1 /* 2 * z_Linux_util.cpp -- platform specific routines. 3 */ 4 5 //===----------------------------------------------------------------------===// 6 // 7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 8 // See https://llvm.org/LICENSE.txt for license information. 9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "kmp.h" 14 #include "kmp_affinity.h" 15 #include "kmp_i18n.h" 16 #include "kmp_io.h" 17 #include "kmp_itt.h" 18 #include "kmp_lock.h" 19 #include "kmp_stats.h" 20 #include "kmp_str.h" 21 #include "kmp_wait_release.h" 22 #include "kmp_wrapper_getpid.h" 23 24 #if !KMP_OS_DRAGONFLY && !KMP_OS_FREEBSD && !KMP_OS_NETBSD && !KMP_OS_OPENBSD 25 #include <alloca.h> 26 #endif 27 #include <math.h> // HUGE_VAL. 28 #if KMP_OS_LINUX 29 #include <semaphore.h> 30 #endif // KMP_OS_LINUX 31 #include <sys/resource.h> 32 #if !KMP_OS_AIX 33 #include <sys/syscall.h> 34 #endif 35 #include <sys/time.h> 36 #include <sys/times.h> 37 #include <unistd.h> 38 39 #if KMP_OS_LINUX 40 #include <sys/sysinfo.h> 41 #if KMP_USE_FUTEX 42 // We should really include <futex.h>, but that causes compatibility problems on 43 // different Linux* OS distributions that either require that you include (or 44 // break when you try to include) <pci/types.h>. Since all we need is the two 45 // macros below (which are part of the kernel ABI, so can't change) we just 46 // define the constants here and don't include <futex.h> 47 #ifndef FUTEX_WAIT 48 #define FUTEX_WAIT 0 49 #endif 50 #ifndef FUTEX_WAKE 51 #define FUTEX_WAKE 1 52 #endif 53 #endif 54 #elif KMP_OS_DARWIN 55 #include <mach/mach.h> 56 #include <sys/sysctl.h> 57 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD 58 #include <sys/types.h> 59 #include <sys/sysctl.h> 60 #include <sys/user.h> 61 #include <pthread_np.h> 62 #if KMP_OS_DRAGONFLY 63 #include <kvm.h> 64 #endif 65 #elif KMP_OS_NETBSD || KMP_OS_OPENBSD 66 #include <sys/types.h> 67 #include <sys/sysctl.h> 68 #elif KMP_OS_SOLARIS 69 #include <sys/loadavg.h> 70 #endif 71 72 #include <ctype.h> 73 #include <dirent.h> 74 #include <fcntl.h> 75 76 struct kmp_sys_timer { 77 struct timespec start; 78 }; 79 80 #ifndef TIMEVAL_TO_TIMESPEC 81 // Convert timeval to timespec. 82 #define TIMEVAL_TO_TIMESPEC(tv, ts) \ 83 do { \ 84 (ts)->tv_sec = (tv)->tv_sec; \ 85 (ts)->tv_nsec = (tv)->tv_usec * 1000; \ 86 } while (0) 87 #endif 88 89 // Convert timespec to nanoseconds. 90 #define TS2NS(timespec) \ 91 (((timespec).tv_sec * (long int)1e9) + (timespec).tv_nsec) 92 93 static struct kmp_sys_timer __kmp_sys_timer_data; 94 95 #if KMP_HANDLE_SIGNALS 96 typedef void (*sig_func_t)(int); 97 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG]; 98 static sigset_t __kmp_sigset; 99 #endif 100 101 static int __kmp_init_runtime = FALSE; 102 103 static int __kmp_fork_count = 0; 104 105 static pthread_condattr_t __kmp_suspend_cond_attr; 106 static pthread_mutexattr_t __kmp_suspend_mutex_attr; 107 108 static kmp_cond_align_t __kmp_wait_cv; 109 static kmp_mutex_align_t __kmp_wait_mx; 110 111 kmp_uint64 __kmp_ticks_per_msec = 1000000; 112 kmp_uint64 __kmp_ticks_per_usec = 1000; 113 114 #ifdef DEBUG_SUSPEND 115 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) { 116 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))", 117 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock, 118 cond->c_cond.__c_waiting); 119 } 120 #endif 121 122 #if ((KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED) 123 124 /* Affinity support */ 125 126 void __kmp_affinity_bind_thread(int which) { 127 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(), 128 "Illegal set affinity operation when not capable"); 129 130 kmp_affin_mask_t *mask; 131 KMP_CPU_ALLOC_ON_STACK(mask); 132 KMP_CPU_ZERO(mask); 133 KMP_CPU_SET(which, mask); 134 __kmp_set_system_affinity(mask, TRUE); 135 KMP_CPU_FREE_FROM_STACK(mask); 136 } 137 138 /* Determine if we can access affinity functionality on this version of 139 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set 140 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */ 141 void __kmp_affinity_determine_capable(const char *env_var) { 142 // Check and see if the OS supports thread affinity. 143 144 #if KMP_OS_LINUX 145 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024) 146 #define KMP_CPU_SET_TRY_SIZE CACHE_LINE 147 #elif KMP_OS_FREEBSD 148 #define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t)) 149 #endif 150 151 int verbose = __kmp_affinity.flags.verbose; 152 int warnings = __kmp_affinity.flags.warnings; 153 enum affinity_type type = __kmp_affinity.type; 154 155 #if KMP_OS_LINUX 156 long gCode; 157 unsigned char *buf; 158 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 159 160 // If the syscall returns a suggestion for the size, 161 // then we don't have to search for an appropriate size. 162 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf); 163 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 164 "initial getaffinity call returned %ld errno = %d\n", 165 gCode, errno)); 166 167 if (gCode < 0 && errno != EINVAL) { 168 // System call not supported 169 if (verbose || 170 (warnings && (type != affinity_none) && (type != affinity_default) && 171 (type != affinity_disabled))) { 172 int error = errno; 173 kmp_msg_t err_code = KMP_ERR(error); 174 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 175 err_code, __kmp_msg_null); 176 if (__kmp_generate_warnings == kmp_warnings_off) { 177 __kmp_str_free(&err_code.str); 178 } 179 } 180 KMP_AFFINITY_DISABLE(); 181 KMP_INTERNAL_FREE(buf); 182 return; 183 } else if (gCode > 0) { 184 // The optimal situation: the OS returns the size of the buffer it expects. 185 KMP_AFFINITY_ENABLE(gCode); 186 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 187 "affinity supported (mask size %d)\n", 188 (int)__kmp_affin_mask_size)); 189 KMP_INTERNAL_FREE(buf); 190 return; 191 } 192 193 // Call the getaffinity system call repeatedly with increasing set sizes 194 // until we succeed, or reach an upper bound on the search. 195 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 196 "searching for proper set size\n")); 197 int size; 198 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) { 199 gCode = syscall(__NR_sched_getaffinity, 0, size, buf); 200 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 201 "getaffinity for mask size %ld returned %ld errno = %d\n", 202 size, gCode, errno)); 203 204 if (gCode < 0) { 205 if (errno == ENOSYS) { 206 // We shouldn't get here 207 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 208 "inconsistent OS call behavior: errno == ENOSYS for mask " 209 "size %d\n", 210 size)); 211 if (verbose || 212 (warnings && (type != affinity_none) && 213 (type != affinity_default) && (type != affinity_disabled))) { 214 int error = errno; 215 kmp_msg_t err_code = KMP_ERR(error); 216 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 217 err_code, __kmp_msg_null); 218 if (__kmp_generate_warnings == kmp_warnings_off) { 219 __kmp_str_free(&err_code.str); 220 } 221 } 222 KMP_AFFINITY_DISABLE(); 223 KMP_INTERNAL_FREE(buf); 224 return; 225 } 226 continue; 227 } 228 229 KMP_AFFINITY_ENABLE(gCode); 230 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 231 "affinity supported (mask size %d)\n", 232 (int)__kmp_affin_mask_size)); 233 KMP_INTERNAL_FREE(buf); 234 return; 235 } 236 #elif KMP_OS_FREEBSD 237 long gCode; 238 unsigned char *buf; 239 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 240 gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT, 241 reinterpret_cast<cpuset_t *>(buf)); 242 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 243 "initial getaffinity call returned %d errno = %d\n", 244 gCode, errno)); 245 if (gCode == 0) { 246 KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT); 247 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 248 "affinity supported (mask size %d)\n", 249 (int)__kmp_affin_mask_size)); 250 KMP_INTERNAL_FREE(buf); 251 return; 252 } 253 #endif 254 KMP_INTERNAL_FREE(buf); 255 256 // Affinity is not supported 257 KMP_AFFINITY_DISABLE(); 258 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 259 "cannot determine mask size - affinity not supported\n")); 260 if (verbose || (warnings && (type != affinity_none) && 261 (type != affinity_default) && (type != affinity_disabled))) { 262 KMP_WARNING(AffCantGetMaskSize, env_var); 263 } 264 } 265 266 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 267 268 #if KMP_USE_FUTEX 269 270 int __kmp_futex_determine_capable() { 271 int loc = 0; 272 long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0); 273 int retval = (rc == 0) || (errno != ENOSYS); 274 275 KA_TRACE(10, 276 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno)); 277 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n", 278 retval ? "" : " not")); 279 280 return retval; 281 } 282 283 #endif // KMP_USE_FUTEX 284 285 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_WASM) && (!KMP_ASM_INTRINS) 286 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to 287 use compare_and_store for these routines */ 288 289 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) { 290 kmp_int8 old_value, new_value; 291 292 old_value = TCR_1(*p); 293 new_value = old_value | d; 294 295 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 296 KMP_CPU_PAUSE(); 297 old_value = TCR_1(*p); 298 new_value = old_value | d; 299 } 300 return old_value; 301 } 302 303 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) { 304 kmp_int8 old_value, new_value; 305 306 old_value = TCR_1(*p); 307 new_value = old_value & d; 308 309 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 310 KMP_CPU_PAUSE(); 311 old_value = TCR_1(*p); 312 new_value = old_value & d; 313 } 314 return old_value; 315 } 316 317 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) { 318 kmp_uint32 old_value, new_value; 319 320 old_value = TCR_4(*p); 321 new_value = old_value | d; 322 323 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 324 KMP_CPU_PAUSE(); 325 old_value = TCR_4(*p); 326 new_value = old_value | d; 327 } 328 return old_value; 329 } 330 331 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) { 332 kmp_uint32 old_value, new_value; 333 334 old_value = TCR_4(*p); 335 new_value = old_value & d; 336 337 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 338 KMP_CPU_PAUSE(); 339 old_value = TCR_4(*p); 340 new_value = old_value & d; 341 } 342 return old_value; 343 } 344 345 #if KMP_ARCH_X86 || KMP_ARCH_WASM 346 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) { 347 kmp_int8 old_value, new_value; 348 349 old_value = TCR_1(*p); 350 new_value = old_value + d; 351 352 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 353 KMP_CPU_PAUSE(); 354 old_value = TCR_1(*p); 355 new_value = old_value + d; 356 } 357 return old_value; 358 } 359 360 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) { 361 kmp_int64 old_value, new_value; 362 363 old_value = TCR_8(*p); 364 new_value = old_value + d; 365 366 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 367 KMP_CPU_PAUSE(); 368 old_value = TCR_8(*p); 369 new_value = old_value + d; 370 } 371 return old_value; 372 } 373 #endif /* KMP_ARCH_X86 */ 374 375 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) { 376 kmp_uint64 old_value, new_value; 377 378 old_value = TCR_8(*p); 379 new_value = old_value | d; 380 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 381 KMP_CPU_PAUSE(); 382 old_value = TCR_8(*p); 383 new_value = old_value | d; 384 } 385 return old_value; 386 } 387 388 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) { 389 kmp_uint64 old_value, new_value; 390 391 old_value = TCR_8(*p); 392 new_value = old_value & d; 393 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 394 KMP_CPU_PAUSE(); 395 old_value = TCR_8(*p); 396 new_value = old_value & d; 397 } 398 return old_value; 399 } 400 401 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */ 402 403 void __kmp_terminate_thread(int gtid) { 404 int status; 405 kmp_info_t *th = __kmp_threads[gtid]; 406 407 if (!th) 408 return; 409 410 #ifdef KMP_CANCEL_THREADS 411 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid)); 412 status = pthread_cancel(th->th.th_info.ds.ds_thread); 413 if (status != 0 && status != ESRCH) { 414 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status), 415 __kmp_msg_null); 416 } 417 #endif 418 KMP_YIELD(TRUE); 419 } // 420 421 /* Set thread stack info according to values returned by pthread_getattr_np(). 