| 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 | #undef WIN32_NO_STATUS |
| 26 | #include <ntstatus.h> |
| 27 | #include <psapi.h> |
| 28 | #ifdef _MSC_VER |
| 29 | #pragma comment(lib, "psapi.lib") |
| 30 | #endif |
| 31 | |
| 32 | enum SYSTEM_INFORMATION_CLASS { |
| 33 | SystemProcessInformation = 5 |
| 34 | }; // SYSTEM_INFORMATION_CLASS |
| 35 | |
| 36 | struct CLIENT_ID { |
| 37 | HANDLE UniqueProcess; |
| 38 | HANDLE UniqueThread; |
| 39 | }; // struct CLIENT_ID |
| 40 | |
| 41 | enum THREAD_STATE { |
| 42 | StateInitialized, |
| 43 | StateReady, |
| 44 | StateRunning, |
| 45 | StateStandby, |
| 46 | StateTerminated, |
| 47 | StateWait, |
| 48 | StateTransition, |
| 49 | StateUnknown |
| 50 | }; // enum THREAD_STATE |
| 51 | |
| 52 | struct VM_COUNTERS { |
| 53 | SIZE_T PeakVirtualSize; |
| 54 | SIZE_T VirtualSize; |
| 55 | ULONG PageFaultCount; |
| 56 | SIZE_T PeakWorkingSetSize; |
| 57 | SIZE_T WorkingSetSize; |
| 58 | SIZE_T QuotaPeakPagedPoolUsage; |
| 59 | SIZE_T QuotaPagedPoolUsage; |
| 60 | SIZE_T QuotaPeakNonPagedPoolUsage; |
| 61 | SIZE_T QuotaNonPagedPoolUsage; |
| 62 | SIZE_T PagefileUsage; |
| 63 | SIZE_T PeakPagefileUsage; |
| 64 | SIZE_T PrivatePageCount; |
| 65 | }; // struct VM_COUNTERS |
| 66 | |
| 67 | struct SYSTEM_THREAD { |
| 68 | LARGE_INTEGER KernelTime; |
| 69 | LARGE_INTEGER UserTime; |
| 70 | LARGE_INTEGER CreateTime; |
| 71 | ULONG WaitTime; |
| 72 | LPVOID StartAddress; |
| 73 | CLIENT_ID ClientId; |
| 74 | DWORD Priority; |
| 75 | LONG BasePriority; |
| 76 | ULONG ContextSwitchCount; |
| 77 | THREAD_STATE State; |
| 78 | ULONG WaitReason; |
| 79 | }; // SYSTEM_THREAD |
| 80 | |
| 81 | KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, KernelTime) == 0); |
| 82 | #if KMP_ARCH_X86 || KMP_ARCH_ARM |
| 83 | KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, StartAddress) == 28); |
| 84 | KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, State) == 52); |
| 85 | #else |
| 86 | KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, StartAddress) == 32); |
| 87 | KMP_BUILD_ASSERT(offsetof(SYSTEM_THREAD, State) == 68); |
| 88 | #endif |
| 89 | |
| 90 | struct SYSTEM_PROCESS_INFORMATION { |
| 91 | ULONG NextEntryOffset; |
| 92 | ULONG NumberOfThreads; |
| 93 | LARGE_INTEGER Reserved[3]; |
| 94 | LARGE_INTEGER CreateTime; |
| 95 | LARGE_INTEGER UserTime; |
| 96 | LARGE_INTEGER KernelTime; |
| 97 | UNICODE_STRING ImageName; |
| 98 | DWORD BasePriority; |
| 99 | HANDLE ProcessId; |
| 100 | HANDLE ParentProcessId; |
| 101 | ULONG HandleCount; |
| 102 | ULONG Reserved2[2]; |
| 103 | VM_COUNTERS VMCounters; |
| 104 | IO_COUNTERS IOCounters; |
| 105 | SYSTEM_THREAD Threads[1]; |
| 106 | }; // SYSTEM_PROCESS_INFORMATION |
| 107 | typedef SYSTEM_PROCESS_INFORMATION *PSYSTEM_PROCESS_INFORMATION; |
| 108 | |
| 109 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, NextEntryOffset) == 0); |
| 110 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, CreateTime) == 32); |
| 111 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, ImageName) == 56); |
| 112 | #if KMP_ARCH_X86 || KMP_ARCH_ARM |
| 113 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, ProcessId) == 68); |
| 114 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, HandleCount) == 76); |
| 115 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, VMCounters) == 88); |
| 116 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, IOCounters) == 136); |
| 117 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, Threads) == 184); |
| 118 | #else |
| 119 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, ProcessId) == 80); |
| 120 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, HandleCount) == 96); |
| 121 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, VMCounters) == 112); |
| 122 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, IOCounters) == 208); |
| 123 | KMP_BUILD_ASSERT(offsetof(SYSTEM_PROCESS_INFORMATION, Threads) == 256); |
| 124 | #endif |
| 125 | |
| 126 | typedef NTSTATUS(NTAPI *NtQuerySystemInformation_t)(SYSTEM_INFORMATION_CLASS, |
| 127 | PVOID, ULONG, PULONG); |
| 128 | NtQuerySystemInformation_t NtQuerySystemInformation = NULL; |
| 129 | |
| 130 | HMODULE ntdll = NULL; |
| 131 | |
| 132 | /* End of NtQuerySystemInformation()-related code */ |
| 133 | |
| 134 | static HMODULE kernel32 = NULL; |
| 135 | |
| 136 | #if KMP_HANDLE_SIGNALS |
| 137 | typedef void (*sig_func_t)(int); |
| 138 | static sig_func_t __kmp_sighldrs[NSIG]; |
| 139 | static int __kmp_siginstalled[NSIG]; |
| 140 | #endif |
| 141 | |
| 142 | #if KMP_USE_MONITOR |
| 143 | static HANDLE __kmp_monitor_ev; |
| 144 | #endif |
| 145 | static kmp_int64 __kmp_win32_time; |
| 146 | double __kmp_win32_tick; |
| 147 | |
| 148 | int __kmp_init_runtime = FALSE; |
| 149 | CRITICAL_SECTION __kmp_win32_section; |
| 150 | |
| 151 | void __kmp_win32_mutex_init(kmp_win32_mutex_t *mx) { |
| 152 | InitializeCriticalSection(&mx->cs); |
| 153 | #if USE_ITT_BUILD |
| 154 | __kmp_itt_system_object_created(&mx->cs, "Critical Section" ); |
| 155 | #endif /* USE_ITT_BUILD */ |
| 156 | } |
| 157 | |
| 158 | void __kmp_win32_mutex_destroy(kmp_win32_mutex_t *mx) { |
| 159 | DeleteCriticalSection(&mx->cs); |
| 160 | } |
| 161 | |
| 162 | void __kmp_win32_mutex_lock(kmp_win32_mutex_t *mx) { |
| 163 | EnterCriticalSection(&mx->cs); |
| 164 | } |
| 165 | |
| 166 | int __kmp_win32_mutex_trylock(kmp_win32_mutex_t *mx) { |
| 167 | return TryEnterCriticalSection(&mx->cs); |
| 168 | } |
| 169 | |
| 170 | void __kmp_win32_mutex_unlock(kmp_win32_mutex_t *mx) { |
| 171 | LeaveCriticalSection(&mx->cs); |
| 172 | } |
| 173 | |
| 174 | void __kmp_win32_cond_init(kmp_win32_cond_t *cv) { |
| 175 | cv->waiters_count_ = 0; |
| 176 | cv->wait_generation_count_ = 0; |
| 177 | cv->release_count_ = 0; |
| 178 | |
| 179 | /* Initialize the critical section */ |
| 180 | __kmp_win32_mutex_init(&cv->waiters_count_lock_); |
| 181 | |
| 182 | /* Create a manual-reset event. */ |
| 183 | cv->event_ = CreateEvent(NULL, // no security |
| 184 | TRUE, // manual-reset |
| 185 | FALSE, // non-signaled initially |
| 186 | NULL); // unnamed |
| 187 | #if USE_ITT_BUILD |
| 188 | __kmp_itt_system_object_created(cv->event_, "Event" ); |
| 189 | #endif /* USE_ITT_BUILD */ |
| 190 | } |
| 191 | |
| 192 | void __kmp_win32_cond_destroy(kmp_win32_cond_t *cv) { |
| 193 | __kmp_win32_mutex_destroy(&cv->waiters_count_lock_); |
| 194 | __kmp_free_handle(cv->event_); |
| 195 | memset(cv, '\0', sizeof(*cv)); |
| 196 | } |
| 197 | |
| 198 | /* TODO associate cv with a team instead of a thread so as to optimize |
| 199 | the case where we wake up a whole team */ |
| 200 | |
| 201 | template <class C> |
| 202 | static void __kmp_win32_cond_wait(kmp_win32_cond_t *cv, kmp_win32_mutex_t *mx, |
| 203 | kmp_info_t *th, C *flag) { |
| 204 | int my_generation; |
| 205 | int last_waiter; |
| 206 | |
| 207 | /* Avoid race conditions */ |
| 208 | __kmp_win32_mutex_lock(&cv->waiters_count_lock_); |
| 209 | |
| 210 | /* Increment count of waiters */ |
| 211 | cv->waiters_count_++; |
| 212 | |
| 213 | /* Store current generation in our activation record. */ |
| 214 | my_generation = cv->wait_generation_count_; |
| 215 | |
| 216 | __kmp_win32_mutex_unlock(&cv->waiters_count_lock_); |
| 217 | __kmp_win32_mutex_unlock(mx); |
| 218 | |
| 219 | for (;;) { |
| 220 | int wait_done = 0; |
| 221 | DWORD res, timeout = 5000; // just tried to quess an appropriate number |
| 222 | /* Wait until the event is signaled */ |
| 223 | res = WaitForSingleObject(cv->event_, timeout); |
| 224 | |
| 225 | if (res == WAIT_OBJECT_0) { |
| 226 | // event signaled |
| 227 | __kmp_win32_mutex_lock(&cv->waiters_count_lock_); |
| 228 | /* Exit the loop when the <cv->event_> is signaled and there are still |
| 229 | waiting threads from this <wait_generation> that haven't been released |
| 230 | from this wait yet. */ |
| 231 | wait_done = (cv->release_count_ > 0) && |
| 232 | (cv->wait_generation_count_ != my_generation); |
| 233 | __kmp_win32_mutex_unlock(&cv->waiters_count_lock_); |
| 234 | } else if (res == WAIT_TIMEOUT || res == WAIT_FAILED) { |
| 235 | // check if the flag and cv counters are in consistent state |
| 236 | // as MS sent us debug dump whith inconsistent state of data |
| 237 | __kmp_win32_mutex_lock(mx); |
| 238 | typename C::flag_t old_f = flag->set_sleeping(); |
| 239 | if (!flag->done_check_val(old_f & ~KMP_BARRIER_SLEEP_STATE)) { |
| 240 | __kmp_win32_mutex_unlock(mx); |
| 241 | continue; |
| 242 | } |
