xref: /freebsd/contrib/llvm-project/openmp/runtime/src/z_Windows_NT_util.cpp (revision 6966ac055c3b7a39266fb982493330df7a097997)
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_win32_mutex_lock(&th->th.th_suspend_mx);
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_win32_mutex_unlock(&th->th.th_suspend_mx);
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_win32_mutex_unlock(&th->th.th_suspend_mx);
441 
442   KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid));
443 }
444 
445 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) {
446   __kmp_suspend_template(th_gtid, flag);
447 }
448 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) {
449   __kmp_suspend_template(th_gtid, flag);
450 }
451 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) {
452   __kmp_suspend_template(th_gtid, flag);
453 }
454 
455 /* This routine signals the thread specified by target_gtid to wake up
456    after setting the sleep bit indicated by the flag argument to FALSE */
457 template <class C>
458 static inline void __kmp_resume_template(int target_gtid, C *flag) {
459   kmp_info_t *th = __kmp_threads[target_gtid];
460   int status;
461 
462 #ifdef KMP_DEBUG
463   int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1;
464 #endif
465 
466   KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n",
467                 gtid, target_gtid));
468 
469   __kmp_suspend_initialize_thread(th);
470   __kmp_win32_mutex_lock(&th->th.th_suspend_mx);
471 
472   if (!flag) { // coming from __kmp_null_resume_wrapper
473     flag = (C *)th->th.th_sleep_loc;
474   }
475 
476   // First, check if the flag is null or its type has changed. If so, someone
477   // else woke it up.
478   if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type
479     // simply shows what
480     // flag was cast to
481     KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
482                  "awake: flag's loc(%p)\n",
483                  gtid, target_gtid, NULL));
484     __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
485     return;
486   } else {
487     typename C::flag_t old_spin = flag->unset_sleeping();
488     if (!flag->is_sleeping_val(old_spin)) {
489       KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already "
490                    "awake: flag's loc(%p): %u => %u\n",
491                    gtid, target_gtid, flag->get(), old_spin, *(flag->get())));
492       __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
493       return;
494     }
495   }
496   TCW_PTR(th->th.th_sleep_loc, NULL);
497   KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset sleep "
498                "bit for flag's loc(%p)\n",
499                gtid, target_gtid, flag->get()));
500 
501   __kmp_win32_cond_signal(&th->th.th_suspend_cv);
502   __kmp_win32_mutex_unlock(&th->th.th_suspend_mx);
503 
504   KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up"
505                 " for T#%d\n",
506                 gtid, target_gtid));
507 }
508 
509 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) {
510   __kmp_resume_template(target_gtid, flag);
511 }
512 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) {
513   __kmp_resume_template(target_gtid, flag);
514 }
515 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) {
516   __kmp_resume_template(target_gtid, flag);
517 }
518 
519 void __kmp_yield() { Sleep(0); }
520 
521 void __kmp_gtid_set_specific(int gtid) {
522   if (__kmp_init_gtid) {
523     KA_TRACE(50, ("__kmp_gtid_set_specific: T#%d key:%d\n", gtid,
524                   __kmp_gtid_threadprivate_key));
525     if (!TlsSetValue(__kmp_gtid_threadprivate_key, (LPVOID)(gtid + 1)))
526       KMP_FATAL(TLSSetValueFailed);
527   } else {
528     KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n"));
529   }
530 }
531 
532 int __kmp_gtid_get_specific() {
533   int gtid;
534   if (!__kmp_init_gtid) {
535     KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning "
536                   "KMP_GTID_SHUTDOWN\n"));
537     return KMP_GTID_SHUTDOWN;
538   }
539   gtid = (int)(kmp_intptr_t)TlsGetValue(__kmp_gtid_threadprivate_key);
540   if (gtid == 0) {
541     gtid = KMP_GTID_DNE;
542   } else {
543     gtid--;
544   }
545   KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n",
546                 __kmp_gtid_threadprivate_key, gtid));
547   return gtid;
548 }
549 
550 void __kmp_affinity_bind_thread(int proc) {
551   if (__kmp_num_proc_groups > 1) {
552     // Form the GROUP_AFFINITY struct directly, rather than filling
553     // out a bit vector and calling __kmp_set_system_affinity().
554     GROUP_AFFINITY ga;
555     KMP_DEBUG_ASSERT((proc >= 0) && (proc < (__kmp_num_proc_groups * CHAR_BIT *
556                                              sizeof(DWORD_PTR))));
557     ga.Group = proc / (CHAR_BIT * sizeof(DWORD_PTR));
558     ga.Mask = (unsigned long long)1 << (proc % (CHAR_BIT * sizeof(DWORD_PTR)));
559     ga.Reserved[0] = ga.Reserved[1] = ga.Reserved[2] = 0;
560 
561     KMP_DEBUG_ASSERT(__kmp_SetThreadGroupAffinity != NULL);
562     if (__kmp_SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL) == 0) {
563       DWORD error = GetLastError();
564       if (__kmp_affinity_verbose) { // AC: continue silently if not verbose
565         kmp_msg_t err_code = KMP_ERR(error);
566         __kmp_msg(kmp_ms_warning, KMP_MSG(CantSetThreadAffMask), err_code,
567                   __kmp_msg_null);
568         if (__kmp_generate_warnings == kmp_warnings_off) {
569           __kmp_str_free(&err_code.str);
570         }
571       }
572     }
573   } else {
574     kmp_affin_mask_t *mask;
575     KMP_CPU_ALLOC_ON_STACK(mask);
576     KMP_CPU_ZERO(mask);
577     KMP_CPU_SET(proc, mask);
578     __kmp_set_system_affinity(mask, TRUE);
579     KMP_CPU_FREE_FROM_STACK(mask);
580   }
581 }
582 
583 void __kmp_affinity_determine_capable(const char *env_var) {
584 // All versions of Windows* OS (since Win '95) support SetThreadAffinityMask().
