, signo);
1367  KMP_MB(); // Flush all pending memory write invalidates.
1368  TCW_4(__kmp_global.g.g_done, TRUE);
1369  KMP_MB(); // Flush all pending memory write invalidates.
1370  }
1371 } // __kmp_team_handler
1372 
1373 static sig_func_t __kmp_signal(int signum, sig_func_t handler) {
1374  sig_func_t old = signal(signum, handler);
1375  if (old == SIG_ERR) {
1376  int error = errno;
1377  __kmp_fatal(KMP_MSG(FunctionError, "signal"), KMP_ERR(error),
1378  __kmp_msg_null);
1379  }
1380  return old;
1381 }
1382 
1383 static void __kmp_install_one_handler(int sig, sig_func_t handler,
1384  int parallel_init) {
1385  sig_func_t old;
1386  KMP_MB(); /* Flush all pending memory write invalidates. */
1387  KB_TRACE(60, ("__kmp_install_one_handler: called: sig=%d\n", sig));
1388  if (parallel_init) {
1389  old = __kmp_signal(sig, handler);
1390  // SIG_DFL on Windows* OS in NULL or 0.
1391  if (old == __kmp_sighldrs[sig]) {
1392  __kmp_siginstalled[sig] = 1;
1393  } else { // Restore/keep user's handler if one previously installed.
1394  old = __kmp_signal(sig, old);
1395  }
1396  } else {
1397  // Save initial/system signal handlers to see if user handlers installed.
1398  // 2009-09-23: It is a dead code. On Windows* OS __kmp_install_signals
1399  // called once with parallel_init == TRUE.
1400  old = __kmp_signal(sig, SIG_DFL);
1401  __kmp_sighldrs[sig] = old;
1402  __kmp_signal(sig, old);
1403  }
1404  KMP_MB(); /* Flush all pending memory write invalidates. */
1405 } // __kmp_install_one_handler
1406 
1407 static void __kmp_remove_one_handler(int sig) {
1408  if (__kmp_siginstalled[sig]) {
1409  sig_func_t old;
1410  KMP_MB(); // Flush all pending memory write invalidates.
1411  KB_TRACE(60, ("__kmp_remove_one_handler: called: sig=%d\n", sig));
1412  old = __kmp_signal(sig, __kmp_sighldrs[sig]);
1413  if (old != __kmp_team_handler) {
1414  KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1415  "restoring: sig=%d\n",
1416  sig));
1417  old = __kmp_signal(sig, old);
1418  }
1419  __kmp_sighldrs[sig] = NULL;
1420  __kmp_siginstalled[sig] = 0;
1421  KMP_MB(); // Flush all pending memory write invalidates.
1422  }
1423 } // __kmp_remove_one_handler
1424 
1425 void __kmp_install_signals(int parallel_init) {
1426  KB_TRACE(10, ("__kmp_install_signals: called\n"));
1427  if (!__kmp_handle_signals) {
1428  KB_TRACE(10, ("__kmp_install_signals: KMP_HANDLE_SIGNALS is false - "
1429  "handlers not installed\n"));
1430  return;
1431  }
1432  __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1433  __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1434  __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1435  __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1436  __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1437  __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1438 } // __kmp_install_signals
1439 
1440 void __kmp_remove_signals(void) {
1441  int sig;
1442  KB_TRACE(10, ("__kmp_remove_signals: called\n"));
1443  for (sig = 1; sig < NSIG; ++sig) {
1444  __kmp_remove_one_handler(sig);
1445  }
1446 } // __kmp_remove_signals
1447 
1448 #endif // KMP_HANDLE_SIGNALS
1449 
1450 /* Put the thread to sleep for a time period */
1451 void __kmp_thread_sleep(int millis) {
1452  DWORD status;
1453 
1454  status = SleepEx((DWORD)millis, FALSE);
1455  if (status) {
1456  DWORD error = GetLastError();
1457  __kmp_fatal(KMP_MSG(FunctionError, "SleepEx()"), KMP_ERR(error),
1458  __kmp_msg_null);
1459  }
1460 }
1461 
1462 // Determine whether the given address is mapped into the current address space.