| 243 | // condition fulfilled, exiting |
| 244 | flag->unset_sleeping(); |
| 245 | TCW_PTR(th->th.th_sleep_loc, NULL); |
| 246 | th->th.th_sleep_loc_type = flag_unset; |
| 247 | KF_TRACE(50, ("__kmp_win32_cond_wait: exiting, condition " |
| 248 | "fulfilled: flag's loc(%p): %u\n" , |
| 249 | flag->get(), (unsigned int)flag->load())); |
| 250 | |
| 251 | __kmp_win32_mutex_lock(&cv->waiters_count_lock_); |
| 252 | KMP_DEBUG_ASSERT(cv->waiters_count_ > 0); |
| 253 | cv->release_count_ = cv->waiters_count_; |
| 254 | cv->wait_generation_count_++; |
| 255 | wait_done = 1; |
| 256 | __kmp_win32_mutex_unlock(&cv->waiters_count_lock_); |
| 257 | |
| 258 | __kmp_win32_mutex_unlock(mx); |
| 259 | } |
| 260 | /* there used to be a semicolon after the if statement, it looked like a |
| 261 | bug, so i removed it */ |
| 262 | if (wait_done) |
| 263 | break; |
| 264 | } |
| 265 | |
| 266 | __kmp_win32_mutex_lock(mx); |
| 267 | __kmp_win32_mutex_lock(&cv->waiters_count_lock_); |
| 268 | |
| 269 | cv->waiters_count_--; |
| 270 | cv->release_count_--; |
| 271 | |
| 272 | last_waiter = (cv->release_count_ == 0); |
| 273 | |
| 274 | __kmp_win32_mutex_unlock(&cv->waiters_count_lock_); |
| 275 | |
| 276 | if (last_waiter) { |
| 277 | /* We're the last waiter to be notified, so reset the manual event. */ |
| 278 | ResetEvent(cv->event_); |
| 279 | } |
| 280 | } |
| 281 | |
| 282 | void __kmp_win32_cond_broadcast(kmp_win32_cond_t *cv) { |
| 283 | __kmp_win32_mutex_lock(&cv->waiters_count_lock_); |
| 284 | |
| 285 | if (cv->waiters_count_ > 0) { |
| 286 | SetEvent(cv->event_); |
| 287 | /* Release all the threads in this generation. */ |
| 288 | |
| 289 | cv->release_count_ = cv->waiters_count_; |
| 290 | |
| 291 | /* Start a new generation. */ |
| 292 | cv->wait_generation_count_++; |
| 293 | } |
| 294 | |
| 295 | __kmp_win32_mutex_unlock(&cv->waiters_count_lock_); |
| 296 | } |
| 297 | |
| 298 | void __kmp_win32_cond_signal(kmp_win32_cond_t *cv) { |
| 299 | __kmp_win32_cond_broadcast(cv); |
| 300 | } |
| 301 | |
| 302 | void __kmp_enable(int new_state) { |
| 303 | if (__kmp_init_runtime) |
| 304 | LeaveCriticalSection(&__kmp_win32_section); |
| 305 | } |
| 306 | |
| 307 | void __kmp_disable(int *old_state) { |
| 308 | *old_state = 0; |
| 309 | |
| 310 | if (__kmp_init_runtime) |
| 311 | EnterCriticalSection(&__kmp_win32_section); |
| 312 | } |
| 313 | |
| 314 | void __kmp_suspend_initialize(void) { /* do nothing */ |
| 315 | } |
| 316 | |
| 317 | void __kmp_suspend_initialize_thread(kmp_info_t *th) { |
| 318 | int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init); |
| 319 | int new_value = TRUE; |
| 320 | // Return if already initialized |
| 321 | if (old_value == new_value) |
| 322 | return; |
| 323 | // Wait, then return if being initialized |
| 324 | if (old_value == -1 || |
| 325 | !__kmp_atomic_compare_store(&th->th.th_suspend_init, old_value, -1)) { |
| 326 | while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init) != new_value) { |
| 327 | KMP_CPU_PAUSE(); |
| 328 | } |
| 329 | } else { |
| 330 | // Claim to be the initializer and do initializations |
| 331 | __kmp_win32_cond_init(&th->th.th_suspend_cv); |
| 332 | __kmp_win32_mutex_init(&th->th.th_suspend_mx); |
| 333 | KMP_ATOMIC_ST_REL(&th->th.th_suspend_init, new_value); |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { |
| 338 | if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init)) { |
| 339 | /* this means we have initialize the suspension pthread objects for this |
| 340 | thread in this instance of the process */ |
| 341 | __kmp_win32_cond_destroy(&th->th.th_suspend_cv); |
| 342 | __kmp_win32_mutex_destroy(&th->th.th_suspend_mx); |
| 343 | KMP_ATOMIC_ST_REL(&th->th.th_suspend_init, FALSE); |
| 344 | } |
| 345 | } |
| 346 | |
| 347 | int __kmp_try_suspend_mx(kmp_info_t *th) { |
| 348 | return __kmp_win32_mutex_trylock(&th->th.th_suspend_mx); |
| 349 | } |
| 350 | |
| 351 | void __kmp_lock_suspend_mx(kmp_info_t *th) { |
| 352 | __kmp_win32_mutex_lock(&th->th.th_suspend_mx); |
| 353 | } |
| 354 | |
| 355 | void __kmp_unlock_suspend_mx(kmp_info_t *th) { |
| 356 | __kmp_win32_mutex_unlock(&th->th.th_suspend_mx); |
| 357 | } |
| 358 | |
| 359 | /* This routine puts the calling thread to sleep after setting the |
| 360 | sleep bit for the indicated flag variable to true. */ |
| 361 | template <class C> |
| 362 | static inline void __kmp_suspend_template(int th_gtid, C *flag) { |
| 363 | kmp_info_t *th = __kmp_threads[th_gtid]; |
| 364 | typename C::flag_t old_spin; |
| 365 | |
| 366 | KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag's loc(%p)\n" , |
| 367 | th_gtid, flag->get())); |
| 368 | |
| 369 | __kmp_suspend_initialize_thread(th); |
| 370 | __kmp_lock_suspend_mx(th); |
| 371 | |
| 372 | KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for flag's" |
| 373 | " loc(%p)\n" , |
| 374 | th_gtid, flag->get())); |
| 375 | |
| 376 | /* TODO: shouldn't this use release semantics to ensure that |
| 377 | __kmp_suspend_initialize_thread gets called first? */ |
| 378 | old_spin = flag->set_sleeping(); |
| 379 | TCW_PTR(th->th.th_sleep_loc, (void *)flag); |
| 380 | th->th.th_sleep_loc_type = flag->get_type(); |
| 381 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && |
| 382 | __kmp_pause_status != kmp_soft_paused) { |
| 383 | flag->unset_sleeping(); |
| 384 | TCW_PTR(th->th.th_sleep_loc, NULL); |
| 385 | th->th.th_sleep_loc_type = flag_unset; |
| 386 | __kmp_unlock_suspend_mx(th); |
| 387 | return; |
| 388 | } |
| 389 | |
| 390 | KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for flag's" |
| 391 | " loc(%p)==%u\n" , |
| 392 | th_gtid, flag->get(), (unsigned int)flag->load())); |
| 393 | |
| 394 | if (flag->done_check_val(old_spin) || flag->done_check()) { |
| 395 | flag->unset_sleeping(); |
| 396 | TCW_PTR(th->th.th_sleep_loc, NULL); |
| 397 | th->th.th_sleep_loc_type = flag_unset; |
| 398 | KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " |
| 399 | "for flag's loc(%p)\n" , |
| 400 | th_gtid, flag->get())); |
| 401 | } else { |
| 402 | #ifdef DEBUG_SUSPEND |
| 403 | __kmp_suspend_count++; |
| 404 | #endif |
| 405 | /* Encapsulate in a loop as the documentation states that this may "with |
| 406 | low probability" return when the condition variable has not been signaled |
| 407 | or broadcast */ |
| 408 | int deactivated = FALSE; |
| 409 | |
| 410 | while (flag->is_sleeping()) { |
| 411 | KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " |
| 412 | "kmp_win32_cond_wait()\n" , |
| 413 | th_gtid)); |
| 414 | // Mark the thread as no longer active (only in the first iteration of the |
| 415 | // loop). |
| 416 | if (!deactivated) { |
| 417 | th->th.th_active = FALSE; |
| 418 | if (th->th.th_active_in_pool) { |
| 419 | th->th.th_active_in_pool = FALSE; |
| 420 | KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); |
| 421 | KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); |
| 422 | } |
| 423 | deactivated = TRUE; |
| 424 | } |
| 425 | |
| 426 | KMP_DEBUG_ASSERT(th->th.th_sleep_loc); |
| 427 | KMP_DEBUG_ASSERT(th->th.th_sleep_loc_type == flag->get_type()); |
| 428 | |
| 429 | __kmp_win32_cond_wait(&th->th.th_suspend_cv, &th->th.th_suspend_mx, th, |
| 430 | flag); |
| 431 | |
| 432 | #ifdef KMP_DEBUG |
| 433 | if (flag->is_sleeping()) { |
| 434 | KF_TRACE(100, |
| 435 | ("__kmp_suspend_template: T#%d spurious wakeup\n" , th_gtid)); |
| 436 | } |
| 437 | #endif /* KMP_DEBUG */ |
| 438 | |
| 439 | } // while |
| 440 | |
| 441 | // We may have had the loop variable set before entering the loop body; |
| 442 | // so we need to reset sleep_loc. |
| 443 | TCW_PTR(th->th.th_sleep_loc, NULL); |
| 444 | th->th.th_sleep_loc_type = flag_unset; |
| 445 | |
| 446 | KMP_DEBUG_ASSERT(!flag->is_sleeping()); |
| 447 | KMP_DEBUG_ASSERT(!th->th.th_sleep_loc); |
| 448 | |
| 449 | // Mark the thread as active again (if it was previous marked as inactive) |
| 450 | if (deactivated) { |
| 451 | th->th.th_active = TRUE; |
| 452 | if (TCR_4(th->th.th_in_pool)) { |
| 453 | KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); |
| 454 | th->th.th_active_in_pool = TRUE; |
| 455 | } |
| 456 | } |
| 457 | } |
| 458 | |
| 459 | __kmp_unlock_suspend_mx(th); |
| 460 | KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n" , th_gtid)); |
| 461 | } |
| 462 | |
| 463 | template <bool C, bool S> |
| 464 | void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) { |
| 465 | __kmp_suspend_template(th_gtid, flag); |
| 466 | } |
| 467 | template <bool C, bool S> |