422 If values are unreasonable, assume call failed and use incremental stack 423 refinement method instead. Returns TRUE if the stack parameters could be 424 determined exactly, FALSE if incremental refinement is necessary. */ 425 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) { 426 int stack_data; 427 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 428 KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_AIX 429 pthread_attr_t attr; 430 int status; 431 size_t size = 0; 432 void *addr = 0; 433 434 /* Always do incremental stack refinement for ubermaster threads since the 435 initial thread stack range can be reduced by sibling thread creation so 436 pthread_attr_getstack may cause thread gtid aliasing */ 437 if (!KMP_UBER_GTID(gtid)) { 438 439 /* Fetch the real thread attributes */ 440 status = pthread_attr_init(&attr); 441 KMP_CHECK_SYSFAIL("pthread_attr_init", status); 442 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD 443 status = pthread_attr_get_np(pthread_self(), &attr); 444 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status); 445 #else 446 status = pthread_getattr_np(pthread_self(), &attr); 447 KMP_CHECK_SYSFAIL("pthread_getattr_np", status); 448 #endif 449 status = pthread_attr_getstack(&attr, &addr, &size); 450 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status); 451 KA_TRACE(60, 452 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:" 453 " %lu, low addr: %p\n", 454 gtid, size, addr)); 455 status = pthread_attr_destroy(&attr); 456 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status); 457 } 458 459 if (size != 0 && addr != 0) { // was stack parameter determination successful? 460 /* Store the correct base and size */ 461 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size)); 462 TCW_PTR(th->th.th_info.ds.ds_stacksize, size); 463 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE); 464 return TRUE; 465 } 466 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD \ 467 || KMP_OS_HURD || KMP_OS_SOLARIS */ 468 /* Use incremental refinement starting from initial conservative estimate */ 469 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0); 470 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data); 471 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE); 472 return FALSE; 473 } 474 475 static void *__kmp_launch_worker(void *thr) { 476 int status, old_type, old_state; 477 #ifdef KMP_BLOCK_SIGNALS 478 sigset_t new_set, old_set; 479 #endif /* KMP_BLOCK_SIGNALS */ 480 void *exit_val; 481 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 482 KMP_OS_OPENBSD || KMP_OS_HURD || KMP_OS_SOLARIS 483 void *volatile padding = 0; 484 #endif 485 int gtid; 486 487 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid; 488 __kmp_gtid_set_specific(gtid); 489 #ifdef KMP_TDATA_GTID 490 __kmp_gtid = gtid; 491 #endif 492 #if KMP_STATS_ENABLED 493 // set thread local index to point to thread-specific stats 494 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats; 495 __kmp_stats_thread_ptr->startLife(); 496 KMP_SET_THREAD_STATE(IDLE); 497 KMP_INIT_PARTITIONED_TIMERS(OMP_idle); 498 #endif 499 500 #if USE_ITT_BUILD 501 __kmp_itt_thread_name(gtid); 502 #endif /* USE_ITT_BUILD */ 503 504 #if KMP_AFFINITY_SUPPORTED 505 __kmp_affinity_bind_init_mask(gtid); 506 #endif 507 508 #ifdef KMP_CANCEL_THREADS 509 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 510 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 511 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 512 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 513 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 514 #endif 515 516 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 517 // Set FP control regs to be a copy of the parallel initialization thread's. 518 __kmp_clear_x87_fpu_status_word(); 519 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); 520 __kmp_load_mxcsr(&__kmp_init_mxcsr); 521 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 522 523 #ifdef KMP_BLOCK_SIGNALS 524 status = sigfillset(&new_set); 525 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 526 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set); 527 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 528 #endif /* KMP_BLOCK_SIGNALS */ 529 530 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 531 KMP_OS_OPENBSD || KMP_OS_HURD || KMP_OS_SOLARIS 532 if (__kmp_stkoffset > 0 && gtid > 0) { 533 padding = KMP_ALLOCA(gtid * __kmp_stkoffset); 534 (void)padding; 535 } 536 #endif 537 538 KMP_MB(); 539 __kmp_set_stack_info(gtid, (kmp_info_t *)thr); 540 541 __kmp_check_stack_overlap((kmp_info_t *)thr); 542 543 exit_val = __kmp_launch_thread((kmp_info_t *)thr); 544 545 #ifdef KMP_BLOCK_SIGNALS 546 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL); 547 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 548 #endif /* KMP_BLOCK_SIGNALS */ 549 550 return exit_val; 551 } 552 553 #if KMP_USE_MONITOR 554 /* The monitor thread controls all of the threads in the complex */ 555 556 static void *__kmp_launch_monitor(void *thr) { 557 int status, old_type, old_state; 558 #ifdef KMP_BLOCK_SIGNALS 559 sigset_t new_set; 560 #endif /* KMP_BLOCK_SIGNALS */ 561 struct timespec interval; 562 563 KMP_MB(); /* Flush all pending memory write invalidates. */ 564 565 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n")); 566 567 /* register us as the monitor thread */ 568 __kmp_gtid_set_specific(KMP_GTID_MONITOR); 569 #ifdef KMP_TDATA_GTID 570 __kmp_gtid = KMP_GTID_MONITOR; 571 #endif 572 573 KMP_MB(); 574 575 #if USE_ITT_BUILD 576 // Instruct Intel(R) Threading Tools to ignore monitor thread. 577 __kmp_itt_thread_ignore(); 578 #endif /* USE_ITT_BUILD */ 579 580 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid, 581 (kmp_info_t *)thr); 582 583 __kmp_check_stack_overlap((kmp_info_t *)thr); 584 585 #ifdef KMP_CANCEL_THREADS 586 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 587 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 588 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 589 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 590 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 591 #endif 592 593 #if KMP_REAL_TIME_FIX 594 // This is a potential fix which allows application with real-time scheduling 595 // policy work. However, decision about the fix is not made yet, so it is 596 // disabled by default. 597 { // Are program started with real-time scheduling policy? 598 int sched = sched_getscheduler(0); 599 if (sched == SCHED_FIFO || sched == SCHED_RR) { 600 // Yes, we are a part of real-time application. Try to increase the 601 // priority of the monitor. 602 struct sched_param param; 603 int max_priority = sched_get_priority_max(sched); 604 int rc; 605 KMP_WARNING(RealTimeSchedNotSupported); 606 sched_getparam(0, ¶m); 607 if (param.sched_priority < max_priority) { 608 param.sched_priority += 1; 609 rc = sched_setscheduler(0, sched, ¶m); 610 if (rc != 0) { 611 int error = errno; 612 kmp_msg_t err_code = KMP_ERR(error); 613 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority), 614 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null); 615 if (__kmp_generate_warnings == kmp_warnings_off) { 616 __kmp_str_free(&err_code.str); 617 } 618 } 619 } else { 620 // We cannot abort here, because number of CPUs may be enough for all 621 // the threads, including the monitor thread, so application could 622 // potentially work... 623 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority), 624 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority), 625 __kmp_msg_null); 626 } 627 } 628 // AC: free thread that waits for monitor started 629 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 630 } 631 #endif // KMP_REAL_TIME_FIX 632 633 KMP_MB(); /* Flush all pending memory write invalidates. */ 634 635 if (__kmp_monitor_wakeups == 1) { 636 interval.tv_sec = 1; 637 interval.tv_nsec = 0; 638 } else { 639 interval.tv_sec = 0; 640 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups); 641 } 642 643 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n")); 644 645 while (!TCR_4(__kmp_global.g.g_done)) { 646 struct timespec now; 647 struct timeval tval; 648 649 /* This thread monitors the state of the system */ 650 651 KA_TRACE(15, ("__kmp_launch_monitor: update\n")); 652 653 status = gettimeofday(&tval, NULL); 654 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 655 TIMEVAL_TO_TIMESPEC(&tval, &now); 656 657 now.tv_sec += interval.tv_sec; 658 now.tv_nsec += interval.tv_nsec; 659 660 if (now.tv_nsec >= KMP_NSEC_PER_SEC) { 661 now.tv_sec += 1; 662 now.tv_nsec -= KMP_NSEC_PER_SEC; 663 } 664 665 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 666 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 667 // AC: the monitor should not fall asleep if g_done has been set 668 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex 669 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond, 670 &__kmp_wait_mx.m_mutex, &now); 671 if (status != 0) { 672 if (status != ETIMEDOUT && status != EINTR) { 673 KMP_SYSFAIL("pthread_cond_timedwait", status); 674 } 675 } 676 } 677 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 678 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 679 680 TCW_4(__kmp_global.g.g_time.dt.t_value, 681 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1); 682 683 KMP_MB(); /* Flush all pending memory write invalidates. */ 684 } 685 686 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n")); 687 688 #ifdef KMP_BLOCK_SIGNALS 689 status = sigfillset(&new_set); 690 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 691 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL); 692 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 693 #endif /* KMP_BLOCK_SIGNALS */ 694 695 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n")); 696 697 if (__kmp_global.g.g_abort != 0) { 698 /* now we need to terminate the worker threads */ 699 /* the value of t_abort is the signal we caught */ 700 701 int gtid; 702 703 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n", 704 __kmp_global.g.g_abort)); 705 706 /* terminate the OpenMP worker threads */ 707 /* TODO this is not valid for sibling threads!! 