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  int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init);
314  int new_value = TRUE;
315  // Return if already initialized
316  if (old_value == new_value)
317  return;
318  // Wait, then return if being initialized
319  if (old_value == -1 ||
320  !__kmp_atomic_compare_store(&th->th.th_suspend_init, old_value, -1)) {
321  while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init) != new_value) {
322  KMP_CPU_PAUSE();
323  }
324  } else {
325  // Claim to be the initializer and do initializations
326  __kmp_win32_cond_init(&th->th.th_suspend_cv);
327  __kmp_win32_mutex_init(&th->th.th_suspend_mx);
328  KMP_ATOMIC_ST_REL(&th->th.th_suspend_init, new_value);
329  }
330 }
331 
332 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) {
333  if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init)) {
334  /* this means we have initialize the suspension pthread objects for this
335  thread in this instance of the process */
336  __kmp_win32_cond_destroy(&th->th.th_suspend_cv);
337  __kmp_win32_mutex_destroy(&th->th.th_suspend_mx);
338  KMP_ATOMIC_ST_REL(&th->th.th_suspend_init, FALSE);
339  }
340 }
341 
342 int __kmp_try_suspend_mx(kmp_info_t *th) {
343  return __kmp_win32_mutex_trylock(&th->th.th_suspend_mx);
344 }
345 
346 void __kmp_lock_suspend_mx(kmp_info_t *th) {
347  __kmp_win32_mutex_lock(&th->th.th_suspend_mx);
348 }
349 
350 void __kmp_unlock_suspend_mx(kmp_info_t *th) {
351  __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
352 }
353 
354 /* This routine puts the calling thread to sleep after setting the
355  sleep bit for the indicated flag variable to true. */
356 template <class C>
357 static inline void __kmp_suspend_template(int th_gtid, C *flag) {
358  kmp_info_t *th = __kmp_threads[th_gtid];
359  int status;
360  typename C::flag_t old_spin;
361 
362  KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag's loc(%p)\n",
363  th_gtid, flag->get()));
364 
365  __kmp_suspend_initialize_thread(th);
366  __kmp_lock_suspend_mx(th);
367 
368  KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for flag's"
369  " loc(%p)\n",
370  th_gtid, flag->get()));
371 
372  /* TODO: shouldn't this use release semantics to ensure that
373  __kmp_suspend_initialize_thread gets called first? */
374  old_spin = flag->set_sleeping();
375  if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME &&
376  __kmp_pause_status != kmp_soft_paused) {
377  flag->unset_sleeping();
378  __kmp_unlock_suspend_mx(th);
379  return;
380  }
381 
382  KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for flag's"
383  " loc(%p)==%d\n",
384  th_gtid, flag->get(), *(flag->get())));
385 
386  if (flag->done_check_val(old_spin)) {
387  old_spin = flag->unset_sleeping();
388  KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit "
389  "for flag's loc(%p)\n",
390  th_gtid, flag->get()));
391  } else {
392 #ifdef DEBUG_SUSPEND
393  __kmp_suspend_count++;
394 #endif
395  /* Encapsulate in a loop as the documentation states that this may "with
396  low probability" return when the condition variable has not been signaled
397  or broadcast */
398  int deactivated = FALSE;
399  TCW_PTR(th->th.th_sleep_loc, (void *)flag);
400  while (flag->is_sleeping()) {
401  KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform "
402  "kmp_win32_cond_wait()\n",
403  th_gtid));
404  // Mark the thread as no longer active (only in the first iteration of the
405  // loop).
406  if (!deactivated) {
407  th->th.th_active = FALSE;
408  if (th->th.th_active_in_pool) {
409  th->th.th_active_in_pool = FALSE;
410  KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth);
411  KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0);
412  }
413  deactivated = TRUE;
414  __kmp_win32_cond_wait(&th->th.th_suspend_cv, &th->th.th_suspend_mx, th,
415  flag);
416  } else {
417  __kmp_win32_cond_wait(&th->th.th_suspend_cv, &th->th.th_suspend_mx, th,
418  flag);
419  }
420 
421 #ifdef KMP_DEBUG
422  if (flag->is_sleeping()) {
423  KF_TRACE(100,
424  ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid));
425  }
426 #endif /* KMP_DEBUG */
427 
428  } // while
429 
430  // Mark the thread as active again (if it was previous marked as inactive)
431  if (deactivated) {
432  th->th.th_active = TRUE;
433  if (TCR_4(th->th.th_in_pool)) {
434  KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth);
435  th->th.th_active_in_pool = TRUE;
436  }
437  }
438  }
439 
440  __kmp_unlock_suspend_mx(th);
441  KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
442 }
443 
444 template <bool C, bool S>
445 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) {
446  __kmp_suspend_template(th_gtid, flag);
447 }
448 template <bool C, bool S>
449 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) {
450  __kmp_suspend_template(th_gtid, flag);
451 }
452 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
453  __kmp_suspend_template(th_gtid, flag);
454 }
455 
456 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *);
457 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *);
458 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *);
459 
460 /* This routine signals the thread specified by target_gtid to wake up
461  after setting the sleep bit indicated by the flag argument to FALSE */
462 template <class C>
463 static inline void __kmp_resume_template(int target_gtid, C *flag) {
464  kmp_info_t *th = __kmp_threads[target_gtid];
465  int status;
466 
467 #ifdef KMP_DEBUG
468  int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
469 #endif
470 
471  KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
472  gtid, target_gtid));
473 
474  __kmp_suspend_initialize_thread(th);
475  __kmp_lock_suspend_mx(th);
476 
477  if (!flag) { // coming from __kmp_null_resume_wrapper
478  flag = (C *)th->th.th_sleep_loc;
479  }
480 
481  // First, check if the flag is null or its type has changed. If so, someone
482  // else woke it up.