585 
586 #if KMP_GROUP_AFFINITY
587   KMP_AFFINITY_ENABLE(__kmp_num_proc_groups * sizeof(DWORD_PTR));
588 #else
589   KMP_AFFINITY_ENABLE(sizeof(DWORD_PTR));
590 #endif
591 
592   KA_TRACE(10, ("__kmp_affinity_determine_capable: "
593                 "Windows* OS affinity interface functional (mask size = "
594                 "%" KMP_SIZE_T_SPEC ").\n",
595                 __kmp_affin_mask_size));
596 }
597 
598 double __kmp_read_cpu_time(void) {
599   FILETIME CreationTime, ExitTime, KernelTime, UserTime;
600   int status;
601   double cpu_time;
602 
603   cpu_time = 0;
604 
605   status = GetProcessTimes(GetCurrentProcess(), &CreationTime, &ExitTime,
606                            &KernelTime, &UserTime);
607 
608   if (status) {
609     double sec = 0;
610 
611     sec += KernelTime.dwHighDateTime;
612     sec += UserTime.dwHighDateTime;
613 
614     /* Shift left by 32 bits */
615     sec *= (double)(1 << 16) * (double)(1 << 16);
616 
617     sec += KernelTime.dwLowDateTime;
618     sec += UserTime.dwLowDateTime;
619 
620     cpu_time += (sec * 100.0) / KMP_NSEC_PER_SEC;
621   }
622 
623   return cpu_time;
624 }
625 
626 int __kmp_read_system_info(struct kmp_sys_info *info) {
627   info->maxrss = 0; /* the maximum resident set size utilized (in kilobytes) */
628   info->minflt = 0; /* the number of page faults serviced without any I/O */
629   info->majflt = 0; /* the number of page faults serviced that required I/O */
630   info->nswap = 0; // the number of times a process was "swapped" out of memory
631   info->inblock = 0; // the number of times the file system had to perform input
632   info->oublock = 0; // number of times the file system had to perform output
633   info->nvcsw = 0; /* the number of times a context switch was voluntarily */
634   info->nivcsw = 0; /* the number of times a context switch was forced */
635 
636   return 1;
637 }
638 
639 void __kmp_runtime_initialize(void) {
640   SYSTEM_INFO info;
641   kmp_str_buf_t path;
642   UINT path_size;
643 
644   if (__kmp_init_runtime) {
645     return;
646   }
647 
648 #if KMP_DYNAMIC_LIB
649   /* Pin dynamic library for the lifetime of application */
650   {
651     // First, turn off error message boxes
652     UINT err_mode = SetErrorMode(SEM_FAILCRITICALERRORS);
653     HMODULE h;
654     BOOL ret = GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS |
655                                      GET_MODULE_HANDLE_EX_FLAG_PIN,
656                                  (LPCTSTR)&__kmp_serial_initialize, &h);
657     KMP_DEBUG_ASSERT2(h && ret, "OpenMP RTL cannot find itself loaded");
658     SetErrorMode(err_mode); // Restore error mode
659     KA_TRACE(10, ("__kmp_runtime_initialize: dynamic library pinned\n"));
660   }
661 #endif
662 
663   InitializeCriticalSection(&__kmp_win32_section);
664 #if USE_ITT_BUILD
665   __kmp_itt_system_object_created(&__kmp_win32_section, "Critical Section");
666 #endif /* USE_ITT_BUILD */
667   __kmp_initialize_system_tick();
668 
669 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
670   if (!__kmp_cpuinfo.initialized) {
671     __kmp_query_cpuid(&__kmp_cpuinfo);
672   }
673 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
674 
675 /* Set up minimum number of threads to switch to TLS gtid */
676 #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB
677   // Windows* OS, static library.
678   /* New thread may use stack space previously used by another thread,
679      currently terminated. On Windows* OS, in case of static linking, we do not
680      know the moment of thread termination, and our structures (__kmp_threads
681      and __kmp_root arrays) are still keep info about dead threads. This leads
682      to problem in __kmp_get_global_thread_id() function: it wrongly finds gtid
683      (by searching through stack addresses of all known threads) for
684      unregistered foreign tread.
685 
686      Setting __kmp_tls_gtid_min to 0 workarounds this problem:
687      __kmp_get_global_thread_id() does not search through stacks, but get gtid
688      from TLS immediately.
689       --ln
690   */
691   __kmp_tls_gtid_min = 0;
692 #else
693   __kmp_tls_gtid_min = KMP_TLS_GTID_MIN;
694 #endif
695 
696   /* for the static library */
697   if (!__kmp_gtid_threadprivate_key) {
698     __kmp_gtid_threadprivate_key = TlsAlloc();
699     if (__kmp_gtid_threadprivate_key == TLS_OUT_OF_INDEXES) {
700       KMP_FATAL(TLSOutOfIndexes);
701     }
702   }
703 
704   // Load ntdll.dll.
705   /* Simple GetModuleHandle( "ntdll.dl" ) is not suitable due to security issue
706      (see http://www.microsoft.com/technet/security/advisory/2269637.mspx). We
707      have to specify full path to the library. */
708   __kmp_str_buf_init(&path);
709   path_size = GetSystemDirectory(path.str, path.size);
710   KMP_DEBUG_ASSERT(path_size > 0);
711   if (path_size >= path.size) {
712     // Buffer is too short.  Expand the buffer and try again.
713     __kmp_str_buf_reserve(&path, path_size);
714     path_size = GetSystemDirectory(path.str, path.size);
715     KMP_DEBUG_ASSERT(path_size > 0);
716   }
717   if (path_size > 0 && path_size < path.size) {
718     // Now we have system directory name in the buffer.
719     // Append backslash and name of dll to form full path,
720     path.used = path_size;
721     __kmp_str_buf_print(&path, "\\%s", "ntdll.dll");
722 
723     // Now load ntdll using full path.
724     ntdll = GetModuleHandle(path.str);
725   }
726 
727   KMP_DEBUG_ASSERT(ntdll != NULL);
728   if (ntdll != NULL) {
729     NtQuerySystemInformation = (NtQuerySystemInformation_t)GetProcAddress(
730         ntdll, "NtQuerySystemInformation");
731   }
732   KMP_DEBUG_ASSERT(NtQuerySystemInformation != NULL);
733 
734 #if KMP_GROUP_AFFINITY
735   // Load kernel32.dll.
736   // Same caveat - must use full system path name.