708 * the uber master might not be 0 anymore.. */ 709 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid) 710 __kmp_terminate_thread(gtid); 711 712 __kmp_cleanup(); 713 714 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n", 715 __kmp_global.g.g_abort)); 716 717 if (__kmp_global.g.g_abort > 0) 718 raise(__kmp_global.g.g_abort); 719 } 720 721 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n")); 722 723 return thr; 724 } 725 #endif // KMP_USE_MONITOR 726 727 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) { 728 pthread_t handle; 729 pthread_attr_t thread_attr; 730 int status; 731 732 th->th.th_info.ds.ds_gtid = gtid; 733 734 #if KMP_STATS_ENABLED 735 // sets up worker thread stats 736 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid); 737 738 // th->th.th_stats is used to transfer thread-specific stats-pointer to 739 // __kmp_launch_worker. So when thread is created (goes into 740 // __kmp_launch_worker) it will set its thread local pointer to 741 // th->th.th_stats 742 if (!KMP_UBER_GTID(gtid)) { 743 th->th.th_stats = __kmp_stats_list->push_back(gtid); 744 } else { 745 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(), 746 // so set the th->th.th_stats field to it. 747 th->th.th_stats = __kmp_stats_thread_ptr; 748 } 749 __kmp_release_tas_lock(&__kmp_stats_lock, gtid); 750 751 #endif // KMP_STATS_ENABLED 752 753 if (KMP_UBER_GTID(gtid)) { 754 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid)); 755 th->th.th_info.ds.ds_thread = pthread_self(); 756 __kmp_set_stack_info(gtid, th); 757 __kmp_check_stack_overlap(th); 758 return; 759 } 760 761 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid)); 762 763 KMP_MB(); /* Flush all pending memory write invalidates. */ 764 765 #ifdef KMP_THREAD_ATTR 766 status = pthread_attr_init(&thread_attr); 767 if (status != 0) { 768 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 769 } 770 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 771 if (status != 0) { 772 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null); 773 } 774 775 /* Set stack size for this thread now. 776 The multiple of 2 is there because on some machines, requesting an unusual 777 stacksize causes the thread to have an offset before the dummy alloca() 778 takes place to create the offset. Since we want the user to have a 779 sufficient stacksize AND support a stack offset, we alloca() twice the 780 offset so that the upcoming alloca() does not eliminate any premade offset, 781 and also gives the user the stack space they requested for all threads */ 782 stack_size += gtid * __kmp_stkoffset * 2; 783 784 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 785 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n", 786 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 787 788 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 789 status = pthread_attr_setstacksize(&thread_attr, stack_size); 790 #ifdef KMP_BACKUP_STKSIZE 791 if (status != 0) { 792 if (!__kmp_env_stksize) { 793 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset; 794 __kmp_stksize = KMP_BACKUP_STKSIZE; 795 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 796 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu " 797 "bytes\n", 798 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 799 status = pthread_attr_setstacksize(&thread_attr, stack_size); 800 } 801 } 802 #endif /* KMP_BACKUP_STKSIZE */ 803 if (status != 0) { 804 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 805 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null); 806 } 807 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 808 809 #endif /* KMP_THREAD_ATTR */ 810 811 status = 812 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th); 813 if (status != 0 || !handle) { // ??? Why do we check handle?? 814 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 815 if (status == EINVAL) { 816 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 817 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null); 818 } 819 if (status == ENOMEM) { 820 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 821 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null); 822 } 823 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 824 if (status == EAGAIN) { 825 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status), 826 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null); 827 } 828 KMP_SYSFAIL("pthread_create", status); 829 } 830 831 th->th.th_info.ds.ds_thread = handle; 832 833 #ifdef KMP_THREAD_ATTR 834 status = pthread_attr_destroy(&thread_attr); 835 if (status) { 836 kmp_msg_t err_code = KMP_ERR(status); 837 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 838 __kmp_msg_null); 839 if (__kmp_generate_warnings == kmp_warnings_off) { 840 __kmp_str_free(&err_code.str); 841 } 842 } 843 #endif /* KMP_THREAD_ATTR */ 844 845 KMP_MB(); /* Flush all pending memory write invalidates. */ 846 847 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid)); 848 849 } // __kmp_create_worker 850 851 #if KMP_USE_MONITOR 852 void __kmp_create_monitor(kmp_info_t *th) { 853 pthread_t handle; 854 pthread_attr_t thread_attr; 855 size_t size; 856 int status; 857 int auto_adj_size = FALSE; 858 859 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { 860 // We don't need monitor thread in case of MAX_BLOCKTIME 861 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of " 862 "MAX blocktime\n")); 863 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op 864 th->th.th_info.ds.ds_gtid = 0; 865 return; 866 } 867 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n")); 868 869 KMP_MB(); /* Flush all pending memory write invalidates. */ 870 871 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR; 872 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR; 873 #if KMP_REAL_TIME_FIX 874 TCW_4(__kmp_global.g.g_time.dt.t_value, 875 -1); // Will use it for synchronization a bit later. 876 #else 877 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 878 #endif // KMP_REAL_TIME_FIX 879 880 #ifdef KMP_THREAD_ATTR 881 if (__kmp_monitor_stksize == 0) { 882 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 883 auto_adj_size = TRUE; 884 } 885 status = pthread_attr_init(&thread_attr); 886 if (status != 0) { 887 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 888 } 889 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 890 if (status != 0) { 891 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null); 892 } 893 894 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 895 status = pthread_attr_getstacksize(&thread_attr, &size); 896 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status); 897 #else 898 size = __kmp_sys_min_stksize; 899 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 900 #endif /* KMP_THREAD_ATTR */ 901 902 if (__kmp_monitor_stksize == 0) { 903 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 904 } 905 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) { 906 __kmp_monitor_stksize = __kmp_sys_min_stksize; 907 } 908 909 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes," 910 "requested stacksize = %lu bytes\n", 911 size, __kmp_monitor_stksize)); 912 913 retry: 914 915 /* Set stack size for this thread now. */ 916 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 917 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,", 918 __kmp_monitor_stksize)); 919 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize); 920 if (status != 0) { 921 if (auto_adj_size) { 922 __kmp_monitor_stksize *= 2; 923 goto retry; 924 } 925 kmp_msg_t err_code = KMP_ERR(status); 926 __kmp_msg(kmp_ms_warning, // should this be fatal? BB 927 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize), 928 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null); 929 if (__kmp_generate_warnings == kmp_warnings_off) { 930 __kmp_str_free(&err_code.str); 931 } 932 } 933 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 934 935 status = 936 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th); 937 938 if (status != 0) { 939 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 940 if (status == EINVAL) { 941 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) { 942 __kmp_monitor_stksize *= 2; 943 goto retry; 944 } 945 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 946 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize), 947 __kmp_msg_null); 948 } 949 if (status == ENOMEM) { 950 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 951 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize), 952 __kmp_msg_null); 953 } 954 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 955 if (status == EAGAIN) { 956 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status), 957 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null); 958 } 959 KMP_SYSFAIL("pthread_create", status); 960 } 961 962 th->th.th_info.ds.ds_thread = handle; 963 964 #if KMP_REAL_TIME_FIX 965 // Wait for the monitor thread is really started and set its *priority*. 966 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 967 sizeof(__kmp_global.g.g_time.dt.t_value)); 968 __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1, 969 &__kmp_neq_4, NULL); 970 #endif // KMP_REAL_TIME_FIX 971 972 #ifdef KMP_THREAD_ATTR 973 status = pthread_attr_destroy(&thread_attr); 974 if (status != 0) { 975 kmp_msg_t err_code = KMP_ERR(status); 976 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 977 __kmp_msg_null); 978 if (__kmp_generate_warnings == kmp_warnings_off) { 979 __kmp_str_free(&err_code.str); 980 } 981 } 982 #endif 983 984 KMP_MB(); /* Flush all pending memory write invalidates. */ 985 986 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n", 987 th->th.th_info.ds.ds_thread)); 988 989 } // __kmp_create_monitor 990 #endif // KMP_USE_MONITOR 991 992 void __kmp_exit_thread(int exit_status) { 993 #if KMP_OS_WASI 994 // TODO: the wasm32-wasi-threads target does not yet support pthread_exit. 995 #else 996 pthread_exit((void *)(intptr_t)exit_status); 997 #endif 998 } // __kmp_exit_thread 999 1000 #if KMP_USE_MONITOR 1001 void __kmp_resume_monitor(); 1002 1003 extern "C" void __kmp_reap_monitor(kmp_info_t *th) { 1004 int status; 1005 void *exit_val; 1006 1007 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle" 1008 " %#.8lx\n", 1009 th->th.th_info.ds.ds_thread)); 1010 1011 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR. 1012 // If both tid and gtid are 0, it means the monitor did not ever start. 1013 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down. 1014 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid); 1015 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) { 1016 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n")); 1017 return; 1018 } 1019 1020 KMP_MB(); /* Flush all pending memory write invalidates. */ 1021 1022 /* First, check to see whether the monitor thread exists to wake it up. This 1023 is to avoid performance problem when the monitor sleeps during 1024 blocktime-size interval */ 1025 1026 status = pthread_kill(th->th.th_info.ds.ds_thread, 0); 1027 if (status != ESRCH) { 1028 __kmp_resume_monitor(); // Wake up the monitor thread 1029 } 1030 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n")); 1031 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1032 if (exit_val != th) { 1033 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null); 1034 } 1035 1036 th->th.th_info.ds.ds_tid = KMP_GTID_DNE; 1037 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE; 1038 1039 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle" 1040 " %#.8lx\n", 1041 th->th.th_info.ds.ds_thread)); 1042 1043 KMP_MB(); /* Flush all pending memory write invalidates. */ 1044 } 1045 #else 1046 // Empty symbol to export (see exports_so.txt) when 1047 // monitor thread feature is disabled 1048 extern "C" void __kmp_reap_monitor(kmp_info_t *th) { (void)th; } 1049 #endif // KMP_USE_MONITOR 1050 1051 void __kmp_reap_worker(kmp_info_t *th) { 1052 int status; 1053 void *exit_val; 1054 1055 KMP_MB(); /* Flush all pending memory write invalidates. */ 1056 1057 KA_TRACE( 1058 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid)); 1059 1060 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1061 #ifdef KMP_DEBUG 1062 /* Don't expose these to the user until we understand when they trigger */ 1063 if (status != 0) { 1064 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null); 1065 } 1066 if (exit_val != th) { 1067 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, " 1068 "exit_val = %p\n", 1069 th->th.th_info.ds.ds_gtid, exit_val)); 1070 } 1071 #else 1072 (void)status; // unused variable 1073 #endif /* KMP_DEBUG */ 1074 1075 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n", 1076 th->th.th_info.ds.ds_gtid)); 1077 1078 KMP_MB(); /* Flush all pending memory write invalidates. */ 1079 } 1080 1081 #if KMP_HANDLE_SIGNALS 1082 1083 static void __kmp_null_handler(int signo) { 1084 // Do nothing, for doing SIG_IGN-type actions. 