483  if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
484  // simply shows what
485  // flag was cast to
486  KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
487  "awake: flag's loc(%p)\n",
488  gtid, target_gtid, NULL));
489  __kmp_unlock_suspend_mx(th);
490  return;
491  } else {
492  typename C::flag_t old_spin = flag->unset_sleeping();
493  if (!flag->is_sleeping_val(old_spin)) {
494  KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
495  "awake: flag's loc(%p): %u => %u\n",
496  gtid, target_gtid, flag->get(), old_spin, *(flag->get())));
497  __kmp_unlock_suspend_mx(th);
498  return;
499  }
500  }
501  TCW_PTR(th->th.th_sleep_loc, NULL);
502  KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset sleep "
503  "bit for flag's loc(%p)\n",
504  gtid, target_gtid, flag->get()));
505 
506  __kmp_win32_cond_signal(&th->th.th_suspend_cv);
507  __kmp_unlock_suspend_mx(th);
508 
509  KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
510  " for T#%d\n",
511  gtid, target_gtid));
512 }
513 
514 template <bool C, bool S>
515 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) {
516  __kmp_resume_template(target_gtid, flag);
517 }
518 template <bool C, bool S>
519 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) {
520  __kmp_resume_template(target_gtid, flag);
521 }
522 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
523  __kmp_resume_template(target_gtid, flag);
524 }
525 
526 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *);
527 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *);
528 
529 void __kmp_yield() { Sleep(0); }
530 
531 void __kmp_gtid_set_specific(int gtid) {
532  if (__kmp_init_gtid) {
533  KA_TRACE(50, ("__kmp_gtid_set_specific: T#%d key:%d\n", gtid,
534  __kmp_gtid_threadprivate_key));
535  if (!TlsSetValue(__kmp_gtid_threadprivate_key, (LPVOID)(gtid + 1)))
536  KMP_FATAL(TLSSetValueFailed);
537  } else {
538  KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
539  }
540 }
541 
542 int __kmp_gtid_get_specific() {
543  int gtid;
544  if (!__kmp_init_gtid) {
545  KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
546  "KMP_GTID_SHUTDOWN\n"));
547  return KMP_GTID_SHUTDOWN;
548  }
549  gtid = (int)(kmp_intptr_t)TlsGetValue(__kmp_gtid_threadprivate_key);
550  if (gtid == 0) {
551  gtid = KMP_GTID_DNE;
552  } else {
553  gtid--;
554  }
555  KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
556  __kmp_gtid_threadprivate_key, gtid));
557  return gtid;
558 }
559 
560 void __kmp_affinity_bind_thread(int proc) {
561  if (__kmp_num_proc_groups > 1) {
562  // Form the GROUP_AFFINITY struct directly, rather than filling
563  // out a bit vector and calling __kmp_set_system_affinity().
564  GROUP_AFFINITY ga;
565  KMP_DEBUG_ASSERT((proc >= 0) && (proc < (__kmp_num_proc_groups * CHAR_BIT *
566  sizeof(DWORD_PTR))));
567  ga.Group = proc / (CHAR_BIT * sizeof(DWORD_PTR));
568  ga.Mask = (unsigned long long)1 << (proc % (CHAR_BIT * sizeof(DWORD_PTR)));
569  ga.Reserved[0] = ga.Reserved[1] = ga.Reserved[2] = 0;
570 
571  KMP_DEBUG_ASSERT(__kmp_SetThreadGroupAffinity != NULL);
572  if (__kmp_SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL) == 0) {
573  DWORD error = GetLastError();
574  if (__kmp_affinity_verbose) { // AC: continue silently if not verbose
575  kmp_msg_t err_code = KMP_ERR(error);
576  __kmp_msg(kmp_ms_warning, KMP_MSG(CantSetThreadAffMask), err_code,
577  __kmp_msg_null);
578  if (__kmp_generate_warnings == kmp_warnings_off) {
579  __kmp_str_free(&err_code.str);
580  }
581  }
582  }
583  } else {
584  kmp_affin_mask_t *mask;
585  KMP_CPU_ALLOC_ON_STACK(mask);
586  KMP_CPU_ZERO(mask);
587  KMP_CPU_SET(proc, mask);
588  __kmp_set_system_affinity(mask, TRUE);
589  KMP_CPU_FREE_FROM_STACK(mask);
590  }
591 }
592 
593 void __kmp_affinity_determine_capable(const char *env_var) {
594 // All versions of Windows* OS (since Win '95) support SetThreadAffinityMask().
595 
596 #if KMP_GROUP_AFFINITY
597  KMP_AFFINITY_ENABLE(__kmp_num_proc_groups * sizeof(DWORD_PTR));
598 #else
599  KMP_AFFINITY_ENABLE(sizeof(DWORD_PTR));
600 #endif
601 
602  KA_TRACE(10, ("__kmp_affinity_determine_capable: "
603  "Windows* OS affinity interface functional (mask size = "
604  "%" KMP_SIZE_T_SPEC ").\n",
605  __kmp_affin_mask_size));
606 }
607 
608 double __kmp_read_cpu_time(void) {
609  FILETIME CreationTime, ExitTime, KernelTime, UserTime;
610  int status;
611  double cpu_time;
612 
613  cpu_time = 0;
614 
615  status = GetProcessTimes(GetCurrentProcess(), &CreationTime, &ExitTime,
616  &KernelTime, &UserTime);
617 
618  if (status) {
619  double sec = 0;
620 
621  sec += KernelTime.dwHighDateTime;
622  sec += UserTime.dwHighDateTime;
623 
624  /* Shift left by 32 bits */
625  sec *= (double)(1 << 16) * (double)(1 << 16);
626 
627  sec += KernelTime.dwLowDateTime;
628  sec += UserTime.dwLowDateTime;
629 
630  cpu_time += (sec * 100.0) / KMP_NSEC_PER_SEC;
631  }
632 
633  return cpu_time;
634 }
635 
636 int __kmp_read_system_info(struct kmp_sys_info *info) {
637  info->maxrss = 0; /* the maximum resident set size utilized (in kilobytes) */
638  info->minflt = 0; /* the number of page faults serviced without any I/O */
639  info->majflt = 0; /* the number of page faults serviced that required I/O */
640  info->nswap = 0; // the number of times a process was "swapped" out of memory
641  info->inblock = 0; // the number of times the file system had to perform input
642  info->oublock = 0; // number of times the file system had to perform output
643  info->nvcsw = 0; /* the number of times a context switch was voluntarily */
644  info->nivcsw = 0; /* the number of times a context switch was forced */
645 
646  return 1;
647 }
648 
649 void __kmp_runtime_initialize(void) {
650  SYSTEM_INFO info;
651  kmp_str_buf_t path;
652  UINT path_size;
653 
654  if (__kmp_init_runtime) {
655  return;
656  }
657 
658 #if KMP_DYNAMIC_LIB
659  /* Pin dynamic library for the lifetime of application */
660  {
661  // First, turn off error message boxes
662  UINT err_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
663  HMODULE h;
664  BOOL ret = GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS |
665  GET_MODULE_HANDLE_EX_FLAG_PIN,
666  (LPCTSTR)&__kmp_serial_initialize, &h);
667  KMP_DEBUG_ASSERT2(h && ret, "OpenMP RTL cannot find itself loaded");
668  SetErrorMode(err_mode); // Restore error mode
669  KA_TRACE(10, ("__kmp_runtime_initialize: dynamic library pinned\n"));
670  }
671 #endif
672 
673  InitializeCriticalSection(&__kmp_win32_section);
674 #if USE_ITT_BUILD
675  __kmp_itt_system_object_created(&__kmp_win32_section, "Critical Section");
676 #endif /* USE_ITT_BUILD */
677  __kmp_initialize_system_tick();
678 
679 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
680  if (!__kmp_cpuinfo.initialized) {
681  __kmp_query_cpuid(&__kmp_cpuinfo);
682  }
683 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
684 
685 /* Set up minimum number of threads to switch to TLS gtid */
686 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
687  // Windows* OS, static library.