737   if (path_size > 0 && path_size < path.size) {
738     // Truncate the buffer back to just the system path length,
739     // discarding "\\ntdll.dll", and replacing it with "kernel32.dll".
740     path.used = path_size;
741     __kmp_str_buf_print(&path, "\\%s", "kernel32.dll");
742 
743     // Load kernel32.dll using full path.
744     kernel32 = GetModuleHandle(path.str);
745     KA_TRACE(10, ("__kmp_runtime_initialize: kernel32.dll = %s\n", path.str));
746 
747     // Load the function pointers to kernel32.dll routines
748     // that may or may not exist on this system.
749     if (kernel32 != NULL) {
750       __kmp_GetActiveProcessorCount =
751           (kmp_GetActiveProcessorCount_t)GetProcAddress(
752               kernel32, "GetActiveProcessorCount");
753       __kmp_GetActiveProcessorGroupCount =
754           (kmp_GetActiveProcessorGroupCount_t)GetProcAddress(
755               kernel32, "GetActiveProcessorGroupCount");
756       __kmp_GetThreadGroupAffinity =
757           (kmp_GetThreadGroupAffinity_t)GetProcAddress(
758               kernel32, "GetThreadGroupAffinity");
759       __kmp_SetThreadGroupAffinity =
760           (kmp_SetThreadGroupAffinity_t)GetProcAddress(
761               kernel32, "SetThreadGroupAffinity");
762 
763       KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_GetActiveProcessorCount"
764                     " = %p\n",
765                     __kmp_GetActiveProcessorCount));
766       KA_TRACE(10, ("__kmp_runtime_initialize: "
767                     "__kmp_GetActiveProcessorGroupCount = %p\n",
768                     __kmp_GetActiveProcessorGroupCount));
769       KA_TRACE(10, ("__kmp_runtime_initialize:__kmp_GetThreadGroupAffinity"
770                     " = %p\n",
771                     __kmp_GetThreadGroupAffinity));
772       KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_SetThreadGroupAffinity"
773                     " = %p\n",
774                     __kmp_SetThreadGroupAffinity));
775       KA_TRACE(10, ("__kmp_runtime_initialize: sizeof(kmp_affin_mask_t) = %d\n",
776                     sizeof(kmp_affin_mask_t)));
777 
778       // See if group affinity is supported on this system.
779       // If so, calculate the #groups and #procs.
780       //
781       // Group affinity was introduced with Windows* 7 OS and
782       // Windows* Server 2008 R2 OS.
783       if ((__kmp_GetActiveProcessorCount != NULL) &&
784           (__kmp_GetActiveProcessorGroupCount != NULL) &&
785           (__kmp_GetThreadGroupAffinity != NULL) &&
786           (__kmp_SetThreadGroupAffinity != NULL) &&
787           ((__kmp_num_proc_groups = __kmp_GetActiveProcessorGroupCount()) >
788            1)) {
789         // Calculate the total number of active OS procs.
790         int i;
791 
792         KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups"
793                       " detected\n",
794                       __kmp_num_proc_groups));
795 
796         __kmp_xproc = 0;
797 
798         for (i = 0; i < __kmp_num_proc_groups; i++) {
799           DWORD size = __kmp_GetActiveProcessorCount(i);
800           __kmp_xproc += size;
801           KA_TRACE(10, ("__kmp_runtime_initialize: proc group %d size = %d\n",
802                         i, size));
803         }
804       } else {
805         KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups"
806                       " detected\n",
807                       __kmp_num_proc_groups));
808       }
809     }
810   }
811   if (__kmp_num_proc_groups <= 1) {
812     GetSystemInfo(&info);
813     __kmp_xproc = info.dwNumberOfProcessors;
814   }
815 #else
816   GetSystemInfo(&info);
817   __kmp_xproc = info.dwNumberOfProcessors;
818 #endif /* KMP_GROUP_AFFINITY */
819 
820   // If the OS said there were 0 procs, take a guess and use a value of 2.
821   // This is done for Linux* OS, also.  Do we need error / warning?
822   if (__kmp_xproc <= 0) {
823     __kmp_xproc = 2;
824   }
825 
826   KA_TRACE(5,
827            ("__kmp_runtime_initialize: total processors = %d\n", __kmp_xproc));
828 
829   __kmp_str_buf_free(&path);
830 
831 #if USE_ITT_BUILD
832   __kmp_itt_initialize();
833 #endif /* USE_ITT_BUILD */
834 
835   __kmp_init_runtime = TRUE;
836 } // __kmp_runtime_initialize
837 
838 void __kmp_runtime_destroy(void) {
839   if (!__kmp_init_runtime) {
840     return;
841   }
842 
843 #if USE_ITT_BUILD
844   __kmp_itt_destroy();
845 #endif /* USE_ITT_BUILD */
846 
847   /* we can't DeleteCriticalsection( & __kmp_win32_section ); */
848   /* due to the KX_TRACE() commands */
849   KA_TRACE(40, ("__kmp_runtime_destroy\n"));
850 
851   if (__kmp_gtid_threadprivate_key) {
852     TlsFree(__kmp_gtid_threadprivate_key);
853     __kmp_gtid_threadprivate_key = 0;
854   }
855 
856   __kmp_affinity_uninitialize();
857   DeleteCriticalSection(&__kmp_win32_section);
858 
859   ntdll = NULL;
860   NtQuerySystemInformation = NULL;
861 
862 #if KMP_ARCH_X86_64
863   kernel32 = NULL;
864   __kmp_GetActiveProcessorCount = NULL;
865   __kmp_GetActiveProcessorGroupCount = NULL;
866   __kmp_GetThreadGroupAffinity = NULL;
867   __kmp_SetThreadGroupAffinity = NULL;
868 #endif // KMP_ARCH_X86_64
869 
870   __kmp_init_runtime = FALSE;
871 }
872 
873 void __kmp_terminate_thread(int gtid) {
874   kmp_info_t *th = __kmp_threads[gtid];
875 
876   if (!th)
877     return;
878 