1085 } // __kmp_null_handler 1086 1087 static void __kmp_team_handler(int signo) { 1088 if (__kmp_global.g.g_abort == 0) { 1089 /* Stage 1 signal handler, let's shut down all of the threads */ 1090 #ifdef KMP_DEBUG 1091 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo); 1092 #endif 1093 switch (signo) { 1094 case SIGHUP: 1095 case SIGINT: 1096 case SIGQUIT: 1097 case SIGILL: 1098 case SIGABRT: 1099 case SIGFPE: 1100 case SIGBUS: 1101 case SIGSEGV: 1102 #ifdef SIGSYS 1103 case SIGSYS: 1104 #endif 1105 case SIGTERM: 1106 if (__kmp_debug_buf) { 1107 __kmp_dump_debug_buffer(); 1108 } 1109 __kmp_unregister_library(); // cleanup shared memory 1110 KMP_MB(); // Flush all pending memory write invalidates. 1111 TCW_4(__kmp_global.g.g_abort, signo); 1112 KMP_MB(); // Flush all pending memory write invalidates. 1113 TCW_4(__kmp_global.g.g_done, TRUE); 1114 KMP_MB(); // Flush all pending memory write invalidates. 1115 break; 1116 default: 1117 #ifdef KMP_DEBUG 1118 __kmp_debug_printf("__kmp_team_handler: unknown signal type"); 1119 #endif 1120 break; 1121 } 1122 } 1123 } // __kmp_team_handler 1124 1125 static void __kmp_sigaction(int signum, const struct sigaction *act, 1126 struct sigaction *oldact) { 1127 int rc = sigaction(signum, act, oldact); 1128 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc); 1129 } 1130 1131 static void __kmp_install_one_handler(int sig, sig_func_t handler_func, 1132 int parallel_init) { 1133 KMP_MB(); // Flush all pending memory write invalidates. 1134 KB_TRACE(60, 1135 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init)); 1136 if (parallel_init) { 1137 struct sigaction new_action; 1138 struct sigaction old_action; 1139 new_action.sa_handler = handler_func; 1140 new_action.sa_flags = 0; 1141 sigfillset(&new_action.sa_mask); 1142 __kmp_sigaction(sig, &new_action, &old_action); 1143 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) { 1144 sigaddset(&__kmp_sigset, sig); 1145 } else { 1146 // Restore/keep user's handler if one previously installed. 1147 __kmp_sigaction(sig, &old_action, NULL); 1148 } 1149 } else { 1150 // Save initial/system signal handlers to see if user handlers installed. 1151 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]); 1152 } 1153 KMP_MB(); // Flush all pending memory write invalidates. 1154 } // __kmp_install_one_handler 1155 1156 static void __kmp_remove_one_handler(int sig) { 1157 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig)); 1158 if (sigismember(&__kmp_sigset, sig)) { 1159 struct sigaction old; 1160 KMP_MB(); // Flush all pending memory write invalidates. 1161 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old); 1162 if ((old.sa_handler != __kmp_team_handler) && 1163 (old.sa_handler != __kmp_null_handler)) { 1164 // Restore the users signal handler. 1165 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, " 1166 "restoring: sig=%d\n", 1167 sig)); 1168 __kmp_sigaction(sig, &old, NULL); 1169 } 1170 sigdelset(&__kmp_sigset, sig); 1171 KMP_MB(); // Flush all pending memory write invalidates. 1172 } 1173 } // __kmp_remove_one_handler 1174 1175 void __kmp_install_signals(int parallel_init) { 1176 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init)); 1177 if (__kmp_handle_signals || !parallel_init) { 1178 // If ! parallel_init, we do not install handlers, just save original 1179 // handlers. Let us do it even __handle_signals is 0. 1180 sigemptyset(&__kmp_sigset); 1181 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init); 1182 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init); 1183 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init); 1184 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init); 1185 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init); 1186 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init); 1187 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init); 1188 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init); 1189 #ifdef SIGSYS 1190 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init); 1191 #endif // SIGSYS 1192 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init); 1193 #ifdef SIGPIPE 1194 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init); 1195 #endif // SIGPIPE 1196 } 1197 } // __kmp_install_signals 1198 1199 void __kmp_remove_signals(void) { 1200 int sig; 1201 KB_TRACE(10, ("__kmp_remove_signals()\n")); 1202 for (sig = 1; sig < NSIG; ++sig) { 1203 __kmp_remove_one_handler(sig); 1204 } 1205 } // __kmp_remove_signals 1206 1207 #endif // KMP_HANDLE_SIGNALS 1208 1209 void __kmp_enable(int new_state) { 1210 #ifdef KMP_CANCEL_THREADS 1211 int status, old_state; 1212 status = pthread_setcancelstate(new_state, &old_state); 1213 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1214 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE); 1215 #endif 1216 } 1217 1218 void __kmp_disable(int *old_state) { 1219 #ifdef KMP_CANCEL_THREADS 1220 int status; 1221 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state); 1222 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1223 #endif 1224 } 1225 1226 static void __kmp_atfork_prepare(void) { 1227 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 1228 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1229 } 1230 1231 static void __kmp_atfork_parent(void) { 1232 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1233 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1234 } 1235 1236 /* Reset the library so execution in the child starts "all over again" with 1237 clean data structures in initial states. Don't worry about freeing memory 1238 allocated by parent, just abandon it to be safe. */ 1239 static void __kmp_atfork_child(void) { 1240 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1241 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1242 /* TODO make sure this is done right for nested/sibling */ 1243 // ATT: Memory leaks are here? TODO: Check it and fix. 1244 /* KMP_ASSERT( 0 ); */ 1245 1246 ++__kmp_fork_count; 1247 1248 #if KMP_AFFINITY_SUPPORTED 1249 #if KMP_OS_LINUX || KMP_OS_FREEBSD 1250 // reset the affinity in the child to the initial thread 1251 // affinity in the parent 1252 kmp_set_thread_affinity_mask_initial(); 1253 #endif 1254 // Set default not to bind threads tightly in the child (we're expecting 1255 // over-subscription after the fork and this can improve things for 1256 // scripting languages that use OpenMP inside process-parallel code). 1257 if (__kmp_nested_proc_bind.bind_types != NULL) { 1258 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false; 1259 } 1260 for (kmp_affinity_t *affinity : __kmp_affinities) 1261 *affinity = KMP_AFFINITY_INIT(affinity->env_var); 1262 __kmp_affin_fullMask = nullptr; 1263 __kmp_affin_origMask = nullptr; 1264 __kmp_topology = nullptr; 1265 #endif // KMP_AFFINITY_SUPPORTED 1266 1267 #if KMP_USE_MONITOR 1268 __kmp_init_monitor = 0; 1269 #endif 1270 __kmp_init_parallel = FALSE; 1271 __kmp_init_middle = FALSE; 1272 __kmp_init_serial = FALSE; 1273 TCW_4(__kmp_init_gtid, FALSE); 1274 __kmp_init_common = FALSE; 1275 1276 TCW_4(__kmp_init_user_locks, FALSE); 1277 #if !KMP_USE_DYNAMIC_LOCK 1278 __kmp_user_lock_table.used = 1; 1279 __kmp_user_lock_table.allocated = 0; 1280 __kmp_user_lock_table.table = NULL; 1281 __kmp_lock_blocks = NULL; 1282 #endif 1283 1284 __kmp_all_nth = 0; 1285 TCW_4(__kmp_nth, 0); 1286 1287 __kmp_thread_pool = NULL; 1288 __kmp_thread_pool_insert_pt = NULL; 1289 __kmp_team_pool = NULL; 1290 1291 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate 1292 here so threadprivate doesn't use stale data */ 1293 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n", 1294 __kmp_threadpriv_cache_list)); 1295 1296 while (__kmp_threadpriv_cache_list != NULL) { 1297 1298 if (*__kmp_threadpriv_cache_list->addr != NULL) { 1299 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n", 1300 &(*__kmp_threadpriv_cache_list->addr))); 1301 1302 *__kmp_threadpriv_cache_list->addr = NULL; 1303 } 1304 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next; 1305 } 1306 1307 __kmp_init_runtime = FALSE; 1308 1309 /* reset statically initialized locks */ 1310 __kmp_init_bootstrap_lock(&__kmp_initz_lock); 1311 __kmp_init_bootstrap_lock(&__kmp_stdio_lock); 1312 __kmp_init_bootstrap_lock(&__kmp_console_lock); 1313 __kmp_init_bootstrap_lock(&__kmp_task_team_lock); 1314 1315 #if USE_ITT_BUILD 1316 __kmp_itt_reset(); // reset ITT's global state 1317 #endif /* USE_ITT_BUILD */ 1318 1319 { 1320 // Child process often get terminated without any use of OpenMP. That might 1321 // cause mapped shared memory file to be left unattended. Thus we postpone 1322 // library registration till middle initialization in the child process. 1323 __kmp_need_register_serial = FALSE; 1324 __kmp_serial_initialize(); 1325 } 1326 1327 /* This is necessary to make sure no stale data is left around */ 1328 /* AC: customers complain that we use unsafe routines in the atfork 1329 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen 1330 in dynamic_link when check the presence of shared tbbmalloc library. 1331 Suggestion is to make the library initialization lazier, similar 1332 to what done for __kmpc_begin(). */ 1333 // TODO: synchronize all static initializations with regular library 1334 // startup; look at kmp_global.cpp and etc. 1335 //__kmp_internal_begin (); 1336 } 1337 1338 void __kmp_register_atfork(void) { 1339 if (__kmp_need_register_atfork) { 1340 #if !KMP_OS_WASI 1341 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent, 1342 __kmp_atfork_child); 1343 KMP_CHECK_SYSFAIL("pthread_atfork", status); 1344 #endif 1345 __kmp_need_register_atfork = FALSE; 1346 } 1347 } 1348 1349 void __kmp_suspend_initialize(void) { 1350 int status; 1351 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr); 1352 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1353 status = pthread_condattr_init(&__kmp_suspend_cond_attr); 1354 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1355 } 1356 1357 void __kmp_suspend_initialize_thread(kmp_info_t *th) { 1358 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count); 1359 int new_value = __kmp_fork_count + 1; 1360 // Return if already initialized 1361 if (old_value == new_value) 1362 return; 1363 // Wait, then return if being initialized 1364 if (old_value == -1 || !__kmp_atomic_compare_store( 1365 &th->th.th_suspend_init_count, old_value, -1)) { 1366 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) { 1367 KMP_CPU_PAUSE(); 1368 } 1369 } else { 1370 // Claim to be the initializer and do initializations 1371 int status; 1372 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond, 1373 &__kmp_suspend_cond_attr); 1374 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1375 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex, 1376 &__kmp_suspend_mutex_attr); 1377 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1378 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value); 1379 } 1380 } 1381 1382 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { 1383 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) { 1384 /* this means we have initialize the suspension pthread objects for this 1385 thread in this instance of the process */ 1386 int status; 1387 1388 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond); 1389 if (status != 0 && status != EBUSY) { 1390 KMP_SYSFAIL("pthread_cond_destroy", status); 1391 } 1392 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex); 1393 if (status != 0 && status != EBUSY) { 1394 KMP_SYSFAIL("pthread_mutex_destroy", status); 1395 } 1396 --th->th.th_suspend_init_count; 1397 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) == 1398 __kmp_fork_count); 1399 } 1400 } 1401 1402 // return true if lock obtained, false otherwise 1403 int __kmp_try_suspend_mx(kmp_info_t *th) { 1404 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0); 1405 } 1406 1407 void __kmp_lock_suspend_mx(kmp_info_t *th) { 1408 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1409 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1410 } 1411 1412 void __kmp_unlock_suspend_mx(kmp_info_t *th) { 