688  /* New thread may use stack space previously used by another thread,
689  currently terminated. On Windows* OS, in case of static linking, we do not
690  know the moment of thread termination, and our structures (__kmp_threads
691  and __kmp_root arrays) are still keep info about dead threads. This leads
692  to problem in __kmp_get_global_thread_id() function: it wrongly finds gtid
693  (by searching through stack addresses of all known threads) for
694  unregistered foreign tread.
695 
696  Setting __kmp_tls_gtid_min to 0 workarounds this problem:
697  __kmp_get_global_thread_id() does not search through stacks, but get gtid
698  from TLS immediately.
699  --ln
700  */
701  __kmp_tls_gtid_min = 0;
702 #else
703  __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
704 #endif
705 
706  /* for the static library */
707  if (!__kmp_gtid_threadprivate_key) {
708  __kmp_gtid_threadprivate_key = TlsAlloc();
709  if (__kmp_gtid_threadprivate_key == TLS_OUT_OF_INDEXES) {
710  KMP_FATAL(TLSOutOfIndexes);
711  }
712  }
713 
714  // Load ntdll.dll.
715  /* Simple GetModuleHandle( "ntdll.dl" ) is not suitable due to security issue
716  (see http://www.microsoft.com/technet/security/advisory/2269637.mspx). We
717  have to specify full path to the library. */
718  __kmp_str_buf_init(&path);
719  path_size = GetSystemDirectory(path.str, path.size);
720  KMP_DEBUG_ASSERT(path_size > 0);
721  if (path_size >= path.size) {
722  // Buffer is too short. Expand the buffer and try again.
723  __kmp_str_buf_reserve(&path, path_size);
724  path_size = GetSystemDirectory(path.str, path.size);
725  KMP_DEBUG_ASSERT(path_size > 0);
726  }
727  if (path_size > 0 && path_size < path.size) {
728  // Now we have system directory name in the buffer.
729  // Append backslash and name of dll to form full path,
730  path.used = path_size;
731  __kmp_str_buf_print(&path, "\\%s", "ntdll.dll");
732 
733  // Now load ntdll using full path.
734  ntdll = GetModuleHandle(path.str);
735  }
736 
737  KMP_DEBUG_ASSERT(ntdll != NULL);
738  if (ntdll != NULL) {
739  NtQuerySystemInformation = (NtQuerySystemInformation_t)GetProcAddress(
740  ntdll, "NtQuerySystemInformation");
741  }
742  KMP_DEBUG_ASSERT(NtQuerySystemInformation != NULL);
743 
744 #if KMP_GROUP_AFFINITY
745  // Load kernel32.dll.
746  // Same caveat - must use full system path name.
747  if (path_size > 0 && path_size < path.size) {
748  // Truncate the buffer back to just the system path length,
749  // discarding "\\ntdll.dll", and replacing it with "kernel32.dll".
750  path.used = path_size;
751  __kmp_str_buf_print(&path, "\\%s", "kernel32.dll");
752 
753  // Load kernel32.dll using full path.
754  kernel32 = GetModuleHandle(path.str);
755  KA_TRACE(10, ("__kmp_runtime_initialize: kernel32.dll = %s\n", path.str));
756 
757  // Load the function pointers to kernel32.dll routines
758  // that may or may not exist on this system.
759  if (kernel32 != NULL) {
760  __kmp_GetActiveProcessorCount =
761  (kmp_GetActiveProcessorCount_t)GetProcAddress(
762  kernel32, "GetActiveProcessorCount");
763  __kmp_GetActiveProcessorGroupCount =
764  (kmp_GetActiveProcessorGroupCount_t)GetProcAddress(
765  kernel32, "GetActiveProcessorGroupCount");
766  __kmp_GetThreadGroupAffinity =
767  (kmp_GetThreadGroupAffinity_t)GetProcAddress(
768  kernel32, "GetThreadGroupAffinity");
769  __kmp_SetThreadGroupAffinity =
770  (kmp_SetThreadGroupAffinity_t)GetProcAddress(
771  kernel32, "SetThreadGroupAffinity");
772 
773  KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_GetActiveProcessorCount"
774  " = %p\n",
775  __kmp_GetActiveProcessorCount));
776  KA_TRACE(10, ("__kmp_runtime_initialize: "
777  "__kmp_GetActiveProcessorGroupCount = %p\n",
778  __kmp_GetActiveProcessorGroupCount));
779  KA_TRACE(10, ("__kmp_runtime_initialize:__kmp_GetThreadGroupAffinity"
780  " = %p\n",
781  __kmp_GetThreadGroupAffinity));
782  KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_SetThreadGroupAffinity"
783  " = %p\n",
784  __kmp_SetThreadGroupAffinity));
785  KA_TRACE(10, ("__kmp_runtime_initialize: sizeof(kmp_affin_mask_t) = %d\n",
786  sizeof(kmp_affin_mask_t)));
787 
788  // See if group affinity is supported on this system.
789  // If so, calculate the #groups and #procs.
790  //
791  // Group affinity was introduced with Windows* 7 OS and
792  // Windows* Server 2008 R2 OS.
793  if ((__kmp_GetActiveProcessorCount != NULL) &&
794  (__kmp_GetActiveProcessorGroupCount != NULL) &&
795  (__kmp_GetThreadGroupAffinity != NULL) &&
796  (__kmp_SetThreadGroupAffinity != NULL) &&
797  ((__kmp_num_proc_groups = __kmp_GetActiveProcessorGroupCount()) >
798  1)) {
799  // Calculate the total number of active OS procs.
800  int i;
801 
802  KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups"
803  " detected\n",
804  __kmp_num_proc_groups));
805 
806  __kmp_xproc = 0;
807 
808  for (i = 0; i < __kmp_num_proc_groups; i++) {
809  DWORD size = __kmp_GetActiveProcessorCount(i);
810  __kmp_xproc += size;
811  KA_TRACE(10, ("__kmp_runtime_initialize: proc group %d size = %d\n",
812  i, size));
813  }
814  } else {
815  KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups"
816  " detected\n",
817  __kmp_num_proc_groups));
818  }
819  }
820  }
821  if (__kmp_num_proc_groups <= 1) {
822  GetSystemInfo(&info);
823  __kmp_xproc = info.dwNumberOfProcessors;
824  }
825 #else
826  GetSystemInfo(&info);
827  __kmp_xproc = info.dwNumberOfProcessors;
828 #endif /* KMP_GROUP_AFFINITY */
829 
830  // If the OS said there were 0 procs, take a guess and use a value of 2.