879   KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid));
880 
881   if (TerminateThread(th->th.th_info.ds.ds_thread, (DWORD)-1) == FALSE) {
882     /* It's OK, the thread may have exited already */
883   }
884   __kmp_free_handle(th->th.th_info.ds.ds_thread);
885 }
886 
887 void __kmp_clear_system_time(void) {
888   BOOL status;
889   LARGE_INTEGER time;
890   status = QueryPerformanceCounter(&time);
891   __kmp_win32_time = (kmp_int64)time.QuadPart;
892 }
893 
894 void __kmp_initialize_system_tick(void) {
895   {
896     BOOL status;
897     LARGE_INTEGER freq;
898 
899     status = QueryPerformanceFrequency(&freq);
900     if (!status) {
901       DWORD error = GetLastError();
902       __kmp_fatal(KMP_MSG(FunctionError, "QueryPerformanceFrequency()"),
903                   KMP_ERR(error), __kmp_msg_null);
904 
905     } else {
906       __kmp_win32_tick = ((double)1.0) / (double)freq.QuadPart;
907     }
908   }
909 }
910 
911 /* Calculate the elapsed wall clock time for the user */
912 
913 void __kmp_elapsed(double *t) {
914   BOOL status;
915   LARGE_INTEGER now;
916   status = QueryPerformanceCounter(&now);
917   *t = ((double)now.QuadPart) * __kmp_win32_tick;
918 }
919 
920 /* Calculate the elapsed wall clock tick for the user */
921 
922 void __kmp_elapsed_tick(double *t) { *t = __kmp_win32_tick; }
923 
924 void __kmp_read_system_time(double *delta) {
925   if (delta != NULL) {
926     BOOL status;
927     LARGE_INTEGER now;
928 
929     status = QueryPerformanceCounter(&now);
930 
931     *delta = ((double)(((kmp_int64)now.QuadPart) - __kmp_win32_time)) *
932              __kmp_win32_tick;
933   }
934 }
935 
936 /* Return the current time stamp in nsec */
937 kmp_uint64 __kmp_now_nsec() {
938   LARGE_INTEGER now;
939   QueryPerformanceCounter(&now);
940   return 1e9 * __kmp_win32_tick * now.QuadPart;
941 }
942 
943 extern "C"
944 void *__stdcall __kmp_launch_worker(void *arg) {
945   volatile void *stack_data;
946   void *exit_val;
947   void *padding = 0;
948   kmp_info_t *this_thr = (kmp_info_t *)arg;
949   int gtid;
950 
951   gtid = this_thr->th.th_info.ds.ds_gtid;
952   __kmp_gtid_set_specific(gtid);
953 #ifdef KMP_TDATA_GTID
954 #error "This define causes problems with LoadLibrary() + declspec(thread) " \
955         "on Windows* OS.  See CQ50564, tests kmp_load_library*.c and this MSDN " \
956         "reference: http://support.microsoft.com/kb/118816"
957 //__kmp_gtid = gtid;
958 #endif
959 
960 #if USE_ITT_BUILD
961   __kmp_itt_thread_name(gtid);
962 #endif /* USE_ITT_BUILD */
963 
964   __kmp_affinity_set_init_mask(gtid, FALSE);
965 
966 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
967   // Set FP control regs to be a copy of the parallel initialization thread's.
968   __kmp_clear_x87_fpu_status_word();
969   __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word);
970   __kmp_load_mxcsr(&__kmp_init_mxcsr);
971 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
972 
973   if (__kmp_stkoffset > 0 && gtid > 0) {
974     padding = KMP_ALLOCA(gtid * __kmp_stkoffset);
975   }
976 
977   KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive);
978   this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
979   TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE);
980 
981   if (TCR_4(__kmp_gtid_mode) <
982       2) { // check stack only if it is used to get gtid
983     TCW_PTR(this_thr->th.th_info.ds.ds_stackbase, &stack_data);
984     KMP_ASSERT(this_thr->th.th_info.ds.ds_stackgrow == FALSE);
985     __kmp_check_stack_overlap(this_thr);
986   }
987   KMP_MB();
988   exit_val = __kmp_launch_thread(this_thr);
989   KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive);
990   TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE);
991   KMP_MB();
992   return exit_val;
993 }
994 
995 #if KMP_USE_MONITOR
996 /* The monitor thread controls all of the threads in the complex */
997 
998 void *__stdcall __kmp_launch_monitor(void *arg) {
999   DWORD wait_status;
1000   kmp_thread_t monitor;
1001   int status;
1002   int interval;
1003   kmp_info_t *this_thr = (kmp_info_t *)arg;
1004 
1005   KMP_DEBUG_ASSERT(__kmp_init_monitor);
1006   TCW_4(__kmp_init_monitor, 2); // AC: Signal library that monitor has started
1007   // TODO: hide "2" in enum (like {true,false,started})
1008   this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
1009   TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE);
1010 
1011   KMP_MB(); /* Flush all pending memory write invalidates.  */
1012   KA_TRACE(10, ("__kmp_launch_monitor: launched\n"));
1013 
1014   monitor = GetCurrentThread();
1015 
1016   /* set thread priority */
1017   status = SetThreadPriority(monitor, THREAD_PRIORITY_HIGHEST);
1018   if (!status) {
1019     DWORD error = GetLastError();
1020     __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null);
1021   }
1022 
1023   /* register us as monitor */
1024   __kmp_gtid_set_specific(KMP_GTID_MONITOR);
1025 #ifdef KMP_TDATA_GTID
1026 #error "This define causes problems with LoadLibrary() + declspec(thread) " \
1027         "on Windows* OS.  See CQ50564, tests kmp_load_library*.c and this MSDN " \
1028         "reference: http://support.microsoft.com/kb/118816"
1029 //__kmp_gtid = KMP_GTID_MONITOR;
1030 #endif
1031 
1032 #if USE_ITT_BUILD
1033   __kmp_itt_thread_ignore(); // Instruct Intel(R) Threading Tools to ignore
1034 // monitor thread.