| 468 | void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) { |
| 469 | __kmp_suspend_template(th_gtid, flag); |
| 470 | } |
| 471 | template <bool C, bool S> |
| 472 | void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) { |
| 473 | __kmp_suspend_template(th_gtid, flag); |
| 474 | } |
| 475 | void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { |
| 476 | __kmp_suspend_template(th_gtid, flag); |
| 477 | } |
| 478 | |
| 479 | template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *); |
| 480 | template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *); |
| 481 | template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *); |
| 482 | template void |
| 483 | __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *); |
| 484 | template void |
| 485 | __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *); |
| 486 | |
| 487 | /* This routine signals the thread specified by target_gtid to wake up |
| 488 | after setting the sleep bit indicated by the flag argument to FALSE */ |
| 489 | template <class C> |
| 490 | static inline void __kmp_resume_template(int target_gtid, C *flag) { |
| 491 | kmp_info_t *th = __kmp_threads[target_gtid]; |
| 492 | |
| 493 | #ifdef KMP_DEBUG |
| 494 | int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; |
| 495 | #endif |
| 496 | |
| 497 | KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n" , |
| 498 | gtid, target_gtid)); |
| 499 | |
| 500 | __kmp_suspend_initialize_thread(th); |
| 501 | __kmp_lock_suspend_mx(th); |
| 502 | |
| 503 | if (!flag || flag != th->th.th_sleep_loc) { |
| 504 | // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a |
| 505 | // different location; wake up at new location |
| 506 | flag = (C *)th->th.th_sleep_loc; |
| 507 | } |
| 508 | |
| 509 | // First, check if the flag is null or its type has changed. If so, someone |
| 510 | // else woke it up. |
| 511 | if (!flag || flag->get_type() != th->th.th_sleep_loc_type) { |
| 512 | // simply shows what flag was cast to |
| 513 | KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " |
| 514 | "awake: flag's loc(%p)\n" , |
| 515 | gtid, target_gtid, NULL)); |
| 516 | __kmp_unlock_suspend_mx(th); |
| 517 | return; |
| 518 | } else { |
| 519 | if (!flag->is_sleeping()) { |
| 520 | KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " |
| 521 | "awake: flag's loc(%p): %u\n" , |
| 522 | gtid, target_gtid, flag->get(), (unsigned int)flag->load())); |
| 523 | __kmp_unlock_suspend_mx(th); |
| 524 | return; |
| 525 | } |
| 526 | } |
| 527 | KMP_DEBUG_ASSERT(flag); |
| 528 | flag->unset_sleeping(); |
| 529 | TCW_PTR(th->th.th_sleep_loc, NULL); |
| 530 | th->th.th_sleep_loc_type = flag_unset; |
| 531 | |
| 532 | KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset sleep " |
| 533 | "bit for flag's loc(%p)\n" , |
| 534 | gtid, target_gtid, flag->get())); |
| 535 | |
| 536 | __kmp_win32_cond_signal(&th->th.th_suspend_cv); |
| 537 | __kmp_unlock_suspend_mx(th); |
| 538 | |
| 539 | KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" |
| 540 | " for T#%d\n" , |
| 541 | gtid, target_gtid)); |
| 542 | } |
| 543 | |
| 544 | template <bool C, bool S> |
| 545 | void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) { |
| 546 | __kmp_resume_template(target_gtid, flag); |
| 547 | } |
| 548 | template <bool C, bool S> |
| 549 | void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) { |
| 550 | __kmp_resume_template(target_gtid, flag); |
| 551 | } |
| 552 | template <bool C, bool S> |
| 553 | void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) { |
| 554 | __kmp_resume_template(target_gtid, flag); |
| 555 | } |
| 556 | void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { |
| 557 | __kmp_resume_template(target_gtid, flag); |
| 558 | } |
| 559 | |
| 560 | template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *); |
| 561 | template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *); |
| 562 | template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *); |
| 563 | template void |
| 564 | __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *); |
| 565 | |
| 566 | void __kmp_yield() { Sleep(0); } |
| 567 | |
| 568 | void __kmp_gtid_set_specific(int gtid) { |
| 569 | if (__kmp_init_gtid) { |
| 570 | KA_TRACE(50, ("__kmp_gtid_set_specific: T#%d key:%d\n" , gtid, |
| 571 | __kmp_gtid_threadprivate_key)); |
| 572 | kmp_intptr_t g = (kmp_intptr_t)gtid; |
| 573 | if (!TlsSetValue(__kmp_gtid_threadprivate_key, (LPVOID)(g + 1))) |
| 574 | KMP_FATAL(TLSSetValueFailed); |
| 575 | } else { |
| 576 | KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n" )); |
| 577 | } |
| 578 | } |
| 579 | |
| 580 | int __kmp_gtid_get_specific() { |
| 581 | int gtid; |
| 582 | if (!__kmp_init_gtid) { |
| 583 | KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " |
| 584 | "KMP_GTID_SHUTDOWN\n" )); |
| 585 | return KMP_GTID_SHUTDOWN; |
| 586 | } |
| 587 | gtid = (int)(kmp_intptr_t)TlsGetValue(__kmp_gtid_threadprivate_key); |
| 588 | if (gtid == 0) { |
| 589 | gtid = KMP_GTID_DNE; |
| 590 | } else { |
| 591 | gtid--; |
| 592 | } |
| 593 | KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n" , |
| 594 | __kmp_gtid_threadprivate_key, gtid)); |
| 595 | return gtid; |
| 596 | } |
| 597 | |
| 598 | void __kmp_affinity_bind_thread(int proc) { |
| 599 | if (__kmp_num_proc_groups > 1) { |
| 600 | // Form the GROUP_AFFINITY struct directly, rather than filling |
| 601 | // out a bit vector and calling __kmp_set_system_affinity(). |
| 602 | GROUP_AFFINITY ga; |
| 603 | KMP_DEBUG_ASSERT((proc >= 0) && (proc < (__kmp_num_proc_groups * CHAR_BIT * |
| 604 | sizeof(DWORD_PTR)))); |
| 605 | ga.Group = proc / (CHAR_BIT * sizeof(DWORD_PTR)); |
| 606 | ga.Mask = (unsigned long long)1 << (proc % (CHAR_BIT * sizeof(DWORD_PTR))); |
| 607 | ga.Reserved[0] = ga.Reserved[1] = ga.Reserved[2] = 0; |
| 608 | |
| 609 | KMP_DEBUG_ASSERT(__kmp_SetThreadGroupAffinity != NULL); |
| 610 | if (__kmp_SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL) == 0) { |
| 611 | DWORD error = GetLastError(); |
| 612 | // AC: continue silently if not verbose |
| 613 | if (__kmp_affinity.flags.verbose) { |
| 614 | kmp_msg_t err_code = KMP_ERR(error); |
| 615 | __kmp_msg(kmp_ms_warning, KMP_MSG(CantSetThreadAffMask), err_code, |
| 616 | __kmp_msg_null); |
| 617 | if (__kmp_generate_warnings == kmp_warnings_off) { |
| 618 | __kmp_str_free(str: &err_code.str); |
| 619 | } |
| 620 | } |
| 621 | } |
| 622 | } else { |
| 623 | kmp_affin_mask_t *mask; |
| 624 | KMP_CPU_ALLOC_ON_STACK(mask); |
| 625 | KMP_CPU_ZERO(mask); |
| 626 | KMP_CPU_SET(proc, mask); |
| 627 | __kmp_set_system_affinity(mask, TRUE); |
| 628 | KMP_CPU_FREE_FROM_STACK(mask); |
| 629 | } |
| 630 | } |
| 631 | |
| 632 | void __kmp_affinity_determine_capable(const char *env_var) { |
| 633 | // All versions of Windows* OS (since Win '95) support |
| 634 | // SetThreadAffinityMask(). |
| 635 | |
| 636 | #if KMP_GROUP_AFFINITY |
| 637 | KMP_AFFINITY_ENABLE(__kmp_num_proc_groups * sizeof(DWORD_PTR)); |
| 638 | #else |
| 639 | KMP_AFFINITY_ENABLE(sizeof(DWORD_PTR)); |
| 640 | #endif |
| 641 | |
| 642 | KA_TRACE(10, ("__kmp_affinity_determine_capable: " |
| 643 | "Windows* OS affinity interface functional (mask size = " |
| 644 | "%" KMP_SIZE_T_SPEC ").\n" , |
| 645 | __kmp_affin_mask_size)); |
| 646 | } |
| 647 | |
| 648 | double __kmp_read_cpu_time(void) { |
| 649 | FILETIME CreationTime, ExitTime, KernelTime, UserTime; |
| 650 | int status; |
| 651 | double cpu_time; |
| 652 | |
| 653 | cpu_time = 0; |
| 654 | |
| 655 | status = GetProcessTimes(GetCurrentProcess(), &CreationTime, &ExitTime, |
| 656 | &KernelTime, &UserTime); |
| 657 | |
| 658 | if (status) { |
| 659 | double sec = 0; |
| 660 | |
| 661 | sec += KernelTime.dwHighDateTime; |
| 662 | sec += UserTime.dwHighDateTime; |
| 663 | |
| 664 | /* Shift left by 32 bits */ |
| 665 | sec *= (double)(1 << 16) * (double)(1 << 16); |
| 666 | |
| 667 | sec += KernelTime.dwLowDateTime; |
| 668 | sec += UserTime.dwLowDateTime; |
| 669 | |
| 670 | cpu_time += (sec * 100.0) / KMP_NSEC_PER_SEC; |
| 671 | } |
| 672 | |
| 673 | return cpu_time; |
| 674 | } |
| 675 | |
| 676 | int __kmp_read_system_info(struct kmp_sys_info *info) { |
| 677 | info->maxrss = 0; /* the maximum resident set size utilized (in kilobytes) */ |
| 678 | info->minflt = 0; /* the number of page faults serviced without any I/O */ |
| 679 | info->majflt = 0; /* the number of page faults serviced that required I/O */ |
| 680 | info->nswap = 0; // the number of times a process was "swapped" out of memory |