831  // This is done for Linux* OS, also. Do we need error / warning?
832  if (__kmp_xproc <= 0) {
833  __kmp_xproc = 2;
834  }
835 
836  KA_TRACE(5,
837  ("__kmp_runtime_initialize: total processors = %d\n", __kmp_xproc));
838 
839  __kmp_str_buf_free(&path);
840 
841 #if USE_ITT_BUILD
842  __kmp_itt_initialize();
843 #endif /* USE_ITT_BUILD */
844 
845  __kmp_init_runtime = TRUE;
846 } // __kmp_runtime_initialize
847 
848 void __kmp_runtime_destroy(void) {
849  if (!__kmp_init_runtime) {
850  return;
851  }
852 
853 #if USE_ITT_BUILD
854  __kmp_itt_destroy();
855 #endif /* USE_ITT_BUILD */
856 
857  /* we can't DeleteCriticalsection( & __kmp_win32_section ); */
858  /* due to the KX_TRACE() commands */
859  KA_TRACE(40, ("__kmp_runtime_destroy\n"));
860 
861  if (__kmp_gtid_threadprivate_key) {
862  TlsFree(__kmp_gtid_threadprivate_key);
863  __kmp_gtid_threadprivate_key = 0;
864  }
865 
866  __kmp_affinity_uninitialize();
867  DeleteCriticalSection(&__kmp_win32_section);
868 
869  ntdll = NULL;
870  NtQuerySystemInformation = NULL;
871 
872 #if KMP_ARCH_X86_64
873  kernel32 = NULL;
874  __kmp_GetActiveProcessorCount = NULL;
875  __kmp_GetActiveProcessorGroupCount = NULL;
876  __kmp_GetThreadGroupAffinity = NULL;
877  __kmp_SetThreadGroupAffinity = NULL;
878 #endif // KMP_ARCH_X86_64
879 
880  __kmp_init_runtime = FALSE;
881 }
882 
883 void __kmp_terminate_thread(int gtid) {
884  kmp_info_t *th = __kmp_threads[gtid];
885 
886  if (!th)
887  return;
888 
889  KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
890 
891  if (TerminateThread(th->th.th_info.ds.ds_thread, (DWORD)-1) == FALSE) {
892  /* It's OK, the thread may have exited already */
893  }
894  __kmp_free_handle(th->th.th_info.ds.ds_thread);
895 }
896 
897 void __kmp_clear_system_time(void) {
898  BOOL status;
899  LARGE_INTEGER time;
900  status = QueryPerformanceCounter(&time);
901  __kmp_win32_time = (kmp_int64)time.QuadPart;
902 }
903 
904 void __kmp_initialize_system_tick(void) {
905  {
906  BOOL status;
907  LARGE_INTEGER freq;
908 
909  status = QueryPerformanceFrequency(&freq);
910  if (!status) {
911  DWORD error = GetLastError();
912  __kmp_fatal(KMP_MSG(FunctionError, "QueryPerformanceFrequency()"),
913  KMP_ERR(error), __kmp_msg_null);
914 
915  } else {
916  __kmp_win32_tick = ((double)1.0) / (double)freq.QuadPart;
917  }
918  }
919 }
920 
921 /* Calculate the elapsed wall clock time for the user */
922 
923 void __kmp_elapsed(double *t) {
924  BOOL status;
925  LARGE_INTEGER now;
926  status = QueryPerformanceCounter(&now);
927  *t = ((double)now.QuadPart) * __kmp_win32_tick;
928 }
929 
930 /* Calculate the elapsed wall clock tick for the user */
931 
932 void __kmp_elapsed_tick(double *t) { *t = __kmp_win32_tick; }
933 
934 void __kmp_read_system_time(double *delta) {
935  if (delta != NULL) {
936  BOOL status;
937  LARGE_INTEGER now;
938 
939  status = QueryPerformanceCounter(&now);
940 
941  *delta = ((double)(((kmp_int64)now.QuadPart) - __kmp_win32_time)) *
942  __kmp_win32_tick;
943  }
944 }
945 
946 /* Return the current time stamp in nsec */
947 kmp_uint64 __kmp_now_nsec() {
948  LARGE_INTEGER now;
949  QueryPerformanceCounter(&now);
950  return 1e9 * __kmp_win32_tick * now.QuadPart;
951 }
952 
953 extern "C"
954 void *__stdcall __kmp_launch_worker(void *arg) {
955  volatile void *stack_data;
956  void *exit_val;
957  void *padding = 0;
958  kmp_info_t *this_thr = (kmp_info_t *)arg;
959  int gtid;
960 
961  gtid = this_thr->th.th_info.ds.ds_gtid;
962  __kmp_gtid_set_specific(gtid);
963 #ifdef KMP_TDATA_GTID
964 #error "This define causes problems with LoadLibrary() + declspec(thread) " \
965  "on Windows* OS. See CQ50564, tests kmp_load_library*.c and this MSDN " \
966  "reference: http://support.microsoft.com/kb/118816"
967 //__kmp_gtid = gtid;
968 #endif
969 
970 #if USE_ITT_BUILD
971  __kmp_itt_thread_name(gtid);
972 #endif /* USE_ITT_BUILD */
973 
974  __kmp_affinity_set_init_mask(gtid, FALSE);
975 
976 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
977  // Set FP control regs to be a copy of the parallel initialization thread's.