1035 #endif /* USE_ITT_BUILD */
1036 
1037   KMP_MB(); /* Flush all pending memory write invalidates.  */
1038 
1039   interval = (1000 / __kmp_monitor_wakeups); /* in milliseconds */
1040 
1041   while (!TCR_4(__kmp_global.g.g_done)) {
1042     /*  This thread monitors the state of the system */
1043 
1044     KA_TRACE(15, ("__kmp_launch_monitor: update\n"));
1045 
1046     wait_status = WaitForSingleObject(__kmp_monitor_ev, interval);
1047 
1048     if (wait_status == WAIT_TIMEOUT) {
1049       TCW_4(__kmp_global.g.g_time.dt.t_value,
1050             TCR_4(__kmp_global.g.g_time.dt.t_value) + 1);
1051     }
1052 
1053     KMP_MB(); /* Flush all pending memory write invalidates.  */
1054   }
1055 
1056   KA_TRACE(10, ("__kmp_launch_monitor: finished\n"));
1057 
1058   status = SetThreadPriority(monitor, THREAD_PRIORITY_NORMAL);
1059   if (!status) {
1060     DWORD error = GetLastError();
1061     __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null);
1062   }
1063 
1064   if (__kmp_global.g.g_abort != 0) {
1065     /* now we need to terminate the worker threads   */
1066     /* the value of t_abort is the signal we caught */
1067     int gtid;
1068 
1069     KA_TRACE(10, ("__kmp_launch_monitor: terminate sig=%d\n",
1070                   (__kmp_global.g.g_abort)));
1071 
1072     /* terminate the OpenMP worker threads */
1073     /* TODO this is not valid for sibling threads!!
1074      * the uber master might not be 0 anymore.. */
1075     for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid)
1076       __kmp_terminate_thread(gtid);
1077 
1078     __kmp_cleanup();
1079 
1080     Sleep(0);
1081 
1082     KA_TRACE(10,
1083              ("__kmp_launch_monitor: raise sig=%d\n", __kmp_global.g.g_abort));
1084 
1085     if (__kmp_global.g.g_abort > 0) {
1086       raise(__kmp_global.g.g_abort);
1087     }
1088   }
1089 
1090   TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE);
1091 
1092   KMP_MB();
1093   return arg;
1094 }
1095 #endif
1096 
1097 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) {
1098   kmp_thread_t handle;
1099   DWORD idThread;
1100 
1101   KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid));
1102 
1103   th->th.th_info.ds.ds_gtid = gtid;
1104 
1105   if (KMP_UBER_GTID(gtid)) {
1106     int stack_data;
1107 
1108     /* TODO: GetCurrentThread() returns a pseudo-handle that is unsuitable for
1109        other threads to use. Is it appropriate to just use GetCurrentThread?
1110        When should we close this handle?  When unregistering the root? */
1111     {
1112       BOOL rc;
1113       rc = DuplicateHandle(GetCurrentProcess(), GetCurrentThread(),
1114                            GetCurrentProcess(), &th->th.th_info.ds.ds_thread, 0,
1115                            FALSE, DUPLICATE_SAME_ACCESS);
1116       KMP_ASSERT(rc);
1117       KA_TRACE(10, (" __kmp_create_worker: ROOT Handle duplicated, th = %p, "
1118                     "handle = %" KMP_UINTPTR_SPEC "\n",
1119                     (LPVOID)th, th->th.th_info.ds.ds_thread));
1120       th->th.th_info.ds.ds_thread_id = GetCurrentThreadId();
1121     }
1122     if (TCR_4(__kmp_gtid_mode) < 2) { // check stack only if used to get gtid
1123       /* we will dynamically update the stack range if gtid_mode == 1 */
1124       TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data);
1125       TCW_PTR(th->th.th_info.ds.ds_stacksize, 0);
1126       TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE);
1127       __kmp_check_stack_overlap(th);
1128     }
1129   } else {
1130     KMP_MB(); /* Flush all pending memory write invalidates.  */
1131 
1132     /* Set stack size for this thread now. */
1133     KA_TRACE(10,
1134              ("__kmp_create_worker: stack_size = %" KMP_SIZE_T_SPEC " bytes\n",
1135               stack_size));
1136 
1137     stack_size += gtid * __kmp_stkoffset;
1138 
1139     TCW_PTR(th->th.th_info.ds.ds_stacksize, stack_size);
1140     TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE);
1141 
1142     KA_TRACE(10,
1143              ("__kmp_create_worker: (before) stack_size = %" KMP_SIZE_T_SPEC
1144               " bytes, &__kmp_launch_worker = %p, th = %p, &idThread = %p\n",
1145               (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker,
1146               (LPVOID)th, &idThread));
1147 
1148     handle = CreateThread(
1149         NULL, (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)__kmp_launch_worker,
1150         (LPVOID)th, STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread);
1151 
1152     KA_TRACE(10,
1153              ("__kmp_create_worker: (after) stack_size = %" KMP_SIZE_T_SPEC
1154               " bytes, &__kmp_launch_worker = %p, th = %p, "
1155               "idThread = %u, handle = %" KMP_UINTPTR_SPEC "\n",
1156               (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker,
1157               (LPVOID)th, idThread, handle));
1158 
1159     if (handle == 0) {
1160       DWORD error = GetLastError();
1161       __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null);
1162     } else {
1163       th->th.th_info.ds.ds_thread = handle;
1164     }
1165 
1166     KMP_MB(); /* Flush all pending memory write invalidates.  */
1167   }
1168 
1169   KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid));
1170 }
1171 
1172 int __kmp_still_running(kmp_info_t *th) {
1173   return (WAIT_TIMEOUT == WaitForSingleObject(th->th.th_info.ds.ds_thread, 0));
1174 }
1175 
1176 #if KMP_USE_MONITOR
1177 void __kmp_create_monitor(kmp_info_t *th) {
1178   kmp_thread_t handle;
1179   DWORD idThread;
1180   int ideal, new_ideal;
1181 
1182   if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) {
1183     // We don't need monitor thread in case of MAX_BLOCKTIME
1184     KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of "
1185                   "MAX blocktime\n"));
1186     th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op
1187     th->th.th_info.ds.ds_gtid = 0;
1188     TCW_4(__kmp_init_monitor, 2); // Signal to stop waiting for monitor creation
1189     return;
1190   }
1191   KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n"));
1192 
1193   KMP_MB(); /* Flush all pending memory write invalidates.  */
1194 
1195   __kmp_monitor_ev = CreateEvent(NULL, TRUE, FALSE, NULL);
1196   if (__kmp_monitor_ev == NULL) {
1197     DWORD error = GetLastError();
1198     __kmp_fatal(KMP_MSG(CantCreateEvent), KMP_ERR(error), __kmp_msg_null);
1199   }
1200 #if USE_ITT_BUILD
1201   __kmp_itt_system_object_created(__kmp_monitor_ev, "Event");
1202 #endif /* USE_ITT_BUILD */
1203 
1204   th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR;
1205   th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR;
1206 
1207   // FIXME - on Windows* OS, if __kmp_monitor_stksize = 0, figure out how
1208   // to automatically expand stacksize based on CreateThread error code.