| 681 | info->inblock = 0; // the number of times the file system had to perform input |
| 682 | info->oublock = 0; // number of times the file system had to perform output |
| 683 | info->nvcsw = 0; /* the number of times a context switch was voluntarily */ |
| 684 | info->nivcsw = 0; /* the number of times a context switch was forced */ |
| 685 | |
| 686 | return 1; |
| 687 | } |
| 688 | |
| 689 | void __kmp_runtime_initialize(void) { |
| 690 | SYSTEM_INFO info; |
| 691 | kmp_str_buf_t path; |
| 692 | UINT path_size; |
| 693 | |
| 694 | if (__kmp_init_runtime) { |
| 695 | return; |
| 696 | } |
| 697 | |
| 698 | #if KMP_DYNAMIC_LIB |
| 699 | /* Pin dynamic library for the lifetime of application */ |
| 700 | { |
| 701 | // First, turn off error message boxes |
| 702 | UINT err_mode = SetErrorMode(SEM_FAILCRITICALERRORS); |
| 703 | HMODULE h; |
| 704 | BOOL ret = GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS | |
| 705 | GET_MODULE_HANDLE_EX_FLAG_PIN, |
| 706 | (LPCTSTR)&__kmp_serial_initialize, &h); |
| 707 | (void)ret; |
| 708 | KMP_DEBUG_ASSERT2(h && ret, "OpenMP RTL cannot find itself loaded" ); |
| 709 | SetErrorMode(err_mode); // Restore error mode |
| 710 | KA_TRACE(10, ("__kmp_runtime_initialize: dynamic library pinned\n" )); |
| 711 | } |
| 712 | #endif |
| 713 | |
| 714 | InitializeCriticalSection(&__kmp_win32_section); |
| 715 | #if USE_ITT_BUILD |
| 716 | __kmp_itt_system_object_created(&__kmp_win32_section, "Critical Section" ); |
| 717 | #endif /* USE_ITT_BUILD */ |
| 718 | __kmp_initialize_system_tick(); |
| 719 | |
| 720 | #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) |
| 721 | if (!__kmp_cpuinfo.initialized) { |
| 722 | __kmp_query_cpuid(p: &__kmp_cpuinfo); |
| 723 | } |
| 724 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
| 725 | |
| 726 | /* Set up minimum number of threads to switch to TLS gtid */ |
| 727 | #if KMP_OS_WINDOWS && !KMP_DYNAMIC_LIB |
| 728 | // Windows* OS, static library. |
| 729 | /* New thread may use stack space previously used by another thread, |
| 730 | currently terminated. On Windows* OS, in case of static linking, we do not |
| 731 | know the moment of thread termination, and our structures (__kmp_threads |
| 732 | and __kmp_root arrays) are still keep info about dead threads. This leads |
| 733 | to problem in __kmp_get_global_thread_id() function: it wrongly finds gtid |
| 734 | (by searching through stack addresses of all known threads) for |
| 735 | unregistered foreign tread. |
| 736 | |
| 737 | Setting __kmp_tls_gtid_min to 0 workarounds this problem: |
| 738 | __kmp_get_global_thread_id() does not search through stacks, but get gtid |
| 739 | from TLS immediately. |
| 740 | --ln |
| 741 | */ |
| 742 | __kmp_tls_gtid_min = 0; |
| 743 | #else |
| 744 | __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; |
| 745 | #endif |
| 746 | |
| 747 | /* for the static library */ |
| 748 | if (!__kmp_gtid_threadprivate_key) { |
| 749 | __kmp_gtid_threadprivate_key = TlsAlloc(); |
| 750 | if (__kmp_gtid_threadprivate_key == TLS_OUT_OF_INDEXES) { |
| 751 | KMP_FATAL(TLSOutOfIndexes); |
| 752 | } |
| 753 | } |
| 754 | |
| 755 | // Load ntdll.dll. |
| 756 | /* Simple GetModuleHandle( "ntdll.dl" ) is not suitable due to security issue |
| 757 | (see http://www.microsoft.com/technet/security/advisory/2269637.mspx). We |
| 758 | have to specify full path to the library. */ |
| 759 | __kmp_str_buf_init(&path); |
| 760 | path_size = GetSystemDirectory(path.str, path.size); |
| 761 | KMP_DEBUG_ASSERT(path_size > 0); |
| 762 | if (path_size >= path.size) { |
| 763 | // Buffer is too short. Expand the buffer and try again. |
| 764 | __kmp_str_buf_reserve(&path, path_size); |
| 765 | path_size = GetSystemDirectory(path.str, path.size); |
| 766 | KMP_DEBUG_ASSERT(path_size > 0); |
| 767 | } |
| 768 | if (path_size > 0 && path_size < path.size) { |
| 769 | // Now we have system directory name in the buffer. |
| 770 | // Append backslash and name of dll to form full path, |
| 771 | path.used = path_size; |
| 772 | __kmp_str_buf_print(buffer: &path, format: "\\%s" , "ntdll.dll" ); |
| 773 | |
| 774 | // Now load ntdll using full path. |
| 775 | ntdll = GetModuleHandle(path.str); |
| 776 | } |
| 777 | |
| 778 | KMP_DEBUG_ASSERT(ntdll != NULL); |
| 779 | if (ntdll != NULL) { |
| 780 | NtQuerySystemInformation = (NtQuerySystemInformation_t)GetProcAddress( |
| 781 | ntdll, "NtQuerySystemInformation" ); |
| 782 | } |
| 783 | KMP_DEBUG_ASSERT(NtQuerySystemInformation != NULL); |
| 784 | |
| 785 | #if KMP_GROUP_AFFINITY |
| 786 | // Load kernel32.dll. |
| 787 | // Same caveat - must use full system path name. |
| 788 | if (path_size > 0 && path_size < path.size) { |
| 789 | // Truncate the buffer back to just the system path length, |
| 790 | // discarding "\\ntdll.dll", and replacing it with "kernel32.dll". |
| 791 | path.used = path_size; |
| 792 | __kmp_str_buf_print(&path, "\\%s" , "kernel32.dll" ); |
| 793 | |
| 794 | // Load kernel32.dll using full path. |
| 795 | kernel32 = GetModuleHandle(path.str); |
| 796 | KA_TRACE(10, ("__kmp_runtime_initialize: kernel32.dll = %s\n" , path.str)); |
| 797 | |
| 798 | // Load the function pointers to kernel32.dll routines |
| 799 | // that may or may not exist on this system. |
| 800 | if (kernel32 != NULL) { |
| 801 | __kmp_GetActiveProcessorCount = |
| 802 | (kmp_GetActiveProcessorCount_t)GetProcAddress( |
| 803 | kernel32, "GetActiveProcessorCount" ); |
| 804 | __kmp_GetActiveProcessorGroupCount = |
| 805 | (kmp_GetActiveProcessorGroupCount_t)GetProcAddress( |
| 806 | kernel32, "GetActiveProcessorGroupCount" ); |
| 807 | __kmp_GetThreadGroupAffinity = |
| 808 | (kmp_GetThreadGroupAffinity_t)GetProcAddress( |
| 809 | kernel32, "GetThreadGroupAffinity" ); |
| 810 | __kmp_SetThreadGroupAffinity = |
| 811 | (kmp_SetThreadGroupAffinity_t)GetProcAddress( |
| 812 | kernel32, "SetThreadGroupAffinity" ); |
| 813 | |
| 814 | KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_GetActiveProcessorCount" |
| 815 | " = %p\n" , |
| 816 | __kmp_GetActiveProcessorCount)); |
| 817 | KA_TRACE(10, ("__kmp_runtime_initialize: " |
| 818 | "__kmp_GetActiveProcessorGroupCount = %p\n" , |
| 819 | __kmp_GetActiveProcessorGroupCount)); |
| 820 | KA_TRACE(10, ("__kmp_runtime_initialize:__kmp_GetThreadGroupAffinity" |
| 821 | " = %p\n" , |
| 822 | __kmp_GetThreadGroupAffinity)); |
| 823 | KA_TRACE(10, ("__kmp_runtime_initialize: __kmp_SetThreadGroupAffinity" |
| 824 | " = %p\n" , |
| 825 | __kmp_SetThreadGroupAffinity)); |
| 826 | KA_TRACE(10, ("__kmp_runtime_initialize: sizeof(kmp_affin_mask_t) = %d\n" , |
| 827 | sizeof(kmp_affin_mask_t))); |
| 828 | |
| 829 | // See if group affinity is supported on this system. |
| 830 | // If so, calculate the #groups and #procs. |
| 831 | // |
| 832 | // Group affinity was introduced with Windows* 7 OS and |
| 833 | // Windows* Server 2008 R2 OS. |
| 834 | if ((__kmp_GetActiveProcessorCount != NULL) && |
| 835 | (__kmp_GetActiveProcessorGroupCount != NULL) && |
| 836 | (__kmp_GetThreadGroupAffinity != NULL) && |
| 837 | (__kmp_SetThreadGroupAffinity != NULL) && |
| 838 | ((__kmp_num_proc_groups = __kmp_GetActiveProcessorGroupCount()) > |
| 839 | 1)) { |
| 840 | // Calculate the total number of active OS procs. |
| 841 | int i; |
| 842 | |
| 843 | KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups" |
| 844 | " detected\n" , |
| 845 | __kmp_num_proc_groups)); |
| 846 | |
| 847 | __kmp_xproc = 0; |
| 848 | |
| 849 | for (i = 0; i < __kmp_num_proc_groups; i++) { |
| 850 | DWORD size = __kmp_GetActiveProcessorCount(i); |
| 851 | __kmp_xproc += size; |
| 852 | KA_TRACE(10, ("__kmp_runtime_initialize: proc group %d size = %d\n" , |
| 853 | i, size)); |
| 854 | } |
| 855 | } else { |
| 856 | KA_TRACE(10, ("__kmp_runtime_initialize: %d processor groups" |
| 857 | " detected\n" , |
| 858 | __kmp_num_proc_groups)); |
| 859 | } |
| 860 | } |
| 861 | } |
| 862 | if (__kmp_num_proc_groups <= 1) { |
| 863 | GetSystemInfo(&info); |
| 864 | __kmp_xproc = info.dwNumberOfProcessors; |
| 865 | } |
| 866 | #else |
| 867 | (void)kernel32; |
| 868 | GetSystemInfo(&info); |
| 869 | __kmp_xproc = info.dwNumberOfProcessors; |
| 870 | #endif /* KMP_GROUP_AFFINITY */ |
| 871 | |
| 872 | // If the OS said there were 0 procs, take a guess and use a value of 2. |
| 873 | // This is done for Linux* OS, also. Do we need error / warning? |
| 874 | if (__kmp_xproc <= 0) { |
| 875 | __kmp_xproc = 2; |
| 876 | } |
| 877 | |