1413 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1414 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1415 } 1416 1417 /* This routine puts the calling thread to sleep after setting the 1418 sleep bit for the indicated flag variable to true. */ 1419 template <class C> 1420 static inline void __kmp_suspend_template(int th_gtid, C *flag) { 1421 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend); 1422 kmp_info_t *th = __kmp_threads[th_gtid]; 1423 int status; 1424 typename C::flag_t old_spin; 1425 1426 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid, 1427 flag->get())); 1428 1429 __kmp_suspend_initialize_thread(th); 1430 1431 __kmp_lock_suspend_mx(th); 1432 1433 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n", 1434 th_gtid, flag->get())); 1435 1436 /* TODO: shouldn't this use release semantics to ensure that 1437 __kmp_suspend_initialize_thread gets called first? */ 1438 old_spin = flag->set_sleeping(); 1439 TCW_PTR(th->th.th_sleep_loc, (void *)flag); 1440 th->th.th_sleep_loc_type = flag->get_type(); 1441 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && 1442 __kmp_pause_status != kmp_soft_paused) { 1443 flag->unset_sleeping(); 1444 TCW_PTR(th->th.th_sleep_loc, NULL); 1445 th->th.th_sleep_loc_type = flag_unset; 1446 __kmp_unlock_suspend_mx(th); 1447 return; 1448 } 1449 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1450 " was %x\n", 1451 th_gtid, flag->get(), flag->load(), old_spin)); 1452 1453 if (flag->done_check_val(old_spin) || flag->done_check()) { 1454 flag->unset_sleeping(); 1455 TCW_PTR(th->th.th_sleep_loc, NULL); 1456 th->th.th_sleep_loc_type = flag_unset; 1457 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1458 "for spin(%p)\n", 1459 th_gtid, flag->get())); 1460 } else { 1461 /* Encapsulate in a loop as the documentation states that this may 1462 "with low probability" return when the condition variable has 1463 not been signaled or broadcast */ 1464 int deactivated = FALSE; 1465 1466 while (flag->is_sleeping()) { 1467 #ifdef DEBUG_SUSPEND 1468 char buffer[128]; 1469 __kmp_suspend_count++; 1470 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1471 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid, 1472 buffer); 1473 #endif 1474 // Mark the thread as no longer active (only in the first iteration of the 1475 // loop). 1476 if (!deactivated) { 1477 th->th.th_active = FALSE; 1478 if (th->th.th_active_in_pool) { 1479 th->th.th_active_in_pool = FALSE; 1480 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); 1481 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 1482 } 1483 deactivated = TRUE; 1484 } 1485 1486 KMP_DEBUG_ASSERT(th->th.th_sleep_loc); 1487 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type); 1488 1489 #if USE_SUSPEND_TIMEOUT 1490 struct timespec now; 1491 struct timeval tval; 1492 int msecs; 1493 1494 status = gettimeofday(&tval, NULL); 1495 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1496 TIMEVAL_TO_TIMESPEC(&tval, &now); 1497 1498 msecs = (4 * __kmp_dflt_blocktime) + 200; 1499 now.tv_sec += msecs / 1000; 1500 now.tv_nsec += (msecs % 1000) * 1000; 1501 1502 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " 1503 "pthread_cond_timedwait\n", 1504 th_gtid)); 1505 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond, 1506 &th->th.th_suspend_mx.m_mutex, &now); 1507 #else 1508 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform" 1509 " pthread_cond_wait\n", 1510 th_gtid)); 1511 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond, 1512 &th->th.th_suspend_mx.m_mutex); 1513 #endif // USE_SUSPEND_TIMEOUT 1514 1515 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) { 1516 KMP_SYSFAIL("pthread_cond_wait", status); 1517 } 1518 1519 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type()); 1520 1521 if (!flag->is_sleeping() && 1522 ((status == EINTR) || (status == ETIMEDOUT))) { 1523 // if interrupt or timeout, and thread is no longer sleeping, we need to 1524 // make sure sleep_loc gets reset; however, this shouldn't be needed if 1525 // we woke up with resume 1526 flag->unset_sleeping(); 1527 TCW_PTR(th->th.th_sleep_loc, NULL); 1528 th->th.th_sleep_loc_type = flag_unset; 1529 } 1530 #ifdef KMP_DEBUG 1531 if (status == ETIMEDOUT) { 1532 if (flag->is_sleeping()) { 1533 KF_TRACE(100, 1534 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid)); 1535 } else { 1536 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1537 "not set!\n", 1538 th_gtid)); 1539 TCW_PTR(th->th.th_sleep_loc, NULL); 1540 th->th.th_sleep_loc_type = flag_unset; 1541 } 1542 } else if (flag->is_sleeping()) { 1543 KF_TRACE(100, 1544 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid)); 1545 } 1546 #endif 1547 } // while 1548 1549 // Mark the thread as active again (if it was previous marked as inactive) 1550 if (deactivated) { 1551 th->th.th_active = TRUE; 1552 if (TCR_4(th->th.th_in_pool)) { 1553 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); 1554 th->th.th_active_in_pool = TRUE; 1555 } 1556 } 1557 } 1558 // We may have had the loop variable set before entering the loop body; 1559 // so we need to reset sleep_loc. 1560 TCW_PTR(th->th.th_sleep_loc, NULL); 1561 th->th.th_sleep_loc_type = flag_unset; 1562 1563 KMP_DEBUG_ASSERT(!flag->is_sleeping()); 1564 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc); 1565 #ifdef DEBUG_SUSPEND 1566 { 1567 char buffer[128]; 1568 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1569 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid, 1570 buffer); 1571 } 1572 #endif 1573 1574 __kmp_unlock_suspend_mx(th); 1575 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid)); 1576 } 1577 1578 template <bool C, bool S> 1579 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) { 1580 __kmp_suspend_template(th_gtid, flag); 1581 } 1582 template <bool C, bool S> 1583 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) { 1584 __kmp_suspend_template(th_gtid, flag); 1585 } 1586 template <bool C, bool S> 1587 void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) { 1588 __kmp_suspend_template(th_gtid, flag); 1589 } 1590 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { 1591 __kmp_suspend_template(th_gtid, flag); 1592 } 1593 1594 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *); 1595 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *); 1596 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *); 1597 template void 1598 __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1599 template void 1600 __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *); 1601 1602 /* This routine signals the thread specified by target_gtid to wake up 1603 after setting the sleep bit indicated by the flag argument to FALSE. 1604 The target thread must already have called __kmp_suspend_template() */ 1605 template <class C> 1606 static inline void __kmp_resume_template(int target_gtid, C *flag) { 1607 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1608 kmp_info_t *th = __kmp_threads[target_gtid]; 1609 int status; 1610 1611 #ifdef KMP_DEBUG 1612 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1613 #endif 1614 1615 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n", 1616 gtid, target_gtid)); 1617 KMP_DEBUG_ASSERT(gtid != target_gtid); 1618 1619 __kmp_suspend_initialize_thread(th); 1620 1621 __kmp_lock_suspend_mx(th); 1622 1623 if (!flag || flag != th->th.th_sleep_loc) { 1624 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a 1625 // different location; wake up at new location 1626 flag = (C *)CCAST(void *, th->th.th_sleep_loc); 1627 } 1628 1629 // First, check if the flag is null or its type has changed. If so, someone 1630 // else woke it up. 1631 if (!flag) { // Thread doesn't appear to be sleeping on anything 1632 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1633 "awake: flag(%p)\n", 1634 gtid, target_gtid, (void *)NULL)); 1635 __kmp_unlock_suspend_mx(th); 1636 return; 1637 } else if (flag->get_type() != th->th.th_sleep_loc_type) { 1638 // Flag type does not appear to match this function template; possibly the 1639 // thread is sleeping on something else. Try null resume again. 1640 KF_TRACE( 1641 5, 1642 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), " 1643 "spin(%p) type=%d ptr_type=%d\n", 1644 gtid, target_gtid, flag, flag->get(), flag->get_type(), 1645 th->th.th_sleep_loc_type)); 1646 __kmp_unlock_suspend_mx(th); 1647 __kmp_null_resume_wrapper(th); 1648 return; 1649 } else { // if multiple threads are sleeping, flag should be internally 1650 // referring to a specific thread here 1651 if (!flag->is_sleeping()) { 1652 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1653 "awake: flag(%p): %u\n", 1654 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1655 __kmp_unlock_suspend_mx(th); 1656 return; 1657 } 1658 } 1659 KMP_DEBUG_ASSERT(flag); 1660 flag->unset_sleeping(); 1661 TCW_PTR(th->th.th_sleep_loc, NULL); 1662 th->th.th_sleep_loc_type = flag_unset; 1663 1664 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1665 "sleep bit for flag's loc(%p): %u\n", 1666 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1667 1668 #ifdef DEBUG_SUSPEND 1669 { 1670 char buffer[128]; 1671 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1672 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid, 1673 target_gtid, buffer); 1674 } 1675 #endif 1676 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond); 1677 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1678 __kmp_unlock_suspend_mx(th); 1679 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" 1680 " for T#%d\n", 1681 gtid, target_gtid)); 1682 } 1683 1684 template <bool C, bool S> 1685 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) { 1686 __kmp_resume_template(target_gtid, flag); 1687 } 1688 template <bool C, bool S> 1689 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) { 1690 __kmp_resume_template(target_gtid, flag); 1691 } 1692 template <bool C, bool S> 1693 void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) { 1694 __kmp_resume_template(target_gtid, flag); 1695 } 1696 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { 1697 __kmp_resume_template(target_gtid, flag); 1698 } 1699 1700 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *); 1701 template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *); 1702 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *); 1703 template void 1704 __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1705 1706 #if KMP_USE_MONITOR 1707 void __kmp_resume_monitor() { 1708 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1709 int status; 1710 #ifdef KMP_DEBUG 1711 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1712 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid, 1713 KMP_GTID_MONITOR)); 1714 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR); 1715 #endif 1716 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 1717 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1718 #ifdef DEBUG_SUSPEND 1719 { 1720 char buffer[128]; 1721 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond); 1722 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid, 1723 KMP_GTID_MONITOR, buffer); 1724 } 1725 #endif 1726 status = pthread_cond_signal(&__kmp_wait_cv.c_cond); 1727 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1728 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 1729 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1730 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up" 1731 " for T#%d\n", 1732 gtid, KMP_GTID_MONITOR)); 1733 } 1734 #endif // KMP_USE_MONITOR 1735 1736 void __kmp_yield() { sched_yield(); } 1737 1738 void __kmp_gtid_set_specific(int gtid) { 1739 if (__kmp_init_gtid) { 1740 int status; 1741 status = pthread_setspecific(__kmp_gtid_threadprivate_key, 1742 (void *)(intptr_t)(gtid + 1)); 1743 KMP_CHECK_SYSFAIL("pthread_setspecific", status); 1744 } else { 1745 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n")); 1746 } 1747 } 1748 1749 int __kmp_gtid_get_specific() { 1750 int gtid; 1751 if (!