1209   if (__kmp_monitor_stksize == 0) {
1210     __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE;
1211   }
1212   if (__kmp_monitor_stksize < __kmp_sys_min_stksize) {
1213     __kmp_monitor_stksize = __kmp_sys_min_stksize;
1214   }
1215 
1216   KA_TRACE(10, ("__kmp_create_monitor: requested stacksize = %d bytes\n",
1217                 (int)__kmp_monitor_stksize));
1218 
1219   TCW_4(__kmp_global.g.g_time.dt.t_value, 0);
1220 
1221   handle =
1222       CreateThread(NULL, (SIZE_T)__kmp_monitor_stksize,
1223                    (LPTHREAD_START_ROUTINE)__kmp_launch_monitor, (LPVOID)th,
1224                    STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread);
1225   if (handle == 0) {
1226     DWORD error = GetLastError();
1227     __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null);
1228   } else
1229     th->th.th_info.ds.ds_thread = handle;
1230 
1231   KMP_MB(); /* Flush all pending memory write invalidates.  */
1232 
1233   KA_TRACE(10, ("__kmp_create_monitor: monitor created %p\n",
1234                 (void *)th->th.th_info.ds.ds_thread));
1235 }
1236 #endif
1237 
1238 /* Check to see if thread is still alive.
1239    NOTE:  The ExitProcess(code) system call causes all threads to Terminate
1240    with a exit_val = code.  Because of this we can not rely on exit_val having
1241    any particular value.  So this routine may return STILL_ALIVE in exit_val
1242    even after the thread is dead. */
1243 
1244 int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val) {
1245   DWORD rc;
1246   rc = GetExitCodeThread(th->th.th_info.ds.ds_thread, exit_val);
1247   if (rc == 0) {
1248     DWORD error = GetLastError();
1249     __kmp_fatal(KMP_MSG(FunctionError, "GetExitCodeThread()"), KMP_ERR(error),
1250                 __kmp_msg_null);
1251   }
1252   return (*exit_val == STILL_ACTIVE);
1253 }
1254 
1255 void __kmp_exit_thread(int exit_status) {
1256   ExitThread(exit_status);
1257 } // __kmp_exit_thread
1258 
1259 // This is a common part for both __kmp_reap_worker() and __kmp_reap_monitor().
1260 static void __kmp_reap_common(kmp_info_t *th) {
1261   DWORD exit_val;
1262 
1263   KMP_MB(); /* Flush all pending memory write invalidates.  */
1264 
1265   KA_TRACE(
1266       10, ("__kmp_reap_common: try to reap (%d)\n", th->th.th_info.ds.ds_gtid));
1267 
1268   /* 2006-10-19:
1269      There are two opposite situations:
1270      1. Windows* OS keep thread alive after it resets ds_alive flag and
1271      exits from thread function. (For example, see C70770/Q394281 "unloading of
1272      dll based on OMP is very slow".)
1273      2. Windows* OS may kill thread before it resets ds_alive flag.
1274 
1275      Right solution seems to be waiting for *either* thread termination *or*
1276      ds_alive resetting. */
1277   {
1278     // TODO: This code is very similar to KMP_WAIT. Need to generalize
1279     // KMP_WAIT to cover this usage also.
1280     void *obj = NULL;
1281     kmp_uint32 spins;
1282 #if USE_ITT_BUILD
1283     KMP_FSYNC_SPIN_INIT(obj, (void *)&th->th.th_info.ds.ds_alive);
1284 #endif /* USE_ITT_BUILD */
1285     KMP_INIT_YIELD(spins);
1286     do {
1287 #if USE_ITT_BUILD
1288       KMP_FSYNC_SPIN_PREPARE(obj);
1289 #endif /* USE_ITT_BUILD */
1290       __kmp_is_thread_alive(th, &exit_val);
1291       KMP_YIELD_OVERSUB_ELSE_SPIN(spins);
1292     } while (exit_val == STILL_ACTIVE && TCR_4(th->th.th_info.ds.ds_alive));
1293 #if USE_ITT_BUILD
1294     if (exit_val == STILL_ACTIVE) {
1295       KMP_FSYNC_CANCEL(obj);
1296     } else {
1297       KMP_FSYNC_SPIN_ACQUIRED(obj);
1298     }
1299 #endif /* USE_ITT_BUILD */
1300   }
1301 
1302   __kmp_free_handle(th->th.th_info.ds.ds_thread);
1303 
1304   /* NOTE:  The ExitProcess(code) system call causes all threads to Terminate
1305      with a exit_val = code.  Because of this we can not rely on exit_val having
1306      any particular value. */
1307   if (exit_val == STILL_ACTIVE) {
1308     KA_TRACE(1, ("__kmp_reap_common: thread still active.\n"));
1309   } else if ((void *)exit_val != (void *)th) {
1310     KA_TRACE(1, ("__kmp_reap_common: ExitProcess / TerminateThread used?\n"));
1311   }
1312 
1313   KA_TRACE(10,
1314            ("__kmp_reap_common: done reaping (%d), handle = %" KMP_UINTPTR_SPEC
1315             "\n",
1316             th->th.th_info.ds.ds_gtid, th->th.th_info.ds.ds_thread));
1317 
1318   th->th.th_info.ds.ds_thread = 0;
1319   th->th.th_info.ds.ds_tid = KMP_GTID_DNE;
1320   th->th.th_info.ds.ds_gtid = KMP_GTID_DNE;
1321   th->th.th_info.ds.ds_thread_id = 0;
1322 
1323   KMP_MB(); /* Flush all pending memory write invalidates.  */
1324 }
1325 
1326 #if KMP_USE_MONITOR
1327 void __kmp_reap_monitor(kmp_info_t *th) {
1328   int status;
1329 
1330   KA_TRACE(10, ("__kmp_reap_monitor: try to reap %p\n",
1331                 (void *)th->th.th_info.ds.ds_thread));
1332 
1333   // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR.