| 878 | KA_TRACE(5, |
| 879 | ("__kmp_runtime_initialize: total processors = %d\n" , __kmp_xproc)); |
| 880 | |
| 881 | __kmp_str_buf_free(buffer: &path); |
| 882 | |
| 883 | #if USE_ITT_BUILD |
| 884 | __kmp_itt_initialize(); |
| 885 | #endif /* USE_ITT_BUILD */ |
| 886 | |
| 887 | __kmp_init_runtime = TRUE; |
| 888 | } // __kmp_runtime_initialize |
| 889 | |
| 890 | void __kmp_runtime_destroy(void) { |
| 891 | if (!__kmp_init_runtime) { |
| 892 | return; |
| 893 | } |
| 894 | |
| 895 | #if USE_ITT_BUILD |
| 896 | __kmp_itt_destroy(); |
| 897 | #endif /* USE_ITT_BUILD */ |
| 898 | |
| 899 | /* we can't DeleteCriticalsection( & __kmp_win32_section ); */ |
| 900 | /* due to the KX_TRACE() commands */ |
| 901 | KA_TRACE(40, ("__kmp_runtime_destroy\n" )); |
| 902 | |
| 903 | if (__kmp_gtid_threadprivate_key) { |
| 904 | TlsFree(__kmp_gtid_threadprivate_key); |
| 905 | __kmp_gtid_threadprivate_key = 0; |
| 906 | } |
| 907 | |
| 908 | __kmp_affinity_uninitialize(); |
| 909 | DeleteCriticalSection(&__kmp_win32_section); |
| 910 | |
| 911 | ntdll = NULL; |
| 912 | NtQuerySystemInformation = NULL; |
| 913 | |
| 914 | #if KMP_ARCH_X86_64 |
| 915 | kernel32 = NULL; |
| 916 | __kmp_GetActiveProcessorCount = NULL; |
| 917 | __kmp_GetActiveProcessorGroupCount = NULL; |
| 918 | __kmp_GetThreadGroupAffinity = NULL; |
| 919 | __kmp_SetThreadGroupAffinity = NULL; |
| 920 | #endif // KMP_ARCH_X86_64 |
| 921 | |
| 922 | __kmp_init_runtime = FALSE; |
| 923 | } |
| 924 | |
| 925 | void __kmp_terminate_thread(int gtid) { |
| 926 | kmp_info_t *th = __kmp_threads[gtid]; |
| 927 | |
| 928 | if (!th) |
| 929 | return; |
| 930 | |
| 931 | KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n" , gtid)); |
| 932 | |
| 933 | if (TerminateThread(th->th.th_info.ds.ds_thread, (DWORD)-1) == FALSE) { |
| 934 | /* It's OK, the thread may have exited already */ |
| 935 | } |
| 936 | __kmp_free_handle(th->th.th_info.ds.ds_thread); |
| 937 | } |
| 938 | |
| 939 | void __kmp_clear_system_time(void) { |
| 940 | LARGE_INTEGER time; |
| 941 | QueryPerformanceCounter(&time); |
| 942 | __kmp_win32_time = (kmp_int64)time.QuadPart; |
| 943 | } |
| 944 | |
| 945 | void __kmp_initialize_system_tick(void) { |
| 946 | { |
| 947 | BOOL status; |
| 948 | LARGE_INTEGER freq; |
| 949 | |
| 950 | status = QueryPerformanceFrequency(&freq); |
| 951 | if (!status) { |
| 952 | DWORD error = GetLastError(); |
| 953 | __kmp_fatal(KMP_MSG(FunctionError, "QueryPerformanceFrequency()" ), |
| 954 | KMP_ERR(error), __kmp_msg_null); |
| 955 | |
| 956 | } else { |
| 957 | __kmp_win32_tick = ((double)1.0) / (double)freq.QuadPart; |
| 958 | } |
| 959 | } |
| 960 | } |
| 961 | |
| 962 | /* Calculate the elapsed wall clock time for the user */ |
| 963 | |
| 964 | void __kmp_elapsed(double *t) { |
| 965 | LARGE_INTEGER now; |
| 966 | QueryPerformanceCounter(&now); |
| 967 | *t = ((double)now.QuadPart) * __kmp_win32_tick; |
| 968 | } |
| 969 | |
| 970 | /* Calculate the elapsed wall clock tick for the user */ |
| 971 | |
| 972 | void __kmp_elapsed_tick(double *t) { *t = __kmp_win32_tick; } |
| 973 | |
| 974 | void __kmp_read_system_time(double *delta) { |
| 975 | if (delta != NULL) { |
| 976 | LARGE_INTEGER now; |
| 977 | QueryPerformanceCounter(&now); |
| 978 | *delta = ((double)(((kmp_int64)now.QuadPart) - __kmp_win32_time)) * |
| 979 | __kmp_win32_tick; |
| 980 | } |
| 981 | } |
| 982 | |
| 983 | /* Return the current time stamp in nsec */ |
| 984 | kmp_uint64 __kmp_now_nsec() { |
| 985 | LARGE_INTEGER now; |
| 986 | QueryPerformanceCounter(&now); |
| 987 | return 1e9 * __kmp_win32_tick * now.QuadPart; |
| 988 | } |
| 989 | |
| 990 | extern "C" void *__stdcall __kmp_launch_worker(void *arg) { |
| 991 | volatile void *stack_data; |
| 992 | void *exit_val; |
| 993 | void *padding = 0; |
| 994 | kmp_info_t *this_thr = (kmp_info_t *)arg; |
| 995 | int gtid; |
| 996 | |
| 997 | gtid = this_thr->th.th_info.ds.ds_gtid; |
| 998 | __kmp_gtid_set_specific(gtid); |
| 999 | #ifdef KMP_TDATA_GTID |
| 1000 | #error "This define causes problems with LoadLibrary() + declspec(thread) " \ |
| 1001 | "on Windows* OS. See CQ50564, tests kmp_load_library*.c and this MSDN " \ |
| 1002 | "reference: http://support.microsoft.com/kb/118816" |
| 1003 | //__kmp_gtid = gtid; |
| 1004 | #endif |
| 1005 | |
| 1006 | #if USE_ITT_BUILD |
| 1007 | __kmp_itt_thread_name(gtid); |
| 1008 | #endif /* USE_ITT_BUILD */ |
| 1009 | |
| 1010 | __kmp_affinity_bind_init_mask(gtid); |
| 1011 | |
| 1012 | #if KMP_ARCH_X86 || KMP_ARCH_X86_64 |
| 1013 | // Set FP control regs to be a copy of the parallel initialization thread's. |
| 1014 | __kmp_clear_x87_fpu_status_word(); |
| 1015 | __kmp_load_x87_fpu_control_word(p: &__kmp_init_x87_fpu_control_word); |
| 1016 | __kmp_load_mxcsr(p: &__kmp_init_mxcsr); |
| 1017 | #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ |
| 1018 | |
| 1019 | if (__kmp_stkoffset > 0 && gtid > 0) { |
| 1020 | padding = KMP_ALLOCA(gtid * __kmp_stkoffset); |
| 1021 | (void)padding; |
| 1022 | } |
| 1023 | |
| 1024 | KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive); |
| 1025 | this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId(); |
| 1026 | TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE); |
| 1027 | |
| 1028 | if (TCR_4(__kmp_gtid_mode) < |
| 1029 | 2) { // check stack only if it is used to get gtid |
| 1030 | TCW_PTR(this_thr->th.th_info.ds.ds_stackbase, &stack_data); |
| 1031 | KMP_ASSERT(this_thr->th.th_info.ds.ds_stackgrow == FALSE); |
| 1032 | __kmp_check_stack_overlap(thr: this_thr); |
| 1033 | } |
| 1034 | KMP_MB(); |
| 1035 | exit_val = __kmp_launch_thread(thr: this_thr); |
| 1036 | KMP_FSYNC_RELEASING(&this_thr->th.th_info.ds.ds_alive); |
| 1037 | TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE); |
| 1038 | KMP_MB(); |
| 1039 | return exit_val; |
| 1040 | } |
| 1041 | |
| 1042 | #if KMP_USE_MONITOR |
| 1043 | /* The monitor thread controls all of the threads in the complex */ |
| 1044 | |
| 1045 | void *__stdcall __kmp_launch_monitor(void *arg) { |
| 1046 | DWORD wait_status; |
| 1047 | kmp_thread_t monitor; |
| 1048 | int status; |
| 1049 | int interval; |
| 1050 | kmp_info_t *this_thr = (kmp_info_t *)arg; |
| 1051 | |
| 1052 | KMP_DEBUG_ASSERT(__kmp_init_monitor); |
| 1053 | TCW_4(__kmp_init_monitor, 2); // AC: Signal library that monitor has started |
| 1054 | // TODO: hide "2" in enum (like {true,false,started}) |
| 1055 | this_thr->th.th_info.ds.ds_thread_id = GetCurrentThreadId(); |
| 1056 | TCW_4(this_thr->th.th_info.ds.ds_alive, TRUE); |
| 1057 | |
| 1058 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1059 | KA_TRACE(10, ("__kmp_launch_monitor: launched\n" )); |
| 1060 | |
| 1061 | monitor = GetCurrentThread(); |
| 1062 | |
| 1063 | /* set thread priority */ |
| 1064 | status = SetThreadPriority(monitor, THREAD_PRIORITY_HIGHEST); |
| 1065 | if (!status) { |
| 1066 | DWORD error = GetLastError(); |
| 1067 | __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null); |
| 1068 | } |
| 1069 | |
| 1070 | /* register us as monitor */ |
| 1071 | __kmp_gtid_set_specific(KMP_GTID_MONITOR); |
| 1072 | #ifdef KMP_TDATA_GTID |
| 1073 | #error "This define causes problems with LoadLibrary() + declspec(thread) " \ |
| 1074 | "on Windows* OS. See CQ50564, tests kmp_load_library*.c and this MSDN " \ |
| 1075 | "reference: http://support.microsoft.com/kb/118816" |
| 1076 | //__kmp_gtid = KMP_GTID_MONITOR; |
| 1077 | #endif |
| 1078 | |
| 1079 | #if USE_ITT_BUILD |
| 1080 | __kmp_itt_thread_ignore(); // Instruct Intel(R) Threading Tools to ignore |
| 1081 | // monitor thread. |
| 1082 | #endif /* USE_ITT_BUILD */ |
| 1083 | |
| 1084 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1085 | |
| 1086 | interval = (1000 / __kmp_monitor_wakeups); /* in milliseconds */ |
| 1087 | |
| 1088 | while (!TCR_4(__kmp_global.g.g_done)) { |
| 1089 | /* This thread monitors the state of the system */ |
| 1090 | |
| 1091 | KA_TRACE(15, ("__kmp_launch_monitor: update\n" )); |
| 1092 | |
| 1093 | wait_status = WaitForSingleObject(__kmp_monitor_ev, interval); |
| 1094 | |
| 1095 | if (wait_status == WAIT_TIMEOUT) { |
| 1096 | TCW_4(__kmp_global.g.g_time.dt.t_value, |
| 1097 | TCR_4(__kmp_global.g.g_time.dt.t_value) + 1); |
| 1098 | } |
| 1099 | |
| 1100 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1101 | } |
| 1102 | |
| 1103 | KA_TRACE(10, ("__kmp_launch_monitor: finished\n" )); |
| 1104 | |
| 1105 | status = SetThreadPriority(monitor, THREAD_PRIORITY_NORMAL); |
| 1106 | if (!status) { |