978  __kmp_clear_x87_fpu_status_word();
979  __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
980  __kmp_load_mxcsr(&__kmp_init_mxcsr);
981 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
982 
983  if (__kmp_stkoffset > 0 && gtid > 0) {
984  padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
985  }
986 
987  KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive);
988  this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
989  TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE);
990 
991  if (TCR_4(__kmp_gtid_mode) <
992  2) { // check stack only if it is used to get gtid
993  TCW_PTR(this_thr->th.th_info.ds.ds_stackbase, &stack_data);
994  KMP_ASSERT(this_thr->th.th_info.ds.ds_stackgrow == FALSE);
995  __kmp_check_stack_overlap(this_thr);
996  }
997  KMP_MB();
998  exit_val = __kmp_launch_thread(this_thr);
999  KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive);
1000  TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE);
1001  KMP_MB();
1002  return exit_val;
1003 }
1004 
1005 #if KMP_USE_MONITOR
1006 /* The monitor thread controls all of the threads in the complex */
1007 
1008 void *__stdcall __kmp_launch_monitor(void *arg) {
1009  DWORD wait_status;
1010  kmp_thread_t monitor;
1011  int status;
1012  int interval;
1013  kmp_info_t *this_thr = (kmp_info_t *)arg;
1014 
1015  KMP_DEBUG_ASSERT(__kmp_init_monitor);
1016  TCW_4(__kmp_init_monitor, 2); // AC: Signal library that monitor has started
1017  // TODO: hide "2" in enum (like {true,false,started})
1018  this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
1019  TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE);
1020 
1021  KMP_MB(); /* Flush all pending memory write invalidates. */
1022  KA_TRACE(10, ("__kmp_launch_monitor: launched\n"));
1023 
1024  monitor = GetCurrentThread();
1025 
1026  /* set thread priority */
1027  status = SetThreadPriority(monitor, THREAD_PRIORITY_HIGHEST);
1028  if (!status) {
1029  DWORD error = GetLastError();
1030  __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null);
1031  }
1032 
1033  /* register us as monitor */
1034  __kmp_gtid_set_specific(KMP_GTID_MONITOR);
1035 #ifdef KMP_TDATA_GTID
1036 #error "This define causes problems with LoadLibrary() + declspec(thread) " \
1037  "on Windows* OS. See CQ50564, tests kmp_load_library*.c and this MSDN " \
1038  "reference: http://support.microsoft.com/kb/118816"
1039 //__kmp_gtid = KMP_GTID_MONITOR;
1040 #endif
1041 
1042 #if USE_ITT_BUILD
1043  __kmp_itt_thread_ignore(); // Instruct Intel(R) Threading Tools to ignore
1044 // monitor thread.
1045 #endif /* USE_ITT_BUILD */
1046 
1047  KMP_MB(); /* Flush all pending memory write invalidates. */
1048 
1049  interval = (1000 / __kmp_monitor_wakeups); /* in milliseconds */
1050 
1051  while (!TCR_4(__kmp_global.g.g_done)) {
1052  /* This thread monitors the state of the system */
1053 
1054  KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
1055 
1056  wait_status = WaitForSingleObject(__kmp_monitor_ev, interval);
1057 
1058  if (wait_status == WAIT_TIMEOUT) {
1059  TCW_4(__kmp_global.g.g_time.dt.t_value,
1060  TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
1061  }
1062 
1063  KMP_MB(); /* Flush all pending memory write invalidates. */
1064  }
1065 
1066  KA_TRACE(10, ("__kmp_launch_monitor: finished\n"));
1067 
1068  status = SetThreadPriority(monitor, THREAD_PRIORITY_NORMAL);
1069  if (!status) {
1070  DWORD error = GetLastError();
1071  __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null);
1072  }
1073 
1074  if (__kmp_global.g.g_abort != 0) {
1075  /* now we need to terminate the worker threads */
1076  /* the value of t_abort is the signal we caught */
1077  int gtid;
1078 
1079  KA_TRACE(10, ("__kmp_launch_monitor: terminate sig=%d\n",
1080  (__kmp_global.g.g_abort)));
1081 
1082  /* terminate the OpenMP worker threads */
1083  /* TODO this is not valid for sibling threads!!
1084  * the uber master might not be 0 anymore.. */
1085  for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
1086  __kmp_terminate_thread(gtid);
1087 
1088  __kmp_cleanup();
1089 
1090  Sleep(0);
1091 
1092  KA_TRACE(10,
1093  ("__kmp_launch_monitor: raise sig=%d\n", __kmp_global.g.g_abort));
1094 
1095  if (__kmp_global.g.g_abort > 0) {
1096  raise(__kmp_global.g.g_abort);
1097  }
1098  }
1099 
1100  TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE);
1101 
1102  KMP_MB();
1103  return arg;
1104 }
1105 #endif
1106 
1107 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
1108  kmp_thread_t handle;
1109  DWORD idThread;
1110 
1111  KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
1112 
1113  th->th.th_info.ds.ds_gtid = gtid;
1114 
1115  if (KMP_UBER_GTID(gtid)) {
1116  int stack_data;
1117 
1118  /* TODO: GetCurrentThread() returns a pseudo-handle that is unsuitable for
1119  other threads to use. Is it appropriate to just use GetCurrentThread?
1120  When should we close this handle? When unregistering the root? */
1121  {
1122  BOOL rc;
1123  rc = DuplicateHandle(GetCurrentProcess(), GetCurrentThread(),
1124  GetCurrentProcess(), &th->th.th_info.ds.ds_thread, 0,
1125  FALSE, DUPLICATE_SAME_ACCESS);
1126  KMP_ASSERT(rc);
1127  KA_TRACE(10, (" __kmp_create_worker: ROOT Handle duplicated, th = %p, "
1128  "handle = %" KMP_UINTPTR_SPEC "\n",
1129  (LPVOID)th, th->th.th_info.ds.ds_thread));
1130  th->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
1131  }
1132  if (TCR_4(__kmp_gtid_mode) < 2) { // check stack only if used to get gtid
1133  /* we will dynamically update the stack range if gtid_mode == 1 */
1134  TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
1135  TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
1136  TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
1137  __kmp_check_stack_overlap(th);
1138  }
1139  } else {
1140  KMP_MB(); /* Flush all pending memory write invalidates. */
1141 
1142  /* Set stack size for this thread now. */
1143  KA_TRACE(10,
1144  ("__kmp_create_worker: stack_size = %" KMP_SIZE_T_SPEC " bytes\n",
1145  stack_size));
1146 
1147  stack_size += gtid * __kmp_stkoffset;
1148 
1149  TCW_PTR(th->th.th_info.ds.ds_stacksize, stack_size);
1150  TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
1151 
1152  KA_TRACE(10,
1153  ("__kmp_create_worker: (before) stack_size = %" KMP_SIZE_T_SPEC
1154  " bytes, &__kmp_launch_worker = %p, th = %p, &idThread = %p\n",
1155  (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker,
1156  (LPVOID)th, &idThread));
1157 
1158  handle = CreateThread(
1159  NULL, (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)__kmp_launch_worker,
1160  (LPVOID)th, STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread);
1161 
1162  KA_TRACE(10,
1163  ("__kmp_create_worker: (after) stack_size = %" KMP_SIZE_T_SPEC
1164  " bytes, &__kmp_launch_worker = %p, th = %p, "
1165  "idThread = %u, handle = %" KMP_UINTPTR_SPEC "\n",
1166  (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker,
1167  (LPVOID)th, idThread, handle));
1168 
1169  if (handle == 0) {
1170  DWORD error = GetLastError();
1171  __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null);
1172  } else {
1173  th->th.th_info.ds.ds_thread = handle;
1174  }
1175 
1176  KMP_MB(); /* Flush all pending memory write invalidates. */
1177  }
1178 
1179  KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
1180 }
1181 
1182 int __kmp_still_running(kmp_info_t *th) {
1183  return (WAIT_TIMEOUT == WaitForSingleObject(th->th.th_info.ds.ds_thread, 0));
1184 }
1185 
1186 #if KMP_USE_MONITOR
1187 void __kmp_create_monitor(kmp_info_t *th) {
1188  kmp_thread_t handle;
1189  DWORD idThread;
1190  int ideal, new_ideal;
1191 
1192  if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
1193  // We don't need monitor thread in case of MAX_BLOCKTIME
1194  KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
1195  "MAX blocktime\n"));
1196  th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
1197  th->th.th_info.ds.ds_gtid = 0;
1198  TCW_4(__kmp_init_monitor, 2); // Signal to stop waiting for monitor creation
1199  return;
1200  }
1201  KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
1202 
1203  KMP_MB(); /* Flush all pending memory write invalidates. */
1204 
1205  __kmp_monitor_ev = CreateEvent(NULL, TRUE, FALSE, NULL);
1206  if (__kmp_monitor_ev == NULL) {
1207  DWORD error = GetLastError();
1208  __kmp_fatal(KMP_MSG(CantCreateEvent), KMP_ERR(error), __kmp_msg_null);
1209  }
1210 #if USE_ITT_BUILD
1211  __kmp_itt_system_object_created(__kmp_monitor_ev, "Event");
1212 #endif /* USE_ITT_BUILD */
1213 
1214  th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
1215  th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
1216 
1217  // FIXME - on Windows* OS, if __kmp_monitor_stksize = 0, figure out how
1218  // to automatically expand stacksize based on CreateThread error code.