1334   // If both tid and gtid are 0, it means the monitor did not ever start.
1335   // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down.
1336   KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid);
1337   if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) {
1338     KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n"));
1339     return;
1340   }
1341 
1342   KMP_MB(); /* Flush all pending memory write invalidates.  */
1343 
1344   status = SetEvent(__kmp_monitor_ev);
1345   if (status == FALSE) {
1346     DWORD error = GetLastError();
1347     __kmp_fatal(KMP_MSG(CantSetEvent), KMP_ERR(error), __kmp_msg_null);
1348   }
1349   KA_TRACE(10, ("__kmp_reap_monitor: reaping thread (%d)\n",
1350                 th->th.th_info.ds.ds_gtid));
1351   __kmp_reap_common(th);
1352 
1353   __kmp_free_handle(__kmp_monitor_ev);
1354 
1355   KMP_MB(); /* Flush all pending memory write invalidates.  */
1356 }
1357 #endif
1358 
1359 void __kmp_reap_worker(kmp_info_t *th) {
1360   KA_TRACE(10, ("__kmp_reap_worker: reaping thread (%d)\n",
1361                 th->th.th_info.ds.ds_gtid));
1362   __kmp_reap_common(th);
1363 }
1364 
1365 #if KMP_HANDLE_SIGNALS
1366 
1367 static void __kmp_team_handler(int signo) {
1368   if (__kmp_global.g.g_abort == 0) {
1369     // Stage 1 signal handler, let's shut down all of the threads.
1370     if (__kmp_debug_buf) {
1371       __kmp_dump_debug_buffer();
1372     }
1373     KMP_MB(); // Flush all pending memory write invalidates.
1374     TCW_4(__kmp_global.g.g_abort, signo);
1375     KMP_MB(); // Flush all pending memory write invalidates.
1376     TCW_4(__kmp_global.g.g_done, TRUE);
1377     KMP_MB(); // Flush all pending memory write invalidates.
1378   }
1379 } // __kmp_team_handler
1380 
1381 static sig_func_t __kmp_signal(int signum, sig_func_t handler) {
1382   sig_func_t old = signal(signum, handler);
1383   if (old == SIG_ERR) {
1384     int error = errno;
1385     __kmp_fatal(KMP_MSG(FunctionError, "signal"), KMP_ERR(error),
1386                 __kmp_msg_null);
1387   }
1388   return old;
1389 }
1390 
1391 static void __kmp_install_one_handler(int sig, sig_func_t handler,
1392                                       int parallel_init) {
1393   sig_func_t old;
1394   KMP_MB(); /* Flush all pending memory write invalidates.  */
1395   KB_TRACE(60, ("__kmp_install_one_handler: called: sig=%d\n", sig));
1396   if (parallel_init) {
1397     old = __kmp_signal(sig, handler);
1398     // SIG_DFL on Windows* OS in NULL or 0.
1399     if (old == __kmp_sighldrs[sig]) {
1400       __kmp_siginstalled[sig] = 1;
1401     } else { // Restore/keep user's handler if one previously installed.
1402       old = __kmp_signal(sig, old);
1403     }
1404   } else {
1405     // Save initial/system signal handlers to see if user handlers installed.
1406     // 2009-09-23: It is a dead code. On Windows* OS __kmp_install_signals
1407     // called once with parallel_init == TRUE.
1408     old = __kmp_signal(sig, SIG_DFL);
1409     __kmp_sighldrs[sig] = old;
1410     __kmp_signal(sig, old);
1411   }
1412   KMP_MB(); /* Flush all pending memory write invalidates.  */
1413 } // __kmp_install_one_handler
1414 
1415 static void __kmp_remove_one_handler(int sig) {
1416   if (__kmp_siginstalled[sig]) {
1417     sig_func_t old;
1418     KMP_MB(); // Flush all pending memory write invalidates.
1419     KB_TRACE(60, ("__kmp_remove_one_handler: called: sig=%d\n", sig));
1420     old = __kmp_signal(sig, __kmp_sighldrs[sig]);
1421     if (old != __kmp_team_handler) {
1422       KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, "
1423                     "restoring: sig=%d\n",
1424                     sig));
1425       old = __kmp_signal(sig, old);
1426     }
1427     __kmp_sighldrs[sig] = NULL;
1428     __kmp_siginstalled[sig] = 0;
1429     KMP_MB(); // Flush all pending memory write invalidates.
1430   }
1431 } // __kmp_remove_one_handler
1432 
1433 void __kmp_install_signals(int parallel_init) {
1434   KB_TRACE(10, ("__kmp_install_signals: called\n"));
1435   if (!__kmp_handle_signals) {
1436     KB_TRACE(10, ("__kmp_install_signals: KMP_HANDLE_SIGNALS is false - "
1437                   "handlers not installed\n"));
1438     return;
1439   }
1440   __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init);
1441   __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init);
1442   __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init);
1443   __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init);
1444   __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init);
1445   __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init);
1446 } // __kmp_install_signals
1447 
1448 void __kmp_remove_signals(void) {
1449   int sig;
1450   KB_TRACE(10, ("__kmp_remove_signals: called\n"));
1451   for (sig = 1; sig < NSIG; ++sig) {
1452     __kmp_remove_one_handler(sig);
1453   }
1454 } // __kmp_remove_signals
1455 
1456 #endif // KMP_HANDLE_SIGNALS
1457 
1458 /* Put the thread to sleep for a time period */
1459 void __kmp_thread_sleep(int millis) {
1460   DWORD status;
1461 
1462   status = SleepEx((DWORD)millis, FALSE);
1463   if (status) {
1464     DWORD error = GetLastError();
1465     __kmp_fatal(KMP_MSG(FunctionError, "SleepEx()"), KMP_ERR(error),
1466                 __kmp_msg_null);
1467   }
1468 }
1469 
1470 // Determine whether the given address is mapped into the current address space.