| 1107 | DWORD error = GetLastError(); |
| 1108 | __kmp_fatal(KMP_MSG(CantSetThreadPriority), KMP_ERR(error), __kmp_msg_null); |
| 1109 | } |
| 1110 | |
| 1111 | if (__kmp_global.g.g_abort != 0) { |
| 1112 | /* now we need to terminate the worker threads */ |
| 1113 | /* the value of t_abort is the signal we caught */ |
| 1114 | int gtid; |
| 1115 | |
| 1116 | KA_TRACE(10, ("__kmp_launch_monitor: terminate sig=%d\n" , |
| 1117 | (__kmp_global.g.g_abort))); |
| 1118 | |
| 1119 | /* terminate the OpenMP worker threads */ |
| 1120 | /* TODO this is not valid for sibling threads!! |
| 1121 | * the uber master might not be 0 anymore.. */ |
| 1122 | for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid) |
| 1123 | __kmp_terminate_thread(gtid); |
| 1124 | |
| 1125 | __kmp_cleanup(); |
| 1126 | |
| 1127 | Sleep(0); |
| 1128 | |
| 1129 | KA_TRACE(10, |
| 1130 | ("__kmp_launch_monitor: raise sig=%d\n" , __kmp_global.g.g_abort)); |
| 1131 | |
| 1132 | if (__kmp_global.g.g_abort > 0) { |
| 1133 | raise(__kmp_global.g.g_abort); |
| 1134 | } |
| 1135 | } |
| 1136 | |
| 1137 | TCW_4(this_thr->th.th_info.ds.ds_alive, FALSE); |
| 1138 | |
| 1139 | KMP_MB(); |
| 1140 | return arg; |
| 1141 | } |
| 1142 | #endif |
| 1143 | |
| 1144 | void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) { |
| 1145 | kmp_thread_t handle; |
| 1146 | DWORD idThread; |
| 1147 | |
| 1148 | KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n" , gtid)); |
| 1149 | |
| 1150 | th->th.th_info.ds.ds_gtid = gtid; |
| 1151 | |
| 1152 | if (KMP_UBER_GTID(gtid)) { |
| 1153 | int stack_data; |
| 1154 | |
| 1155 | /* TODO: GetCurrentThread() returns a pseudo-handle that is unsuitable for |
| 1156 | other threads to use. Is it appropriate to just use GetCurrentThread? |
| 1157 | When should we close this handle? When unregistering the root? */ |
| 1158 | { |
| 1159 | BOOL rc; |
| 1160 | rc = DuplicateHandle(GetCurrentProcess(), GetCurrentThread(), |
| 1161 | GetCurrentProcess(), &th->th.th_info.ds.ds_thread, 0, |
| 1162 | FALSE, DUPLICATE_SAME_ACCESS); |
| 1163 | KMP_ASSERT(rc); |
| 1164 | KA_TRACE(10, (" __kmp_create_worker: ROOT Handle duplicated, th = %p, " |
| 1165 | "handle = %" KMP_UINTPTR_SPEC "\n" , |
| 1166 | (LPVOID)th, th->th.th_info.ds.ds_thread)); |
| 1167 | th->th.th_info.ds.ds_thread_id = GetCurrentThreadId(); |
| 1168 | } |
| 1169 | if (TCR_4(__kmp_gtid_mode) < 2) { // check stack only if used to get gtid |
| 1170 | /* we will dynamically update the stack range if gtid_mode == 1 */ |
| 1171 | TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data); |
| 1172 | TCW_PTR(th->th.th_info.ds.ds_stacksize, 0); |
| 1173 | TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE); |
| 1174 | __kmp_check_stack_overlap(thr: th); |
| 1175 | } |
| 1176 | } else { |
| 1177 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1178 | |
| 1179 | /* Set stack size for this thread now. */ |
| 1180 | KA_TRACE(10, |
| 1181 | ("__kmp_create_worker: stack_size = %" KMP_SIZE_T_SPEC " bytes\n" , |
| 1182 | stack_size)); |
| 1183 | |
| 1184 | stack_size += gtid * __kmp_stkoffset; |
| 1185 | |
| 1186 | TCW_PTR(th->th.th_info.ds.ds_stacksize, stack_size); |
| 1187 | TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE); |
| 1188 | |
| 1189 | KA_TRACE(10, |
| 1190 | ("__kmp_create_worker: (before) stack_size = %" KMP_SIZE_T_SPEC |
| 1191 | " bytes, &__kmp_launch_worker = %p, th = %p, &idThread = %p\n" , |
| 1192 | (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker, |
| 1193 | (LPVOID)th, &idThread)); |
| 1194 | |
| 1195 | handle = CreateThread( |
| 1196 | NULL, (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)__kmp_launch_worker, |
| 1197 | (LPVOID)th, STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread); |
| 1198 | |
| 1199 | KA_TRACE(10, |
| 1200 | ("__kmp_create_worker: (after) stack_size = %" KMP_SIZE_T_SPEC |
| 1201 | " bytes, &__kmp_launch_worker = %p, th = %p, " |
| 1202 | "idThread = %u, handle = %" KMP_UINTPTR_SPEC "\n" , |
| 1203 | (SIZE_T)stack_size, (LPTHREAD_START_ROUTINE)&__kmp_launch_worker, |
| 1204 | (LPVOID)th, idThread, handle)); |
| 1205 | |
| 1206 | if (handle == 0) { |
| 1207 | DWORD error = GetLastError(); |
| 1208 | __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null); |
| 1209 | } else { |
| 1210 | th->th.th_info.ds.ds_thread = handle; |
| 1211 | } |
| 1212 | |
| 1213 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1214 | } |
| 1215 | |
| 1216 | KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n" , gtid)); |
| 1217 | } |
| 1218 | |
| 1219 | int __kmp_still_running(kmp_info_t *th) { |
| 1220 | return (WAIT_TIMEOUT == WaitForSingleObject(th->th.th_info.ds.ds_thread, 0)); |
| 1221 | } |
| 1222 | |
| 1223 | #if KMP_USE_MONITOR |
| 1224 | void __kmp_create_monitor(kmp_info_t *th) { |
| 1225 | kmp_thread_t handle; |
| 1226 | DWORD idThread; |
| 1227 | int ideal, new_ideal; |
| 1228 | |
| 1229 | if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { |
| 1230 | // We don't need monitor thread in case of MAX_BLOCKTIME |
| 1231 | KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of " |
| 1232 | "MAX blocktime\n" )); |
| 1233 | th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op |
| 1234 | th->th.th_info.ds.ds_gtid = 0; |
| 1235 | TCW_4(__kmp_init_monitor, 2); // Signal to stop waiting for monitor creation |
| 1236 | return; |
| 1237 | } |
| 1238 | KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n" )); |
| 1239 | |
| 1240 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1241 | |
| 1242 | __kmp_monitor_ev = CreateEvent(NULL, TRUE, FALSE, NULL); |
| 1243 | if (__kmp_monitor_ev == NULL) { |
| 1244 | DWORD error = GetLastError(); |
| 1245 | __kmp_fatal(KMP_MSG(CantCreateEvent), KMP_ERR(error), __kmp_msg_null); |
| 1246 | } |
| 1247 | #if USE_ITT_BUILD |
| 1248 | __kmp_itt_system_object_created(__kmp_monitor_ev, "Event" ); |
| 1249 | #endif /* USE_ITT_BUILD */ |
| 1250 | |
| 1251 | th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR; |
| 1252 | th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR; |
| 1253 | |
| 1254 | // FIXME - on Windows* OS, if __kmp_monitor_stksize = 0, figure out how |
| 1255 | // to automatically expand stacksize based on CreateThread error code. |
| 1256 | if (__kmp_monitor_stksize == 0) { |
| 1257 | __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; |
| 1258 | } |
| 1259 | if (__kmp_monitor_stksize < __kmp_sys_min_stksize) { |
| 1260 | __kmp_monitor_stksize = __kmp_sys_min_stksize; |
| 1261 | } |
| 1262 | |
| 1263 | KA_TRACE(10, ("__kmp_create_monitor: requested stacksize = %d bytes\n" , |
| 1264 | (int)__kmp_monitor_stksize)); |
| 1265 | |
| 1266 | TCW_4(__kmp_global.g.g_time.dt.t_value, 0); |
| 1267 | |
| 1268 | handle = |
| 1269 | CreateThread(NULL, (SIZE_T)__kmp_monitor_stksize, |
| 1270 | (LPTHREAD_START_ROUTINE)__kmp_launch_monitor, (LPVOID)th, |
| 1271 | STACK_SIZE_PARAM_IS_A_RESERVATION, &idThread); |
| 1272 | if (handle == 0) { |
| 1273 | DWORD error = GetLastError(); |
| 1274 | __kmp_fatal(KMP_MSG(CantCreateThread), KMP_ERR(error), __kmp_msg_null); |
| 1275 | } else |
| 1276 | th->th.th_info.ds.ds_thread = handle; |
| 1277 | |
| 1278 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1279 | |
| 1280 | KA_TRACE(10, ("__kmp_create_monitor: monitor created %p\n" , |
| 1281 | (void *)th->th.th_info.ds.ds_thread)); |
| 1282 | } |
| 1283 | #endif |
| 1284 | |
| 1285 | /* Check to see if thread is still alive. |
| 1286 | NOTE: The ExitProcess(code) system call causes all threads to Terminate |
| 1287 | with a exit_val = code. Because of this we can not rely on exit_val having |
| 1288 | any particular value. So this routine may return STILL_ALIVE in exit_val |
| 1289 | even after the thread is dead. */ |
| 1290 | |
| 1291 | int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val) { |
| 1292 | DWORD rc; |
| 1293 | rc = GetExitCodeThread(th->th.th_info.ds.ds_thread, exit_val); |
| 1294 | if (rc == 0) { |
| 1295 | DWORD error = GetLastError(); |
| 1296 | __kmp_fatal(KMP_MSG(FunctionError, "GetExitCodeThread()" ), KMP_ERR(error), |
| 1297 | __kmp_msg_null); |
| 1298 | } |
| 1299 | return (*exit_val == STILL_ACTIVE); |
| 1300 | } |
| 1301 | |
| 1302 | void __kmp_exit_thread(int exit_status) { |
| 1303 | ExitThread(exit_status); |
| 1304 | } // __kmp_exit_thread |
| 1305 | |
| 1306 | // This is a common part for both __kmp_reap_worker() and __kmp_reap_monitor(). |
| 1307 | static void __kmp_reap_common(kmp_info_t *th) { |
| 1308 | DWORD exit_val; |
| 1309 | |