1219  if (__kmp_monitor_stksize == 0) {
1220  __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
1221  }
1222  if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
1223  __kmp_monitor_stksize = __kmp_sys_min_stksize;
1224  }
1225 
1226  KA_TRACE(10, ("__kmp_create_monitor: requested stacksize = %d bytes\n",
1227  (int)__kmp_monitor_stksize));
1228 
1229  TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
1230 
1231  handle =
1232  CreateThread(NULL, (SIZE_T)__kmp_monitor_stksize,
1233  (LPTHREAD_START_ROUTINE)__kmp_launch_monitor, (LPVOID)th,
1234  STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread);
1235  if (handle == 0) {
1236  DWORD error = GetLastError();
1237  __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null);
1238  } else
1239  th->th.th_info.ds.ds_thread = handle;
1240 
1241  KMP_MB(); /* Flush all pending memory write invalidates. */
1242 
1243  KA_TRACE(10, ("__kmp_create_monitor: monitor created %p\n",
1244  (void *)th->th.th_info.ds.ds_thread));
1245 }
1246 #endif
1247 
1248 /* Check to see if thread is still alive.
1249  NOTE: The ExitProcess(code) system call causes all threads to Terminate
1250  with a exit_val = code. Because of this we can not rely on exit_val having
1251  any particular value. So this routine may return STILL_ALIVE in exit_val
1252  even after the thread is dead. */
1253 
1254 int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val) {
1255  DWORD rc;
1256  rc = GetExitCodeThread(th->th.th_info.ds.ds_thread, exit_val);
1257  if (rc == 0) {
1258  DWORD error = GetLastError();
1259  __kmp_fatal(KMP_MSG(FunctionError, "GetExitCodeThread()"), KMP_ERR(error),
1260  __kmp_msg_null);
1261  }
1262  return (*exit_val == STILL_ACTIVE);
1263 }
1264 
1265 void __kmp_exit_thread(int exit_status) {
1266  ExitThread(exit_status);
1267 } // __kmp_exit_thread
1268 
1269 // This is a common part for both __kmp_reap_worker() and __kmp_reap_monitor().
1270 static void __kmp_reap_common(kmp_info_t *th) {
1271  DWORD exit_val;
1272 
1273  KMP_MB(); /* Flush all pending memory write invalidates. */
1274 
1275  KA_TRACE(
1276  10, ("__kmp_reap_common: try to reap (%d)\n", th->th.th_info.ds.ds_gtid));
1277 
1278  /* 2006-10-19:
1279  There are two opposite situations:
1280  1. Windows* OS keep thread alive after it resets ds_alive flag and
1281  exits from thread function. (For example, see C70770/Q394281 "unloading of
1282  dll based on OMP is very slow".)
1283  2. Windows* OS may kill thread before it resets ds_alive flag.
1284 
1285  Right solution seems to be waiting for *either* thread termination *or*
1286  ds_alive resetting. */
1287  {
1288  // TODO: This code is very similar to KMP_WAIT. Need to generalize
1289  // KMP_WAIT to cover this usage also.
1290  void *obj = NULL;
1291  kmp_uint32 spins;
1292 #if USE_ITT_BUILD
1293  KMP_FSYNC_SPIN_INIT(obj, (void *)&th->th.th_info.ds.ds_alive);
1294 #endif /* USE_ITT_BUILD */
1295  KMP_INIT_YIELD(spins);
1296  do {
1297 #if USE_ITT_BUILD
1298  KMP_FSYNC_SPIN_PREPARE(obj);
1299 #endif /* USE_ITT_BUILD */
1300  __kmp_is_thread_alive(th, &exit_val);
1301  KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
1302  } while (exit_val == STILL_ACTIVE && TCR_4(th->th.th_info.ds.ds_alive));
1303 #if USE_ITT_BUILD
1304  if (exit_val == STILL_ACTIVE) {
1305  KMP_FSYNC_CANCEL(obj);
1306  } else {
1307  KMP_FSYNC_SPIN_ACQUIRED(obj);
1308  }
1309 #endif /* USE_ITT_BUILD */
1310  }
1311 
1312  __kmp_free_handle(th->th.th_info.ds.ds_thread);
1313 
1314  /* NOTE: The ExitProcess(code) system call causes all threads to Terminate
1315  with a exit_val = code. Because of this we can not rely on exit_val having
1316  any particular value. */
1317  if (exit_val == STILL_ACTIVE) {
1318  KA_TRACE(1, ("__kmp_reap_common: thread still active.\n"));
1319  } else if ((void *)exit_val != (void *)th) {
1320  KA_TRACE(1, ("__kmp_reap_common: ExitProcess / TerminateThread used?\n"));
1321  }
1322 
1323  KA_TRACE(10,
1324  ("__kmp_reap_common: done reaping (%d), handle = %" KMP_UINTPTR_SPEC
1325  "\n",
1326  th->th.th_info.ds.ds_gtid, th->th.th_info.ds.ds_thread));
1327 
1328  th->th.th_info.ds.ds_thread = 0;
1329  th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1330  th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1331  th->th.th_info.ds.ds_thread_id = 0;
1332 
1333  KMP_MB(); /* Flush all pending memory write invalidates. */
1334 }
1335 
1336 #if KMP_USE_MONITOR
1337 void __kmp_reap_monitor(kmp_info_t *th) {
1338  int status;
1339 
1340  KA_TRACE(10, ("__kmp_reap_monitor: try to reap %p\n",
1341  (void *)th->th.th_info.ds.ds_thread));
1342 
1343  // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1344  // If both tid and gtid are 0, it means the monitor did not ever start.