1471 int __kmp_is_address_mapped(void *addr) {
1472   DWORD status;
1473   MEMORY_BASIC_INFORMATION lpBuffer;
1474   SIZE_T dwLength;
1475 
1476   dwLength = sizeof(MEMORY_BASIC_INFORMATION);
1477 
1478   status = VirtualQuery(addr, &lpBuffer, dwLength);
1479 
1480   return !(((lpBuffer.State == MEM_RESERVE) || (lpBuffer.State == MEM_FREE)) ||
1481            ((lpBuffer.Protect == PAGE_NOACCESS) ||
1482             (lpBuffer.Protect == PAGE_EXECUTE)));
1483 }
1484 
1485 kmp_uint64 __kmp_hardware_timestamp(void) {
1486   kmp_uint64 r = 0;
1487 
1488   QueryPerformanceCounter((LARGE_INTEGER *)&r);
1489   return r;
1490 }
1491 
1492 /* Free handle and check the error code */
1493 void __kmp_free_handle(kmp_thread_t tHandle) {
1494   /* called with parameter type HANDLE also, thus suppose kmp_thread_t defined
1495    * as HANDLE */
1496   BOOL rc;
1497   rc = CloseHandle(tHandle);
1498   if (!rc) {
1499     DWORD error = GetLastError();
1500     __kmp_fatal(KMP_MSG(CantCloseHandle), KMP_ERR(error), __kmp_msg_null);
1501   }
1502 }
1503 
1504 int __kmp_get_load_balance(int max) {
1505   static ULONG glb_buff_size = 100 * 1024;
1506 
1507   // Saved count of the running threads for the thread balance algortihm
1508   static int glb_running_threads = 0;
1509   static double glb_call_time = 0; /* Thread balance algorithm call time */
1510 
1511   int running_threads = 0; // Number of running threads in the system.
1512   NTSTATUS status = 0;
1513   ULONG buff_size = 0;
1514   ULONG info_size = 0;
1515   void *buffer = NULL;
1516   PSYSTEM_PROCESS_INFORMATION spi = NULL;
1517   int first_time = 1;
1518 
1519   double call_time = 0.0; // start, finish;
1520 
1521   __kmp_elapsed(&call_time);
1522 
1523   if (glb_call_time &&
1524       (call_time - glb_call_time < __kmp_load_balance_interval)) {
1525     running_threads = glb_running_threads;
1526     goto finish;
1527   }
1528   glb_call_time = call_time;
1529 
1530   // Do not spend time on running algorithm if we have a permanent error.
1531   if (NtQuerySystemInformation == NULL) {
1532     running_threads = -1;
1533     goto finish;
1534   }
1535 
1536   if (max <= 0) {
1537     max = INT_MAX;
1538   }
1539 
1540   do {
1541 
1542     if (first_time) {
1543       buff_size = glb_buff_size;
1544     } else {
1545       buff_size = 2 * buff_size;
1546     }
1547 
1548     buffer = KMP_INTERNAL_REALLOC(buffer, buff_size);
1549     if (buffer == NULL) {
1550       running_threads = -1;
1551       goto finish;
1552     }
1553     status = NtQuerySystemInformation(SystemProcessInformation, buffer,
1554                                       buff_size, &info_size);
1555     first_time = 0;
1556 
1557   } while (status == STATUS_INFO_LENGTH_MISMATCH);
1558   glb_buff_size = buff_size;
1559 
1560 #define CHECK(cond)                                                            \
1561   {                                                                            \
1562     KMP_DEBUG_ASSERT(cond);                                                    \
1563     if (!(cond)) {                                                             \
1564       running_threads = -1;                                                    \
1565       goto finish;                                                             \
1566     }                                                                          \
1567   }
1568 
1569   CHECK(buff_size >= info_size);
1570   spi = PSYSTEM_PROCESS_INFORMATION(buffer);
1571   for (;;) {
1572     ptrdiff_t offset = uintptr_t(spi) - uintptr_t(buffer);
1573     CHECK(0 <= offset &&
1574           offset + sizeof(SYSTEM_PROCESS_INFORMATION) < info_size);
1575     HANDLE pid = spi->ProcessId;
1576     ULONG num = spi->NumberOfThreads;
1577     CHECK(num >= 1);
1578     size_t spi_size =
1579         sizeof(SYSTEM_PROCESS_INFORMATION) + sizeof(SYSTEM_THREAD) * (num - 1);
1580     CHECK(offset + spi_size <
1581           info_size); // Make sure process info record fits the buffer.
1582     if (spi->NextEntryOffset != 0) {
1583       CHECK(spi_size <=
1584             spi->NextEntryOffset); // And do not overlap with the next record.
1585     }
1586     // pid == 0 corresponds to the System Idle Process. It always has running
1587     // threads on all cores. So, we don't consider the running threads of this
1588     // process.
1589     if (pid != 0) {
1590       for (int i = 0; i < num; ++i) {
1591         THREAD_STATE state = spi->Threads[i].State;
1592         // Count threads that have Ready or Running state.
1593         // !!! TODO: Why comment does not match the code???
1594         if (state == StateRunning) {
1595           ++running_threads;
1596           // Stop counting running threads if the number is already greater than
1597           // the number of available cores
1598           if (running_threads >= max) {
1599             goto finish;
1600           }
1601         }
1602       }
1603     }
1604     if (spi->NextEntryOffset == 0) {
1605       break;
1606     }
1607     spi = PSYSTEM_PROCESS_INFORMATION(uintptr_t(spi) + spi->NextEntryOffset);
1608   }
1609 
1610 #undef CHECK
1611 
1612 finish: // Clean up and exit.
1613 
1614   if (buffer != NULL) {
1615     KMP_INTERNAL_FREE(buffer);
1616   }
1617 
1618   glb_running_threads = running_threads;
1619 
1620   return running_threads;
1621 } //__kmp_get_load_balance()
1622