| 1310 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1311 | |
| 1312 | KA_TRACE( |
| 1313 | 10, ("__kmp_reap_common: try to reap (%d)\n" , th->th.th_info.ds.ds_gtid)); |
| 1314 | |
| 1315 | /* 2006-10-19: |
| 1316 | There are two opposite situations: |
| 1317 | 1. Windows* OS keep thread alive after it resets ds_alive flag and |
| 1318 | exits from thread function. (For example, see C70770/Q394281 "unloading of |
| 1319 | dll based on OMP is very slow".) |
| 1320 | 2. Windows* OS may kill thread before it resets ds_alive flag. |
| 1321 | |
| 1322 | Right solution seems to be waiting for *either* thread termination *or* |
| 1323 | ds_alive resetting. */ |
| 1324 | { |
| 1325 | // TODO: This code is very similar to KMP_WAIT. Need to generalize |
| 1326 | // KMP_WAIT to cover this usage also. |
| 1327 | void *obj = NULL; |
| 1328 | kmp_uint32 spins; |
| 1329 | kmp_uint64 time; |
| 1330 | #if USE_ITT_BUILD |
| 1331 | KMP_FSYNC_SPIN_INIT(obj, (void *)&th->th.th_info.ds.ds_alive); |
| 1332 | #endif /* USE_ITT_BUILD */ |
| 1333 | KMP_INIT_YIELD(spins); |
| 1334 | KMP_INIT_BACKOFF(time); |
| 1335 | do { |
| 1336 | #if USE_ITT_BUILD |
| 1337 | KMP_FSYNC_SPIN_PREPARE(obj); |
| 1338 | #endif /* USE_ITT_BUILD */ |
| 1339 | __kmp_is_thread_alive(th, &exit_val); |
| 1340 | KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time); |
| 1341 | } while (exit_val == STILL_ACTIVE && TCR_4(th->th.th_info.ds.ds_alive)); |
| 1342 | #if USE_ITT_BUILD |
| 1343 | if (exit_val == STILL_ACTIVE) { |
| 1344 | KMP_FSYNC_CANCEL(obj); |
| 1345 | } else { |
| 1346 | KMP_FSYNC_SPIN_ACQUIRED(obj); |
| 1347 | } |
| 1348 | #endif /* USE_ITT_BUILD */ |
| 1349 | } |
| 1350 | |
| 1351 | __kmp_free_handle(th->th.th_info.ds.ds_thread); |
| 1352 | |
| 1353 | /* NOTE: The ExitProcess(code) system call causes all threads to Terminate |
| 1354 | with a exit_val = code. Because of this we can not rely on exit_val having |
| 1355 | any particular value. */ |
| 1356 | kmp_intptr_t e = (kmp_intptr_t)exit_val; |
| 1357 | if (exit_val == STILL_ACTIVE) { |
| 1358 | KA_TRACE(1, ("__kmp_reap_common: thread still active.\n" )); |
| 1359 | } else if ((void *)e != (void *)th) { |
| 1360 | KA_TRACE(1, ("__kmp_reap_common: ExitProcess / TerminateThread used?\n" )); |
| 1361 | } |
| 1362 | |
| 1363 | KA_TRACE(10, |
| 1364 | ("__kmp_reap_common: done reaping (%d), handle = %" KMP_UINTPTR_SPEC |
| 1365 | "\n" , |
| 1366 | th->th.th_info.ds.ds_gtid, th->th.th_info.ds.ds_thread)); |
| 1367 | |
| 1368 | th->th.th_info.ds.ds_thread = 0; |
| 1369 | th->th.th_info.ds.ds_tid = KMP_GTID_DNE; |
| 1370 | th->th.th_info.ds.ds_gtid = KMP_GTID_DNE; |
| 1371 | th->th.th_info.ds.ds_thread_id = 0; |
| 1372 | |
| 1373 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1374 | } |
| 1375 | |
| 1376 | #if KMP_USE_MONITOR |
| 1377 | void __kmp_reap_monitor(kmp_info_t *th) { |
| 1378 | int status; |
| 1379 | |
| 1380 | KA_TRACE(10, ("__kmp_reap_monitor: try to reap %p\n" , |
| 1381 | (void *)th->th.th_info.ds.ds_thread)); |
| 1382 | |
| 1383 | // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR. |
| 1384 | // If both tid and gtid are 0, it means the monitor did not ever start. |
| 1385 | // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down. |
| 1386 | KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid); |
| 1387 | if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) { |
| 1388 | KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n" )); |
| 1389 | return; |
| 1390 | } |
| 1391 | |
| 1392 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1393 | |
| 1394 | status = SetEvent(__kmp_monitor_ev); |
| 1395 | if (status == FALSE) { |
| 1396 | DWORD error = GetLastError(); |
| 1397 | __kmp_fatal(KMP_MSG(CantSetEvent), KMP_ERR(error), __kmp_msg_null); |
| 1398 | } |
| 1399 | KA_TRACE(10, ("__kmp_reap_monitor: reaping thread (%d)\n" , |
| 1400 | th->th.th_info.ds.ds_gtid)); |
| 1401 | __kmp_reap_common(th); |
| 1402 | |
| 1403 | __kmp_free_handle(__kmp_monitor_ev); |
| 1404 | |
| 1405 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1406 | } |
| 1407 | #endif |
| 1408 | |
| 1409 | void __kmp_reap_worker(kmp_info_t *th) { |
| 1410 | KA_TRACE(10, ("__kmp_reap_worker: reaping thread (%d)\n" , |
| 1411 | th->th.th_info.ds.ds_gtid)); |
| 1412 | __kmp_reap_common(th); |
| 1413 | } |
| 1414 | |
| 1415 | #if KMP_HANDLE_SIGNALS |
| 1416 | |
| 1417 | static void __kmp_team_handler(int signo) { |
| 1418 | if (__kmp_global.g.g_abort == 0) { |
| 1419 | // Stage 1 signal handler, let's shut down all of the threads. |
| 1420 | if (__kmp_debug_buf) { |
| 1421 | __kmp_dump_debug_buffer(); |
| 1422 | } |
| 1423 | KMP_MB(); // Flush all pending memory write invalidates. |
| 1424 | TCW_4(__kmp_global.g.g_abort, signo); |
| 1425 | KMP_MB(); // Flush all pending memory write invalidates. |
| 1426 | TCW_4(__kmp_global.g.g_done, TRUE); |
| 1427 | KMP_MB(); // Flush all pending memory write invalidates. |
| 1428 | } |
| 1429 | } // __kmp_team_handler |
| 1430 | |
| 1431 | static sig_func_t __kmp_signal(int signum, sig_func_t handler) { |
| 1432 | sig_func_t old = signal(sig: signum, handler: handler); |
| 1433 | if (old == SIG_ERR) { |
| 1434 | int error = errno; |
| 1435 | __kmp_fatal(KMP_MSG(FunctionError, "signal" ), KMP_ERR(error), |
| 1436 | __kmp_msg_null); |
| 1437 | } |
| 1438 | return old; |
| 1439 | } |
| 1440 | |
| 1441 | static void __kmp_install_one_handler(int sig, sig_func_t handler, |
| 1442 | int parallel_init) { |
| 1443 | sig_func_t old; |
| 1444 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1445 | KB_TRACE(60, ("__kmp_install_one_handler: called: sig=%d\n" , sig)); |
| 1446 | if (parallel_init) { |
| 1447 | old = __kmp_signal(signum: sig, handler); |
| 1448 | // SIG_DFL on Windows* OS in NULL or 0. |
| 1449 | if (old == __kmp_sighldrs[sig]) { |
| 1450 | __kmp_siginstalled[sig] = 1; |
| 1451 | } else { // Restore/keep user's handler if one previously installed. |
| 1452 | old = __kmp_signal(signum: sig, handler: old); |
| 1453 | } |
| 1454 | } else { |
| 1455 | // Save initial/system signal handlers to see if user handlers installed. |
| 1456 | // 2009-09-23: It is a dead code. On Windows* OS __kmp_install_signals |
| 1457 | // called once with parallel_init == TRUE. |
| 1458 | old = __kmp_signal(signum: sig, SIG_DFL); |
| 1459 | __kmp_sighldrs[sig] = old; |
| 1460 | __kmp_signal(signum: sig, handler: old); |
| 1461 | } |
| 1462 | KMP_MB(); /* Flush all pending memory write invalidates. */ |
| 1463 | } // __kmp_install_one_handler |
| 1464 | |
| 1465 | static void __kmp_remove_one_handler(int sig) { |
| 1466 | if (__kmp_siginstalled[sig]) { |
| 1467 | sig_func_t old; |
| 1468 | KMP_MB(); // Flush all pending memory write invalidates. |
| 1469 | KB_TRACE(60, ("__kmp_remove_one_handler: called: sig=%d\n" , sig)); |
| 1470 | old = __kmp_signal(signum: sig, handler: __kmp_sighldrs[sig]); |
| 1471 | if (old != __kmp_team_handler) { |
| 1472 | KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, " |
| 1473 | "restoring: sig=%d\n" , |
| 1474 | sig)); |
| 1475 | old = __kmp_signal(signum: sig, handler: old); |
| 1476 | } |
| 1477 | __kmp_sighldrs[sig] = NULL; |
| 1478 | __kmp_siginstalled[sig] = 0; |
| 1479 | KMP_MB(); // Flush all pending memory write invalidates. |
| 1480 | } |
| 1481 | } // __kmp_remove_one_handler |
| 1482 | |
| 1483 | void __kmp_install_signals(int parallel_init) { |
| 1484 | KB_TRACE(10, ("__kmp_install_signals: called\n" )); |
| 1485 | if (!__kmp_handle_signals) { |
| 1486 | KB_TRACE(10, ("__kmp_install_signals: KMP_HANDLE_SIGNALS is false - " |
| 1487 | "handlers not installed\n" )); |
| 1488 | return; |
| 1489 | } |
| 1490 | __kmp_install_one_handler(SIGINT, handler: __kmp_team_handler, parallel_init); |
| 1491 | __kmp_install_one_handler(SIGILL, handler: __kmp_team_handler, parallel_init); |
| 1492 | __kmp_install_one_handler(SIGABRT, handler: __kmp_team_handler, parallel_init); |
| 1493 | __kmp_install_one_handler(SIGFPE, handler: __kmp_team_handler, parallel_init); |
| 1494 | __kmp_install_one_handler(SIGSEGV, handler: __kmp_team_handler, parallel_init); |
| 1495 | __kmp_install_one_handler(SIGTERM, handler: __kmp_team_handler, parallel_init); |
| 1496 | } // __kmp_install_signals |
| 1497 | |
| 1498 | void __kmp_remove_signals(void) { |
| 1499 | int sig; |
| 1500 | KB_TRACE(10, ("__kmp_remove_signals: called\n" )); |
| 1501 | for (sig = 1; sig < NSIG; ++sig) { |
| 1502 | __kmp_remove_one_handler(sig); |
| 1503 | } |
| 1504 | } // __kmp_remove_signals |
| 1505 | |