1345  // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1346  KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1347  if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1348  KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1349  return;
1350  }
1351 
1352  KMP_MB(); /* Flush all pending memory write invalidates. */
1353 
1354  status = SetEvent(__kmp_monitor_ev);
1355  if (status == FALSE) {
1356  DWORD error = GetLastError();
1357  __kmp_fatal(KMP_MSG(CantSetEvent), KMP_ERR(error), __kmp_msg_null);
1358  }
1359  KA_TRACE(10, ("__kmp_reap_monitor: reaping thread (%d)\n",
1360  th->th.th_info.ds.ds_gtid));
1361  __kmp_reap_common(th);
1362 
1363  __kmp_free_handle(__kmp_monitor_ev);
1364 
1365  KMP_MB(); /* Flush all pending memory write invalidates. */
1366 }
1367 #endif
1368 
1369 void __kmp_reap_worker(kmp_info_t *th) {
1370  KA_TRACE(10, ("__kmp_reap_worker: reaping thread (%d)\n",
1371  th->th.th_info.ds.ds_gtid));
1372  __kmp_reap_common(th);
1373 }
1374 
1375 #if KMP_HANDLE_SIGNALS
1376 
1377 static void __kmp_team_handler(int signo) {
1378  if (__kmp_global.g.g_abort == 0) {
1379  // Stage 1 signal handler, let's shut down all of the threads.
1380  if (__kmp_debug_buf) {
1381  __kmp_dump_debug_buffer();
1382  }
1383  KMP_MB(); // Flush all pending memory write invalidates.
1384  TCW_4(__kmp_global.g.g_abort, signo);
1385  KMP_MB(); // Flush all pending memory write invalidates.
1386  TCW_4(__kmp_global.g.g_done, TRUE);
1387  KMP_MB(); // Flush all pending memory write invalidates.
1388  }
1389 } // __kmp_team_handler
1390 
1391 static sig_func_t __kmp_signal(int signum, sig_func_t handler) {
1392  sig_func_t old = signal(signum, handler);
1393  if (old == SIG_ERR) {
1394  int error = errno;
1395  __kmp_fatal(KMP_MSG(FunctionError, "signal"), KMP_ERR(error),
1396  __kmp_msg_null);
1397  }
1398  return old;
1399 }
1400 
1401 static void __kmp_install_one_handler(int sig, sig_func_t handler,
1402  int parallel_init) {
1403  sig_func_t old;
1404  KMP_MB(); /* Flush all pending memory write invalidates. */
1405  KB_TRACE(60, ("__kmp_install_one_handler: called: sig=%d\n", sig));
1406  if (parallel_init) {
1407  old = __kmp_signal(sig, handler);
1408  // SIG_DFL on Windows* OS in NULL or 0.
1409  if (old == __kmp_sighldrs[sig]) {
1410  __kmp_siginstalled[sig] = 1;
1411  } else { // Restore/keep user's handler if one previously installed.
1412  old = __kmp_signal(sig, old);
1413  }
1414  } else {
1415  // Save initial/system signal handlers to see if user handlers installed.
1416  // 2009-09-23: It is a dead code. On Windows* OS __kmp_install_signals
1417  // called once with parallel_init == TRUE.
1418  old = __kmp_signal(sig, SIG_DFL);
1419  __kmp_sighldrs[sig] = old;
1420  __kmp_signal(sig, old);
1421  }
1422  KMP_MB(); /* Flush all pending memory write invalidates. */
1423 } // __kmp_install_one_handler
1424 
1425 static void __kmp_remove_one_handler(int sig) {
1426  if (__kmp_siginstalled[sig]) {
1427  sig_func_t old;
1428  KMP_MB(); // Flush all pending memory write invalidates.
1429  KB_TRACE(60, ("__kmp_remove_one_handler: called: sig=%d\n", sig));
1430  old = __kmp_signal(sig, __kmp_sighldrs[sig]);
1431  if (old != __kmp_team_handler) {
1432  KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1433  "restoring: sig=%d\n",
1434  sig));
1435  old = __kmp_signal(sig, old);
1436  }
1437  __kmp_sighldrs[sig] = NULL;
1438  __kmp_siginstalled[sig] = 0;
1439  KMP_MB(); // Flush all pending memory write invalidates.
1440  }
1441 } // __kmp_remove_one_handler
1442 
1443 void __kmp_install_signals(int parallel_init) {
1444  KB_TRACE(10, ("__kmp_install_signals: called\n"));
1445  if (!__kmp_handle_signals) {
1446  KB_TRACE(10, ("__kmp_install_signals: KMP_HANDLE_SIGNALS is false - "
1447  "handlers not installed\n"));
1448  return;
1449  }
1450  __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1451  __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1452  __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1453  __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1454  __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1455  __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1456 } // __kmp_install_signals
1457 
1458 void __kmp_remove_signals(void) {
1459  int sig;
1460  KB_TRACE(10, ("__kmp_remove_signals: called\n"));
1461  for (sig = 1; sig < NSIG; ++sig) {
1462  __kmp_remove_one_handler(sig);
1463  }
1464 } // __kmp_remove_signals
1465 
1466 #endif // KMP_HANDLE_SIGNALS
1467 
1468 /* Put the thread to sleep for a time period */
1469 void __kmp_thread_sleep(int millis) {
1470  DWORD status;
1471 
1472  status = SleepEx((DWORD)millis, FALSE);
1473  if (status) {
1474  DWORD error = GetLastError();
1475  __kmp_fatal(KMP_MSG(FunctionError, "SleepEx()"), KMP_ERR(error),
1476  __kmp_msg_null);
1477  }
1478 }
1479 
1480 // Determine whether the given address is mapped into the current address space.