1463 int __kmp_is_address_mapped(void *addr) {
1464  DWORD status;
1465  MEMORY_BASIC_INFORMATION lpBuffer;
1466  SIZE_T dwLength;
1467 
1468  dwLength = sizeof(MEMORY_BASIC_INFORMATION);
1469 
1470  status = VirtualQuery(addr, &lpBuffer, dwLength);
1471 
1472  return !(((lpBuffer.State == MEM_RESERVE) || (lpBuffer.State == MEM_FREE)) ||
1473  ((lpBuffer.Protect == PAGE_NOACCESS) ||
1474  (lpBuffer.Protect == PAGE_EXECUTE)));
1475 }
1476 
1477 kmp_uint64 __kmp_hardware_timestamp(void) {
1478  kmp_uint64 r = 0;
1479 
1480  QueryPerformanceCounter((LARGE_INTEGER *)&r);
1481  return r;
1482 }
1483 
1484 /* Free handle and check the error code */
1485 void __kmp_free_handle(kmp_thread_t tHandle) {
1486  /* called with parameter type HANDLE also, thus suppose kmp_thread_t defined
1487  * as HANDLE */
1488  BOOL rc;
1489  rc = CloseHandle(tHandle);
1490  if (!rc) {
1491  DWORD error = GetLastError();
1492  __kmp_fatal(KMP_MSG(CantCloseHandle), KMP_ERR(error), __kmp_msg_null);
1493  }
1494 }
1495 
1496 int __kmp_get_load_balance(int max) {
1497  static ULONG glb_buff_size = 100 * 1024;
1498 
1499  // Saved count of the running threads for the thread balance algortihm
1500  static int glb_running_threads = 0;
1501  static double glb_call_time = 0; /* Thread balance algorithm call time */
1502 
1503  int running_threads = 0; // Number of running threads in the system.
1504  NTSTATUS status = 0;
1505  ULONG buff_size = 0;
1506  ULONG info_size = 0;
1507  void *buffer = NULL;
1508  PSYSTEM_PROCESS_INFORMATION spi = NULL;
1509  int first_time = 1;
1510 
1511  double call_time = 0.0; // start, finish;
1512 
1513  __kmp_elapsed(&call_time);
1514 
1515  if (glb_call_time &&
1516  (call_time - glb_call_time < __kmp_load_balance_interval)) {
1517  running_threads = glb_running_threads;
1518  goto finish;
1519  }
1520  glb_call_time = call_time;
1521 
1522  // Do not spend time on running algorithm if we have a permanent error.
1523  if (NtQuerySystemInformation == NULL) {
1524  running_threads = -1;
1525  goto finish;
1526  }
1527 
1528  if (max <= 0) {
1529  max = INT_MAX;
1530  }
1531 
1532  do {
1533 
1534  if (first_time) {
1535  buff_size = glb_buff_size;
1536  } else {
1537  buff_size = 2 * buff_size;
1538  }
1539 
1540  buffer = KMP_INTERNAL_REALLOC(buffer, buff_size);
1541  if (buffer == NULL) {
1542  running_threads = -1;
1543  goto finish;
1544  }
1545  status = NtQuerySystemInformation(SystemProcessInformation, buffer,
1546  buff_size, &info_size);
1547  first_time = 0;
1548 
1549  } while (status == STATUS_INFO_LENGTH_MISMATCH);
1550  glb_buff_size = buff_size;
1551 
1552 #define CHECK(cond) \
1553  { \
1554  KMP_DEBUG_ASSERT(cond); \
1555  if (!(cond)) { \
1556  running_threads = -1; \
1557  goto finish; \
1558  } \
1559  }
1560 
1561  CHECK(buff_size >= info_size);
1562  spi = PSYSTEM_PROCESS_INFORMATION(buffer);
1563  for (;;) {
1564  ptrdiff_t offset = uintptr_t(spi) - uintptr_t(buffer);
1565  CHECK(0 <= offset &&
1566  offset + sizeof(SYSTEM_PROCESS_INFORMATION) < info_size);
1567  HANDLE pid = spi->ProcessId;
1568  ULONG num = spi->NumberOfThreads;
1569  CHECK(num >= 1);
1570  size_t spi_size =
1571  sizeof(SYSTEM_PROCESS_INFORMATION) + sizeof(SYSTEM_THREAD) * (num - 1);
1572  CHECK(offset + spi_size <
1573  info_size); // Make sure process info record fits the buffer.
1574  if (spi->NextEntryOffset != 0) {
1575  CHECK(spi_size <=
1576  spi->NextEntryOffset); // And do not overlap with the next record.
1577  }
1578  // pid == 0 corresponds to the System Idle Process. It always has running
1579  // threads on all cores. So, we don't consider the running threads of this
1580  // process.
1581  if (pid != 0) {
1582  for (int i = 0; i < num; ++i) {
1583  THREAD_STATE state = spi->Threads[i].State;
1584  // Count threads that have Ready or Running state.
1585  // !!! TODO: Why comment does not match the code???
1586  if (state == StateRunning) {
1587  ++running_threads;
1588  // Stop counting running threads if the number is already greater than
1589  // the number of available cores
1590  if (running_threads >= max) {
1591  goto finish;
1592  }
1593  }
1594  }
1595  }
1596  if (spi->NextEntryOffset == 0) {
1597  break;
1598  }
1599  spi = PSYSTEM_PROCESS_INFORMATION(uintptr_t(spi) + spi->NextEntryOffset);
1600  }
1601 
1602 #undef CHECK
1603 
1604 finish: // Clean up and exit.
1605 
1606  if (buffer != NULL) {
1607  KMP_INTERNAL_FREE(buffer);
1608  }
1609 
1610  glb_running_threads = running_threads;
1611 
1612  return running_threads;
1613 } //__kmp_get_load_balance()