__kmp_init_gtid) { 1752 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " 1753 "KMP_GTID_SHUTDOWN\n")); 1754 return KMP_GTID_SHUTDOWN; 1755 } 1756 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key); 1757 if (gtid == 0) { 1758 gtid = KMP_GTID_DNE; 1759 } else { 1760 gtid--; 1761 } 1762 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n", 1763 __kmp_gtid_threadprivate_key, gtid)); 1764 return gtid; 1765 } 1766 1767 double __kmp_read_cpu_time(void) { 1768 /*clock_t t;*/ 1769 struct tms buffer; 1770 1771 /*t =*/times(&buffer); 1772 1773 return (double)(buffer.tms_utime + buffer.tms_cutime) / 1774 (double)CLOCKS_PER_SEC; 1775 } 1776 1777 int __kmp_read_system_info(struct kmp_sys_info *info) { 1778 int status; 1779 struct rusage r_usage; 1780 1781 memset(info, 0, sizeof(*info)); 1782 1783 status = getrusage(RUSAGE_SELF, &r_usage); 1784 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status); 1785 1786 #if !KMP_OS_WASI 1787 // The maximum resident set size utilized (in kilobytes) 1788 info->maxrss = r_usage.ru_maxrss; 1789 // The number of page faults serviced without any I/O 1790 info->minflt = r_usage.ru_minflt; 1791 // The number of page faults serviced that required I/O 1792 info->majflt = r_usage.ru_majflt; 1793 // The number of times a process was "swapped" out of memory 1794 info->nswap = r_usage.ru_nswap; 1795 // The number of times the file system had to perform input 1796 info->inblock = r_usage.ru_inblock; 1797 // The number of times the file system had to perform output 1798 info->oublock = r_usage.ru_oublock; 1799 // The number of times a context switch was voluntarily 1800 info->nvcsw = r_usage.ru_nvcsw; 1801 // The number of times a context switch was forced 1802 info->nivcsw = r_usage.ru_nivcsw; 1803 #endif 1804 1805 return (status != 0); 1806 } 1807 1808 void __kmp_read_system_time(double *delta) { 1809 double t_ns; 1810 struct timeval tval; 1811 struct timespec stop; 1812 int status; 1813 1814 status = gettimeofday(&tval, NULL); 1815 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1816 TIMEVAL_TO_TIMESPEC(&tval, &stop); 1817 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start)); 1818 *delta = (t_ns * 1e-9); 1819 } 1820 1821 void __kmp_clear_system_time(void) { 1822 struct timeval tval; 1823 int status; 1824 status = gettimeofday(&tval, NULL); 1825 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1826 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start); 1827 } 1828 1829 static int __kmp_get_xproc(void) { 1830 1831 int r = 0; 1832 1833 #if KMP_OS_LINUX 1834 1835 __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r)); 1836 1837 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \ 1838 KMP_OS_HURD || KMP_OS_SOLARIS || KMP_OS_WASI || KMP_OS_AIX 1839 1840 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r)); 1841 1842 #elif KMP_OS_DARWIN 1843 1844 // Bug C77011 High "OpenMP Threads and number of active cores". 1845 1846 // Find the number of available CPUs. 1847 kern_return_t rc; 1848 host_basic_info_data_t info; 1849 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT; 1850 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num); 1851 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) { 1852 // Cannot use KA_TRACE() here because this code works before trace support 1853 // is initialized. 1854 r = info.avail_cpus; 1855 } else { 1856 KMP_WARNING(CantGetNumAvailCPU); 1857 KMP_INFORM(AssumedNumCPU); 1858 } 1859 1860 #else 1861 1862 #error "Unknown or unsupported OS." 1863 1864 #endif 1865 1866 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */ 1867 1868 } // __kmp_get_xproc 1869 1870 int __kmp_read_from_file(char const *path, char const *format, ...) { 1871 int result; 1872 va_list args; 1873 1874 va_start(args, format); 1875 FILE *f = fopen(path, "rb"); 1876 if (f == NULL) { 1877 va_end(args); 1878 return 0; 1879 } 1880 result = vfscanf(f, format, args); 1881 fclose(f); 1882 va_end(args); 1883 1884 return result; 1885 } 1886 1887 void __kmp_runtime_initialize(void) { 1888 int status; 1889 pthread_mutexattr_t mutex_attr; 1890 pthread_condattr_t cond_attr; 1891 1892 if (__kmp_init_runtime) { 1893 return; 1894 } 1895 1896 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1897 if (!__kmp_cpuinfo.initialized) { 1898 __kmp_query_cpuid(&__kmp_cpuinfo); 1899 } 1900 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1901 1902 __kmp_xproc = __kmp_get_xproc(); 1903 1904 #if !KMP_32_BIT_ARCH 1905 struct rlimit rlim; 1906 // read stack size of calling thread, save it as default for worker threads; 1907 // this should be done before reading environment variables 1908 status = getrlimit(RLIMIT_STACK, &rlim); 1909 if (status == 0) { // success? 1910 __kmp_stksize = rlim.rlim_cur; 1911 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed 1912 } 1913 #endif /* KMP_32_BIT_ARCH */ 1914 1915 if (sysconf(_SC_THREADS)) { 1916 1917 /* Query the maximum number of threads */ 1918 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth)); 1919 #ifdef __ve__ 1920 if (__kmp_sys_max_nth == -1) { 1921 // VE's pthread supports only up to 64 threads per a VE process. 1922 // So we use that KMP_MAX_NTH (predefined as 64) here. 1923 __kmp_sys_max_nth = KMP_MAX_NTH; 1924 } 1925 #else 1926 if (__kmp_sys_max_nth == -1) { 1927 /* Unlimited threads for NPTL */ 1928 __kmp_sys_max_nth = INT_MAX; 1929 } else if (__kmp_sys_max_nth <= 1) { 1930 /* Can't tell, just use PTHREAD_THREADS_MAX */ 1931 __kmp_sys_max_nth = KMP_MAX_NTH; 1932 } 1933 #endif 1934 1935 /* Query the minimum stack size */ 1936 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN); 1937 if (__kmp_sys_min_stksize <= 1) { 1938 __kmp_sys_min_stksize = KMP_MIN_STKSIZE; 1939 } 1940 } 1941 1942 /* Set up minimum number of threads to switch to TLS gtid */ 1943 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; 1944 1945 status = pthread_key_create(&__kmp_gtid_threadprivate_key, 1946 __kmp_internal_end_dest); 1947 KMP_CHECK_SYSFAIL("pthread_key_create", status); 1948 status = pthread_mutexattr_init(&mutex_attr); 1949 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1950 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr); 1951 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1952 status = pthread_mutexattr_destroy(&mutex_attr); 1953 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status); 1954 status = pthread_condattr_init(&cond_attr); 1955 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1956 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr); 1957 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1958 status = pthread_condattr_destroy(&cond_attr); 1959 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status); 1960 #if USE_ITT_BUILD 1961 __kmp_itt_initialize(); 1962 #endif /* USE_ITT_BUILD */ 1963 1964 __kmp_init_runtime = TRUE; 1965 } 1966 1967 void __kmp_runtime_destroy(void) { 1968 int status; 1969 1970 if (!__kmp_init_runtime) { 1971 return; // Nothing to do. 1972 } 1973 1974 #if USE_ITT_BUILD 1975 __kmp_itt_destroy(); 1976 #endif /* USE_ITT_BUILD */ 1977 1978 status = pthread_key_delete(__kmp_gtid_threadprivate_key); 1979 KMP_CHECK_SYSFAIL("pthread_key_delete", status); 1980 1981 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex); 1982 if (status != 0 && status != EBUSY) { 1983 KMP_SYSFAIL("pthread_mutex_destroy", status); 1984 } 1985 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond); 1986 if (status != 0 && status != EBUSY) { 1987 KMP_SYSFAIL("pthread_cond_destroy", status); 1988 } 1989 #if KMP_AFFINITY_SUPPORTED 1990 __kmp_affinity_uninitialize(); 1991 #endif 1992 1993 __kmp_init_runtime = FALSE; 1994 } 1995 1996 /* Put the thread to sleep for a time period */ 1997 /* NOTE: not currently used anywhere */ 1998 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); } 1999 2000 /* Calculate the elapsed wall clock time for the user */ 2001 void __kmp_elapsed(double *t) { 2002 int status; 2003 #ifdef FIX_SGI_CLOCK 2004 struct timespec ts; 2005 2006 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); 2007 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status); 2008 *t = 2009 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec; 2010 #else 2011 struct timeval tv; 2012 2013 status = gettimeofday(&tv, NULL); 2014 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 2015 *t = 2016 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec; 2017 #endif 2018 } 2019 2020 /* Calculate the elapsed wall clock tick for the user */ 2021 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; } 2022 2023 /* Return the current time stamp in nsec */ 2024 kmp_uint64 __kmp_now_nsec() { 2025 struct timeval t; 2026 gettimeofday(&t, NULL); 2027 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec + 2028 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec; 2029 return nsec; 2030 } 2031 2032 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2033 /* Measure clock ticks per millisecond */ 2034 void __kmp_initialize_system_tick() { 2035 kmp_uint64 now, nsec2, diff; 2036 kmp_uint64 delay = 1000000; // ~450 usec on most machines. 2037 kmp_uint64 nsec = __kmp_now_nsec(); 2038 kmp_uint64 goal = __kmp_hardware_timestamp() + delay; 2039 while ((now = __kmp_hardware_timestamp()) < goal) 2040 ; 2041 nsec2 = __kmp_now_nsec(); 2042 diff = nsec2 - nsec; 2043 if (diff > 0) { 2044 double tpus = 1000.0 * (double)(delay + (now - goal)) / (double)diff; 2045 if (tpus > 0.0) { 2046 __kmp_ticks_per_msec = (kmp_uint64)(tpus * 1000.0); 2047 __kmp_ticks_per_usec = (kmp_uint64)tpus; 2048 } 2049 } 2050 } 2051 #endif 2052 2053 /* Determine whether the given address is mapped into the current address 2054 space. */ 2055 2056 int __kmp_is_address_mapped(void *addr) { 2057 2058 int found = 0; 2059 int rc; 2060 2061 #if KMP_OS_LINUX || KMP_OS_HURD 2062 2063 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the 2064 address ranges mapped into the address space. */ 2065 2066 char *name = __kmp_str_format("/proc/%d/maps", getpid()); 2067 FILE *file = NULL; 2068 2069 file = fopen(name, "r"); 2070 KMP_ASSERT(file != NULL); 2071 2072 for (;;) { 2073 2074 void *beginning = NULL; 2075 void *ending = NULL; 2076 char perms[5]; 2077 2078 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms); 2079 if (rc == EOF) { 2080 break; 2081 } 2082 KMP_ASSERT(rc == 3 && 2083 KMP_STRLEN(perms) == 4); // Make sure all fields are read. 2084 2085 // Ending address is not included in the region, but beginning is. 2086 if ((addr >= beginning) && (addr < ending)) { 2087 perms[2] = 0; // 3th and 4th character does not matter. 2088 if (strcmp(perms, "rw") == 0) { 2089 // Memory we are looking for should be readable and writable. 2090 found = 1; 2091 } 2092 break; 2093 } 2094 } 2095 2096 // Free resources. 