| 1506 | #endif // KMP_HANDLE_SIGNALS |
| 1507 | |
| 1508 | /* Put the thread to sleep for a time period */ |
| 1509 | void __kmp_thread_sleep(int millis) { |
| 1510 | DWORD status; |
| 1511 | |
| 1512 | status = SleepEx((DWORD)millis, FALSE); |
| 1513 | if (status) { |
| 1514 | DWORD error = GetLastError(); |
| 1515 | __kmp_fatal(KMP_MSG(FunctionError, "SleepEx()" ), KMP_ERR(error), |
| 1516 | __kmp_msg_null); |
| 1517 | } |
| 1518 | } |
| 1519 | |
| 1520 | // Determine whether the given address is mapped into the current address space. |
| 1521 | int __kmp_is_address_mapped(void *addr) { |
| 1522 | MEMORY_BASIC_INFORMATION lpBuffer; |
| 1523 | SIZE_T dwLength; |
| 1524 | |
| 1525 | dwLength = sizeof(MEMORY_BASIC_INFORMATION); |
| 1526 | |
| 1527 | VirtualQuery(addr, &lpBuffer, dwLength); |
| 1528 | |
| 1529 | return !(((lpBuffer.State == MEM_RESERVE) || (lpBuffer.State == MEM_FREE)) || |
| 1530 | ((lpBuffer.Protect == PAGE_NOACCESS) || |
| 1531 | (lpBuffer.Protect == PAGE_EXECUTE))); |
| 1532 | } |
| 1533 | |
| 1534 | kmp_uint64 __kmp_hardware_timestamp(void) { |
| 1535 | kmp_uint64 r = 0; |
| 1536 | |
| 1537 | QueryPerformanceCounter((LARGE_INTEGER *)&r); |
| 1538 | return r; |
| 1539 | } |
| 1540 | |
| 1541 | /* Free handle and check the error code */ |
| 1542 | void __kmp_free_handle(kmp_thread_t tHandle) { |
| 1543 | /* called with parameter type HANDLE also, thus suppose kmp_thread_t defined |
| 1544 | * as HANDLE */ |
| 1545 | BOOL rc; |
| 1546 | rc = CloseHandle(tHandle); |
| 1547 | if (!rc) { |
| 1548 | DWORD error = GetLastError(); |
| 1549 | __kmp_fatal(KMP_MSG(CantCloseHandle), KMP_ERR(error), __kmp_msg_null); |
| 1550 | } |
| 1551 | } |
| 1552 | |
| 1553 | int __kmp_get_load_balance(int max) { |
| 1554 | static ULONG glb_buff_size = 100 * 1024; |
| 1555 | |
| 1556 | // Saved count of the running threads for the thread balance algorithm |
| 1557 | static int glb_running_threads = 0; |
| 1558 | static double glb_call_time = 0; /* Thread balance algorithm call time */ |
| 1559 | |
| 1560 | int running_threads = 0; // Number of running threads in the system. |
| 1561 | NTSTATUS status = 0; |
| 1562 | ULONG buff_size = 0; |
| 1563 | ULONG info_size = 0; |
| 1564 | void *buffer = NULL; |
| 1565 | PSYSTEM_PROCESS_INFORMATION spi = NULL; |
| 1566 | int first_time = 1; |
| 1567 | |
| 1568 | double call_time = 0.0; // start, finish; |
| 1569 | |
| 1570 | __kmp_elapsed(t: &call_time); |
| 1571 | |
| 1572 | if (glb_call_time && |
| 1573 | (call_time - glb_call_time < __kmp_load_balance_interval)) { |
| 1574 | running_threads = glb_running_threads; |
| 1575 | goto finish; |
| 1576 | } |
| 1577 | glb_call_time = call_time; |
| 1578 | |
| 1579 | // Do not spend time on running algorithm if we have a permanent error. |
| 1580 | if (NtQuerySystemInformation == NULL) { |
| 1581 | running_threads = -1; |
| 1582 | goto finish; |
| 1583 | } |
| 1584 | |
| 1585 | if (max <= 0) { |
| 1586 | max = INT_MAX; |
| 1587 | } |
| 1588 | |
| 1589 | do { |
| 1590 | |
| 1591 | if (first_time) { |
| 1592 | buff_size = glb_buff_size; |
| 1593 | } else { |
| 1594 | buff_size = 2 * buff_size; |
| 1595 | } |
| 1596 | |
| 1597 | buffer = KMP_INTERNAL_REALLOC(buffer, buff_size); |
| 1598 | if (buffer == NULL) { |
| 1599 | running_threads = -1; |
| 1600 | goto finish; |
| 1601 | } |
| 1602 | status = NtQuerySystemInformation(SystemProcessInformation, buffer, |
| 1603 | buff_size, &info_size); |
| 1604 | first_time = 0; |
| 1605 | |
| 1606 | } while (status == STATUS_INFO_LENGTH_MISMATCH); |
| 1607 | glb_buff_size = buff_size; |
| 1608 | |
| 1609 | #define CHECK(cond) \ |
| 1610 | { \ |
| 1611 | KMP_DEBUG_ASSERT(cond); \ |
| 1612 | if (!(cond)) { \ |
| 1613 | running_threads = -1; \ |
| 1614 | goto finish; \ |
| 1615 | } \ |
| 1616 | } |
| 1617 | |
| 1618 | CHECK(buff_size >= info_size); |
| 1619 | spi = PSYSTEM_PROCESS_INFORMATION(buffer); |
| 1620 | for (;;) { |
| 1621 | ptrdiff_t offset = uintptr_t(spi) - uintptr_t(buffer); |
| 1622 | CHECK(0 <= offset && |
| 1623 | offset + sizeof(SYSTEM_PROCESS_INFORMATION) < info_size); |
| 1624 | HANDLE pid = spi->ProcessId; |
| 1625 | ULONG num = spi->NumberOfThreads; |
| 1626 | CHECK(num >= 1); |
| 1627 | size_t spi_size = |
| 1628 | sizeof(SYSTEM_PROCESS_INFORMATION) + sizeof(SYSTEM_THREAD) * (num - 1); |
| 1629 | CHECK(offset + spi_size < |
| 1630 | info_size); // Make sure process info record fits the buffer. |
| 1631 | if (spi->NextEntryOffset != 0) { |
| 1632 | CHECK(spi_size <= |
| 1633 | spi->NextEntryOffset); // And do not overlap with the next record. |
| 1634 | } |
| 1635 | // pid == 0 corresponds to the System Idle Process. It always has running |
| 1636 | // threads on all cores. So, we don't consider the running threads of this |
| 1637 | // process. |
| 1638 | if (pid != 0) { |
| 1639 | for (ULONG i = 0; i < num; ++i) { |
| 1640 | THREAD_STATE state = spi->Threads[i].State; |
| 1641 | // Count threads that have Ready or Running state. |
| 1642 | // !!! TODO: Why comment does not match the code??? |
| 1643 | if (state == StateRunning) { |
| 1644 | ++running_threads; |
| 1645 | // Stop counting running threads if the number is already greater than |
| 1646 | // the number of available cores |
| 1647 | if (running_threads >= max) { |
| 1648 | goto finish; |
| 1649 | } |
| 1650 | } |
| 1651 | } |
| 1652 | } |
| 1653 | if (spi->NextEntryOffset == 0) { |
| 1654 | break; |
| 1655 | } |
| 1656 | spi = PSYSTEM_PROCESS_INFORMATION(uintptr_t(spi) + spi->NextEntryOffset); |
| 1657 | } |
| 1658 | |
| 1659 | #undef CHECK |
| 1660 | |
| 1661 | finish: // Clean up and exit. |
| 1662 | |
| 1663 | if (buffer != NULL) { |
| 1664 | KMP_INTERNAL_FREE(buffer); |
| 1665 | } |
| 1666 | |
| 1667 | glb_running_threads = running_threads; |
| 1668 | |
| 1669 | return running_threads; |
| 1670 | } //__kmp_get_load_balance() |
| 1671 | |
| 1672 | // Find symbol from the loaded modules |
| 1673 | void *__kmp_lookup_symbol(const char *name, bool next) { |
| 1674 | HANDLE process = GetCurrentProcess(); |
| 1675 | DWORD needed; |
| 1676 | HMODULE *modules = nullptr; |
| 1677 | if (!EnumProcessModules(process, modules, 0, &needed)) |
| 1678 | return nullptr; |
| 1679 | DWORD num_modules = needed / sizeof(HMODULE); |
| 1680 | modules = (HMODULE *)malloc(num_modules * sizeof(HMODULE)); |
| 1681 | if (!EnumProcessModules(process, modules, needed, &needed)) { |
| 1682 | free(modules); |
| 1683 | return nullptr; |
| 1684 | } |
| 1685 | HMODULE curr_module = nullptr; |
| 1686 | if (next) { |
| 1687 | // Current module needs to be skipped if next flag is true |
| 1688 | if (!GetModuleHandleEx(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS, |
| 1689 | (LPCTSTR)&__kmp_lookup_symbol, &curr_module)) { |
| 1690 | free(modules); |
| 1691 | return nullptr; |
| 1692 | } |
| 1693 | } |
| 1694 | void *proc = nullptr; |
| 1695 | for (uint32_t i = 0; i < num_modules; i++) { |
| 1696 | if (next && modules[i] == curr_module) |
| 1697 | continue; |
| 1698 | proc = (void *)GetProcAddress(modules[i], name); |
| 1699 | if (proc) |
| 1700 | break; |
| 1701 | } |
| 1702 | free(modules); |
| 1703 | return proc; |
| 1704 | } |
| 1705 | |
| 1706 | // Functions for hidden helper task |
| 1707 | void __kmp_hidden_helper_worker_thread_wait() { |
| 1708 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1709 | } |
| 1710 | |
| 1711 | void __kmp_do_initialize_hidden_helper_threads() { |
| 1712 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1713 | } |
| 1714 | |
| 1715 | void __kmp_hidden_helper_threads_initz_wait() { |
| 1716 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1717 | } |
| 1718 | |
| 1719 | void __kmp_hidden_helper_initz_release() { |
| 1720 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1721 | } |
| 1722 | |
| 1723 | void __kmp_hidden_helper_main_thread_wait() { |
| 1724 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1725 | } |
| 1726 | |
| 1727 | void __kmp_hidden_helper_main_thread_release() { |
| 1728 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1729 | } |
| 1730 | |
| 1731 | void __kmp_hidden_helper_worker_thread_signal() { |
| 1732 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1733 | } |
| 1734 | |
| 1735 | void __kmp_hidden_helper_threads_deinitz_wait() { |
| 1736 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1737 | } |
| 1738 | |
| 1739 | void __kmp_hidden_helper_threads_deinitz_release() { |
| 1740 | KMP_ASSERT(0 && "Hidden helper task is not supported on Windows" ); |
| 1741 | } |
| 1742 | |