1481 int __kmp_is_address_mapped(void *addr) {
1482  DWORD status;
1483  MEMORY_BASIC_INFORMATION lpBuffer;
1484  SIZE_T dwLength;
1485 
1486  dwLength = sizeof(MEMORY_BASIC_INFORMATION);
1487 
1488  status = VirtualQuery(addr, &lpBuffer, dwLength);
1489 
1490  return !(((lpBuffer.State == MEM_RESERVE) || (lpBuffer.State == MEM_FREE)) ||
1491  ((lpBuffer.Protect == PAGE_NOACCESS) ||
1492  (lpBuffer.Protect == PAGE_EXECUTE)));
1493 }
1494 
1495 kmp_uint64 __kmp_hardware_timestamp(void) {
1496  kmp_uint64 r = 0;
1497 
1498  QueryPerformanceCounter((LARGE_INTEGER *)&r);
1499  return r;
1500 }
1501 
1502 /* Free handle and check the error code */
1503 void __kmp_free_handle(kmp_thread_t tHandle) {
1504  /* called with parameter type HANDLE also, thus suppose kmp_thread_t defined
1505  * as HANDLE */
1506  BOOL rc;
1507  rc = CloseHandle(tHandle);
1508  if (!rc) {
1509  DWORD error = GetLastError();
1510  __kmp_fatal(KMP_MSG(CantCloseHandle), KMP_ERR(error), __kmp_msg_null);
1511  }
1512 }
1513 
1514 int __kmp_get_load_balance(int max) {
1515  static ULONG glb_buff_size = 100 * 1024;
1516 
1517  // Saved count of the running threads for the thread balance algorithm
1518  static int glb_running_threads = 0;
1519  static double glb_call_time = 0; /* Thread balance algorithm call time */
1520 
1521  int running_threads = 0; // Number of running threads in the system.
1522  NTSTATUS status = 0;
1523  ULONG buff_size = 0;
1524  ULONG info_size = 0;
1525  void *buffer = NULL;
1526  PSYSTEM_PROCESS_INFORMATION spi = NULL;
1527  int first_time = 1;
1528 
1529  double call_time = 0.0; // start, finish;
1530 
1531  __kmp_elapsed(&call_time);
1532 
1533  if (glb_call_time &&
1534  (call_time - glb_call_time < __kmp_load_balance_interval)) {
1535  running_threads = glb_running_threads;
1536  goto finish;
1537  }
1538  glb_call_time = call_time;
1539 
1540  // Do not spend time on running algorithm if we have a permanent error.
1541  if (NtQuerySystemInformation == NULL) {
1542  running_threads = -1;
1543  goto finish;
1544  }
1545 
1546  if (max <= 0) {
1547  max = INT_MAX;
1548  }
1549 
1550  do {
1551 
1552  if (first_time) {
1553  buff_size = glb_buff_size;
1554  } else {
1555  buff_size = 2 * buff_size;
1556  }
1557 
1558  buffer = KMP_INTERNAL_REALLOC(buffer, buff_size);
1559  if (buffer == NULL) {
1560  running_threads = -1;
1561  goto finish;
1562  }
1563  status = NtQuerySystemInformation(SystemProcessInformation, buffer,
1564  buff_size, &info_size);
1565  first_time = 0;
1566 
1567  } while (status == STATUS_INFO_LENGTH_MISMATCH);
1568  glb_buff_size = buff_size;
1569 
1570 #define CHECK(cond) \
1571  { \
1572  KMP_DEBUG_ASSERT(cond); \
1573  if (!(cond)) { \
1574  running_threads = -1; \
1575  goto finish; \
1576  } \
1577  }
1578 
1579  CHECK(buff_size >= info_size);
1580  spi = PSYSTEM_PROCESS_INFORMATION(buffer);
1581  for (;;) {
1582  ptrdiff_t offset = uintptr_t(spi) - uintptr_t(buffer);
1583  CHECK(0 <= offset &&
1584  offset + sizeof(SYSTEM_PROCESS_INFORMATION) < info_size);
1585  HANDLE pid = spi->ProcessId;
1586  ULONG num = spi->NumberOfThreads;
1587  CHECK(num >= 1);
1588  size_t spi_size =
1589  sizeof(SYSTEM_PROCESS_INFORMATION) + sizeof(SYSTEM_THREAD) * (num - 1);
1590  CHECK(offset + spi_size <
1591  info_size); // Make sure process info record fits the buffer.
1592  if (spi->NextEntryOffset != 0) {
1593  CHECK(spi_size <=
1594  spi->NextEntryOffset); // And do not overlap with the next record.
1595  }
1596  // pid == 0 corresponds to the System Idle Process. It always has running
1597  // threads on all cores. So, we don't consider the running threads of this
1598  // process.
1599  if (pid != 0) {
1600  for (int i = 0; i < num; ++i) {
1601  THREAD_STATE state = spi->Threads[i].State;
1602  // Count threads that have Ready or Running state.
1603  // !!! TODO: Why comment does not match the code???
1604  if (state == StateRunning) {
1605  ++running_threads;
1606  // Stop counting running threads if the number is already greater than
1607  // the number of available cores
1608  if (running_threads >= max) {
1609  goto finish;
1610  }
1611  }
1612  }
1613  }
1614  if (spi->NextEntryOffset == 0) {
1615  break;
1616  }
1617  spi = PSYSTEM_PROCESS_INFORMATION(uintptr_t(spi) + spi->NextEntryOffset);
1618  }
1619 
1620 #undef CHECK
1621 
1622 finish: // Clean up and exit.
1623 
1624  if (buffer != NULL) {
1625  KMP_INTERNAL_FREE(buffer);
1626  }
1627 
1628  glb_running_threads = running_threads;
1629 
1630  return running_threads;
1631 } //__kmp_get_load_balance()
1632 
1633 // Functions for hidden helper task
1634 void __kmp_hidden_helper_worker_thread_wait() {
1635  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1636 }
1637 
1638 void __kmp_do_initialize_hidden_helper_threads() {
1639  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1640 }
1641 
1642 void __kmp_hidden_helper_threads_initz_wait() {
1643  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1644 }
1645 
1646 void __kmp_hidden_helper_initz_release() {
1647  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1648 }
1649 
1650 void __kmp_hidden_helper_main_thread_wait() {
1651  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1652 }
1653 
1654 void __kmp_hidden_helper_main_thread_release() {
1655  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1656 }
1657 
1658 void __kmp_hidden_helper_worker_thread_signal() {
1659  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1660 }
1661 
1662 void __kmp_hidden_helper_threads_deinitz_wait() {
1663  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1664 }
1665 
1666 void __kmp_hidden_helper_threads_deinitz_release() {
1667  KMP_ASSERT(0 && "Hidden helper task is not supported on Windows");
1668 }