2097 fclose(file); 2098 KMP_INTERNAL_FREE(name); 2099 #elif KMP_OS_FREEBSD 2100 char *buf; 2101 size_t lstsz; 2102 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()}; 2103 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0); 2104 if (rc < 0) 2105 return 0; 2106 // We pass from number of vm entry's semantic 2107 // to size of whole entry map list. 2108 lstsz = lstsz * 4 / 3; 2109 buf = reinterpret_cast<char *>(kmpc_malloc(lstsz)); 2110 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0); 2111 if (rc < 0) { 2112 kmpc_free(buf); 2113 return 0; 2114 } 2115 2116 char *lw = buf; 2117 char *up = buf + lstsz; 2118 2119 while (lw < up) { 2120 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw); 2121 size_t cursz = cur->kve_structsize; 2122 if (cursz == 0) 2123 break; 2124 void *start = reinterpret_cast<void *>(cur->kve_start); 2125 void *end = reinterpret_cast<void *>(cur->kve_end); 2126 // Readable/Writable addresses within current map entry 2127 if ((addr >= start) && (addr < end)) { 2128 if ((cur->kve_protection & KVME_PROT_READ) != 0 && 2129 (cur->kve_protection & KVME_PROT_WRITE) != 0) { 2130 found = 1; 2131 break; 2132 } 2133 } 2134 lw += cursz; 2135 } 2136 kmpc_free(buf); 2137 #elif KMP_OS_DRAGONFLY 2138 char err[_POSIX2_LINE_MAX]; 2139 kinfo_proc *proc; 2140 vmspace sp; 2141 vm_map *cur; 2142 vm_map_entry entry, *c; 2143 struct proc p; 2144 kvm_t *fd; 2145 uintptr_t uaddr; 2146 int num; 2147 2148 fd = kvm_openfiles(nullptr, nullptr, nullptr, O_RDONLY, err); 2149 if (!fd) { 2150 return 0; 2151 } 2152 2153 proc = kvm_getprocs(fd, KERN_PROC_PID, getpid(), &num); 2154 2155 if (kvm_read(fd, static_cast<uintptr_t>(proc->kp_paddr), &p, sizeof(p)) != 2156 sizeof(p) || 2157 kvm_read(fd, reinterpret_cast<uintptr_t>(p.p_vmspace), &sp, sizeof(sp)) != 2158 sizeof(sp)) { 2159 kvm_close(fd); 2160 return 0; 2161 } 2162 2163 (void)rc; 2164 cur = &sp.vm_map; 2165 uaddr = reinterpret_cast<uintptr_t>(addr); 2166 for (c = kvm_vm_map_entry_first(fd, cur, &entry); c; 2167 c = kvm_vm_map_entry_next(fd, c, &entry)) { 2168 if ((uaddr >= entry.ba.start) && (uaddr <= entry.ba.end)) { 2169 if ((entry.protection & VM_PROT_READ) != 0 && 2170 (entry.protection & VM_PROT_WRITE) != 0) { 2171 found = 1; 2172 break; 2173 } 2174 } 2175 } 2176 2177 kvm_close(fd); 2178 #elif KMP_OS_DARWIN 2179 2180 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory 2181 using vm interface. */ 2182 2183 int buffer; 2184 vm_size_t count; 2185 rc = vm_read_overwrite( 2186 mach_task_self(), // Task to read memory of. 2187 (vm_address_t)(addr), // Address to read from. 2188 1, // Number of bytes to be read. 2189 (vm_address_t)(&buffer), // Address of buffer to save read bytes in. 2190 &count // Address of var to save number of read bytes in. 2191 ); 2192 if (rc == 0) { 2193 // Memory successfully read. 2194 found = 1; 2195 } 2196 2197 #elif KMP_OS_NETBSD 2198 2199 int mib[5]; 2200 mib[0] = CTL_VM; 2201 mib[1] = VM_PROC; 2202 mib[2] = VM_PROC_MAP; 2203 mib[3] = getpid(); 2204 mib[4] = sizeof(struct kinfo_vmentry); 2205 2206 size_t size; 2207 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0); 2208 KMP_ASSERT(!rc); 2209 KMP_ASSERT(size); 2210 2211 size = size * 4 / 3; 2212 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size); 2213 KMP_ASSERT(kiv); 2214 2215 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0); 2216 KMP_ASSERT(!rc); 2217 KMP_ASSERT(size); 2218 2219 for (size_t i = 0; i < size; i++) { 2220 if (kiv[i].kve_start >= (uint64_t)addr && 2221 kiv[i].kve_end <= (uint64_t)addr) { 2222 found = 1; 2223 break; 2224 } 2225 } 2226 KMP_INTERNAL_FREE(kiv); 2227 #elif KMP_OS_OPENBSD 2228 2229 int mib[3]; 2230 mib[0] = CTL_KERN; 2231 mib[1] = KERN_PROC_VMMAP; 2232 mib[2] = getpid(); 2233 2234 size_t size; 2235 uint64_t end; 2236 rc = sysctl(mib, 3, NULL, &size, NULL, 0); 2237 KMP_ASSERT(!rc); 2238 KMP_ASSERT(size); 2239 end = size; 2240 2241 struct kinfo_vmentry kiv = {.kve_start = 0}; 2242 2243 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) { 2244 KMP_ASSERT(size); 2245 if (kiv.kve_end == end) 2246 break; 2247 2248 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) { 2249 found = 1; 2250 break; 2251 } 2252 kiv.kve_start += 1; 2253 } 2254 #elif KMP_OS_WASI 2255 found = (int)addr < (__builtin_wasm_memory_size(0) * PAGESIZE); 2256 #elif KMP_OS_SOLARIS || KMP_OS_AIX 2257 2258 // FIXME(Solaris, AIX): Implement this 2259 found = 1; 2260 2261 #else 2262 2263 #error "Unknown or unsupported OS" 2264 2265 #endif 2266 2267 return found; 2268 2269 } // __kmp_is_address_mapped 2270 2271 #ifdef USE_LOAD_BALANCE 2272 2273 #if KMP_OS_DARWIN || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 2274 KMP_OS_OPENBSD || KMP_OS_SOLARIS 2275 2276 // The function returns the rounded value of the system load average 2277 // during given time interval which depends on the value of 2278 // __kmp_load_balance_interval variable (default is 60 sec, other values 2279 // may be 300 sec or 900 sec). 2280 // It returns -1 in case of error. 2281 int __kmp_get_load_balance(int max) { 2282 double averages[3]; 2283 int ret_avg = 0; 2284 2285 int res = getloadavg(averages, 3); 2286 2287 // Check __kmp_load_balance_interval to determine which of averages to use. 2288 // getloadavg() may return the number of samples less than requested that is 2289 // less than 3. 2290 if (__kmp_load_balance_interval < 180 && (res >= 1)) { 2291 ret_avg = (int)averages[0]; // 1 min 2292 } else if ((__kmp_load_balance_interval >= 180 && 2293 __kmp_load_balance_interval < 600) && 2294 (res >= 2)) { 2295 ret_avg = (int)averages[1]; // 5 min 2296 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) { 2297 ret_avg = (int)averages[2]; // 15 min 2298 } else { // Error occurred 2299 return -1; 2300 } 2301 2302 return ret_avg; 2303 } 2304 2305 #else // Linux* OS 2306 2307 // The function returns number of running (not sleeping) threads, or -1 in case 2308 // of error. Error could be reported if Linux* OS kernel too old (without 2309 // "/proc" support). Counting running threads stops if max running threads 2310 // encountered. 2311 int __kmp_get_load_balance(int max) { 2312 static int permanent_error = 0; 2313 static int glb_running_threads = 0; // Saved count of the running threads for 2314 // the thread balance algorithm 2315 static double glb_call_time = 0; /* Thread balance algorithm call time */ 2316 2317 int running_threads = 0; // Number of running threads in the system. 2318 2319 DIR *proc_dir = NULL; // Handle of "/proc/" directory. 2320 struct dirent *proc_entry = NULL; 2321 2322 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path. 2323 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory. 2324 struct dirent *task_entry = NULL; 2325 int task_path_fixed_len; 2326 2327 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path. 2328 int stat_file = -1; 2329 int stat_path_fixed_len; 2330 2331 #ifdef KMP_DEBUG 2332 int total_processes = 0; // Total number of processes in system. 2333 #endif 2334 2335 double call_time = 0.0; 2336 2337 __kmp_str_buf_init(&task_path); 2338 __kmp_str_buf_init(&stat_path); 2339 2340 __kmp_elapsed(&call_time); 2341 2342 if (glb_call_time && 2343 (call_time - glb_call_time < __kmp_load_balance_interval)) { 2344 running_threads = glb_running_threads; 2345 goto finish; 2346 } 2347 2348 glb_call_time = call_time; 2349 2350 // Do not spend time on scanning "/proc/" if we have a permanent error. 2351 if (permanent_error) { 2352 running_threads = -1; 2353 goto finish; 2354 } 2355 2356 if (max <= 0) { 2357 max = INT_MAX; 2358 } 2359 2360 // Open "/proc/" directory. 2361 proc_dir = opendir("/proc"); 2362 if (proc_dir == NULL) { 2363 // Cannot open "/proc/". Probably the kernel does not support it. Return an 2364 // error now and in subsequent calls. 2365 running_threads = -1; 2366 permanent_error = 1; 2367 goto finish; 2368 } 2369 2370 // Initialize fixed part of task_path. This part will not change. 2371 __kmp_str_buf_cat(&task_path, "/proc/", 6); 2372 task_path_fixed_len = task_path.used; // Remember number of used characters. 2373 2374 proc_entry = readdir(proc_dir); 2375 while (proc_entry != NULL) { 2376 #if KMP_OS_AIX 2377 // Proc entry name starts with a digit. Assume it is a process' directory. 2378 if (isdigit(proc_entry->d_name[0])) { 2379 #else 2380 // Proc entry is a directory and name starts with a digit. Assume it is a 2381 // process' directory. 2382 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) { 2383 #endif 2384 2385 #ifdef KMP_DEBUG 2386 ++total_processes; 2387 #endif 2388 // Make sure init process is the very first in "/proc", so we can replace 2389 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes == 2390 // 1. We are going to check that total_processes == 1 => d_name == "1" is 2391 // true (where "=>" is implication). Since C++ does not have => operator, 2392 // let us replace it with its equivalent: a => b == ! a || b. 2393 KMP_DEBUG_ASSERT(total_processes != 1 || 2394 strcmp(proc_entry->d_name, "1") == 0); 2395 2396 // Construct task_path. 2397 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/". 2398 __kmp_str_buf_cat(&task_path, proc_entry->d_name, 2399 KMP_STRLEN(proc_entry->d_name)); 2400 __kmp_str_buf_cat(&task_path, "/task", 5); 2401 2402 task_dir = opendir(task_path.str); 2403 if (task_dir == NULL) { 2404 // Process can finish between reading "/proc/" directory entry and 2405 // opening process' "task/" directory. So, in general case we should not 2406 // complain, but have to skip this process and read the next one. But on 2407 // systems with no "task/" support we will spend lot of time to scan 2408 // "/proc/" tree again and again without any benefit. "init" process 2409 // (its pid is 1) should exist always, so, if we cannot open 2410 // "/proc/1/task/" directory, it means "task/" is not supported by 2411 // kernel. Report an error now and in the future. 2412 if (strcmp(proc_entry->d_name, "1") == 0) { 2413 running_threads = -1; 2414 permanent_error = 1; 2415 goto finish; 2416 } 2417 } else { 2418 // Construct fixed part of stat file path. 2419 __kmp_str_buf_clear(&stat_path); 2420 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used); 2421 __kmp_str_buf_cat(&stat_path, "/", 1); 2422 stat_path_fixed_len = stat_path.used; 2423 2424 task_entry = readdir(task_dir); 2425 while (task_entry != NULL) { 2426 // It is a directory and name starts with a digit. 2427 #if KMP_OS_AIX 2428 if (isdigit(task_entry->d_name[0])) { 2429 #else 2430 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) { 2431 #endif 2432 2433 // Construct complete stat file path. Easiest way would be: 2434 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str, 2435 // task_entry->d_name ); 2436 // but seriae of __kmp_str_buf_cat works a bit faster. 2437 stat_path.used = 2438 stat_path_fixed_len; // Reset stat path to its fixed part. 2439 __kmp_str_buf_cat(&stat_path, task_entry->d_name, 2440 KMP_STRLEN(task_entry->d_name)); 2441 __kmp_str_buf_cat(&stat_path, "/stat", 5); 2442 2443 // Note: Low-level API (open/read/close) is used. High-level API 2444 // (fopen/fclose) works ~ 30 % slower. 2445 stat_file = open(stat_path.str, O_RDONLY); 2446 if (stat_file == -1) { 2447 // We cannot report an error because task (thread) can terminate 2448 // just before reading this file. 2449 } else { 2450 /* Content of "stat" file looks like: 2451 24285 (program) S ... 2452 2453 It is a single line (if program name does not include funny 2454 symbols). First number is a thread id, then name of executable 2455 file name in paretheses, then state of the thread. We need just 2456 thread state. 2457 2458 Good news: Length of program name is 15 characters max. Longer 2459 names are truncated. 2460 2461 Thus, we need rather short buffer: 15 chars for program name + 2462 2 parenthesis, + 3 spaces + ~7 digits of pid = 37. 2463 2464 Bad news: Program name may contain special symbols like space, 2465 closing parenthesis, or even new line. This makes parsing 2466 "stat" file not 100 % reliable. In case of fanny program names 2467 parsing may fail (report incorrect thread state). 2468 2469 Parsing "status" file looks more promissing (due to different 2470 file structure and escaping special symbols) but reading and 2471 parsing of "status" file works slower. 2472 -- ln 2473 */ 2474 char buffer[65]; 2475 ssize_t len; 2476 len = read(stat_file, buffer, sizeof(buffer) - 1); 2477 if (len >= 0) { 2478 buffer[len] = 0; 2479 // Using scanf: 2480 // sscanf( buffer, "%*d (%*s) %c ", & state ); 2481 // looks very nice, but searching for a closing parenthesis 2482 // works a bit faster. 2483 char *close_parent = strstr(buffer, ") "); 2484 if (close_parent != NULL) { 2485 char state = *(close_parent + 2); 2486 if (state == 'R') { 2487 ++running_threads; 2488 if (running_threads >= max) { 2489 goto finish; 2490 } 2491 } 2492 } 2493 } 2494 close(stat_file); 2495 stat_file = -1; 2496 } 2497 } 2498 task_entry = readdir(task_dir); 2499 } 2500 closedir(task_dir); 2501 task_dir = NULL; 2502 } 2503 } 2504 proc_entry = readdir(proc_dir); 2505 } 2506 2507 // There _might_ be a timing hole where the thread executing this 2508 // code get skipped in the load balance, and running_threads is 0. 2509 // Assert in the debug builds only!!! 2510 KMP_DEBUG_ASSERT(running_threads > 0); 2511 if (running_threads <= 0) { 2512 running_threads = 1; 2513 } 2514 2515 finish: // Clean up and exit. 2516 if (proc_dir != NULL) { 2517 closedir(proc_dir); 2518 } 2519 __kmp_str_buf_free(&task_path); 2520 if (task_dir != NULL) { 2521 closedir(task_dir); 2522 } 2523 __kmp_str_buf_free(&stat_path); 2524 if (stat_file != -1) { 2525 close(stat_file); 2526 } 2527 2528 glb_running_threads = running_threads; 2529 2530 return running_threads; 2531 2532 } // __kmp_get_load_balance 2533 2534 #endif // KMP_OS_DARWIN 2535 2536 #endif // USE_LOAD_BALANCE 2537 2538 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2539 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \ 2540 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \ 2541 KMP_ARCH_ARM || KMP_ARCH_VE || KMP_ARCH_S390X || KMP_ARCH_PPC_XCOFF) 2542 2543 // we really only need the case with 1 argument, because CLANG always build 2544 // a struct of pointers to shared variables referenced in the outlined function 2545 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc, 2546 void *p_argv[] 2547 #if OMPT_SUPPORT 2548 , 2549 void **exit_frame_ptr 2550 #endif 2551 ) { 2552 #if OMPT_SUPPORT 2553 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0); 2554 #endif 2555 2556 switch (argc) { 2557 default: 2558 fprintf(stderr, "Too many args to microtask: %d!\n", argc); 2559 fflush(stderr); 2560 exit(-1); 2561 case 0: 2562 (*pkfn)(>id, &tid); 2563 break; 2564 case 1: 2565 (*pkfn)(>id, &tid, p_argv[0]); 2566 break; 2567 case 2: 2568 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]); 2569 break; 2570 case 3: 2571 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]); 2572 break; 2573 case 4: 2574 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]); 2575 break; 2576 case 5: 2577 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]); 2578 break; 2579 case 6: 2580 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2581 p_argv[5]); 2582 break; 2583 case 7: 2584 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2585 p_argv[5], p_argv[6]); 2586 break; 2587 case 8: 2588 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2589 p_argv[5], p_argv[6], p_argv[7]); 2590 break; 2591 case 9: 2592 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2593 p_argv[5], p_argv[6], p_argv[7], p_argv[8]); 2594 break; 2595 case 10: 2596 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2597 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]); 2598 break; 2599 case 11: 2600 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2601 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]); 2602 break; 2603 case 12: 2604 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2605 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2606 p_argv[11]); 2607 break; 2608 case 13: 2609 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2610 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2611 p_argv[11], p_argv[12]); 2612 break; 2613 case 14: 2614 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2615 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2616 p_argv[11], p_argv[12], p_argv[13]); 2617 break; 2618 case 15: 2619 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2620 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2621 p_argv[11], p_argv[12], p_argv[13], p_argv[14]); 2622 break; 2623 } 2624 2625 return 1; 2626 } 2627 2628 #endif 2629 2630 #if KMP_OS_LINUX 2631 // Functions for hidden helper task 2632 namespace { 2633 // Condition variable for initializing hidden helper team 2634 pthread_cond_t hidden_helper_threads_initz_cond_var; 2635 pthread_mutex_t hidden_helper_threads_initz_lock; 2636 volatile int hidden_helper_initz_signaled = FALSE; 2637 2638 // Condition variable for deinitializing hidden helper team 2639 pthread_cond_t hidden_helper_threads_deinitz_cond_var; 2640 pthread_mutex_t hidden_helper_threads_deinitz_lock; 2641 volatile int hidden_helper_deinitz_signaled = FALSE; 2642 2643 // Condition variable for the wrapper function of main thread 2644 pthread_cond_t hidden_helper_main_thread_cond_var; 2645 pthread_mutex_t hidden_helper_main_thread_lock; 2646 volatile int hidden_helper_main_thread_signaled = FALSE; 2647 2648 // Semaphore for worker threads. We don't use condition variable here in case 2649 // that when multiple signals are sent at the same time, only one thread might 2650 // be waken. 2651 sem_t hidden_helper_task_sem; 2652 } // namespace 2653 2654 void __kmp_hidden_helper_worker_thread_wait() { 2655 int status = sem_wait(&hidden_helper_task_sem); 2656 KMP_CHECK_SYSFAIL("sem_wait", status); 2657 } 2658 2659 void __kmp_do_initialize_hidden_helper_threads() { 2660 // Initialize condition variable 2661 int status = 2662 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr); 2663 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2664 2665 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr); 2666 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2667 2668 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr); 2669 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2670 2671 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr); 2672 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2673 2674 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr); 2675 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2676 2677 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr); 2678 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2679 2680 // Initialize the semaphore 2681 status = sem_init(&hidden_helper_task_sem, 0, 0); 2682 KMP_CHECK_SYSFAIL("sem_init", status); 2683 2684 // Create a new thread to finish initialization 2685 pthread_t handle; 2686 status = pthread_create( 2687 &handle, nullptr, 2688 [](void *) -> void * { 2689 __kmp_hidden_helper_threads_initz_routine(); 2690 return nullptr; 2691 }, 2692 nullptr); 2693 KMP_CHECK_SYSFAIL("pthread_create", status); 2694 } 2695 2696 void __kmp_hidden_helper_threads_initz_wait() { 2697 // Initial thread waits here for the completion of the initialization. The 2698 // condition variable will be notified by main thread of hidden helper teams. 2699 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2700 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2701 2702 if (!TCR_4(hidden_helper_initz_signaled)) { 2703 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var, 2704 &hidden_helper_threads_initz_lock); 2705 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2706 } 2707 2708 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2709 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2710 } 2711 2712 void __kmp_hidden_helper_initz_release() { 2713 // After all initialization, reset __kmp_init_hidden_helper_threads to false. 2714 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2715 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2716 2717 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var); 2718 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2719 2720 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE); 2721 2722 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2723 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2724 } 2725 2726 void __kmp_hidden_helper_main_thread_wait() { 2727 // The main thread of hidden helper team will be blocked here. The 2728 // condition variable can only be signal in the destructor of RTL. 2729 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2730 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2731 2732 if (!TCR_4(hidden_helper_main_thread_signaled)) { 2733 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var, 2734 &hidden_helper_main_thread_lock); 2735 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2736 } 2737 2738 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2739 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2740 } 2741 2742 void __kmp_hidden_helper_main_thread_release() { 2743 // The initial thread of OpenMP RTL should call this function to wake up the 2744 // main thread of hidden helper team. 2745 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2746 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2747 2748 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var); 2749 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 2750 2751 // The hidden helper team is done here 2752 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE); 2753 2754 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2755 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2756 } 2757 2758 void __kmp_hidden_helper_worker_thread_signal() { 2759 int status = sem_post(&hidden_helper_task_sem); 2760 KMP_CHECK_SYSFAIL("sem_post", status); 2761 } 2762 2763 void __kmp_hidden_helper_threads_deinitz_wait() { 2764 // Initial thread waits here for the completion of the deinitialization. The 2765 // condition variable will be notified by main thread of hidden helper teams. 2766 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2767 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2768 2769 if (!TCR_4(hidden_helper_deinitz_signaled)) { 2770 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var, 2771 &hidden_helper_threads_deinitz_lock); 2772 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2773 } 2774 2775 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2776 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2777 } 2778 2779 void __kmp_hidden_helper_threads_deinitz_release() { 2780 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2781 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2782 2783 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var); 2784 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2785 2786 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE); 2787 2788 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2789 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2790 } 2791 #else // KMP_OS_LINUX 2792 void __kmp_hidden_helper_worker_thread_wait() { 2793 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2794 } 2795 2796 void __kmp_do_initialize_hidden_helper_threads() { 2797 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2798 } 2799 2800 void __kmp_hidden_helper_threads_initz_wait() { 2801 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2802 } 2803 2804 void __kmp_hidden_helper_initz_release() { 2805 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2806 } 2807 2808 void __kmp_hidden_helper_main_thread_wait() { 2809 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2810 } 2811 2812 void __kmp_hidden_helper_main_thread_release() { 2813 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2814 } 2815 2816 void __kmp_hidden_helper_worker_thread_signal() { 2817 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2818 } 2819 2820 void __kmp_hidden_helper_threads_deinitz_wait() { 2821 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2822 } 2823 2824 void __kmp_hidden_helper_threads_deinitz_release() { 2825 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2826 } 2827 #endif // KMP_OS_LINUX 2828 2829 bool __kmp_detect_shm() { 2830 DIR *dir = opendir("/dev/shm"); 2831 if (dir) { // /dev/shm exists 2832 closedir(dir); 2833 return true; 2834 } else if (ENOENT == errno) { // /dev/shm does not exist 2835 return false; 2836 } else { // opendir() failed 2837 return false; 2838 } 2839 } 2840 2841 bool __kmp_detect_tmp() { 2842 DIR *dir = opendir("/tmp"); 2843 if (dir) { // /tmp exists 2844 closedir(dir); 2845 return true; 2846 } else if (ENOENT == errno) { // /tmp does not exist 2847 return false; 2848 } else { // opendir() failed 2849 return false; 2850 } 2851 } 2852 2853 // end of file // 2854