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 #include <sys/syscall.h> 33 #include <sys/time.h> 34 #include <sys/times.h> 35 #include <unistd.h> 36 37 #if KMP_OS_LINUX 38 #include <sys/sysinfo.h> 39 #if KMP_USE_FUTEX 40 // We should really include <futex.h>, but that causes compatibility problems on 41 // different Linux* OS distributions that either require that you include (or 42 // break when you try to include) <pci/types.h>. Since all we need is the two 43 // macros below (which are part of the kernel ABI, so can't change) we just 44 // define the constants here and don't include <futex.h> 45 #ifndef FUTEX_WAIT 46 #define FUTEX_WAIT 0 47 #endif 48 #ifndef FUTEX_WAKE 49 #define FUTEX_WAKE 1 50 #endif 51 #endif 52 #elif KMP_OS_DARWIN 53 #include <mach/mach.h> 54 #include <sys/sysctl.h> 55 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD 56 #include <sys/types.h> 57 #include <sys/sysctl.h> 58 #include <sys/user.h> 59 #include <pthread_np.h> 60 #elif KMP_OS_NETBSD || KMP_OS_OPENBSD 61 #include <sys/types.h> 62 #include <sys/sysctl.h> 63 #endif 64 65 #include <ctype.h> 66 #include <dirent.h> 67 #include <fcntl.h> 68 69 struct kmp_sys_timer { 70 struct timespec start; 71 }; 72 73 // Convert timespec to nanoseconds. 74 #define TS2NS(timespec) \ 75 (((timespec).tv_sec * (long int)1e9) + (timespec).tv_nsec) 76 77 static struct kmp_sys_timer __kmp_sys_timer_data; 78 79 #if KMP_HANDLE_SIGNALS 80 typedef void (*sig_func_t)(int); 81 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG]; 82 static sigset_t __kmp_sigset; 83 #endif 84 85 static int __kmp_init_runtime = FALSE; 86 87 static int __kmp_fork_count = 0; 88 89 static pthread_condattr_t __kmp_suspend_cond_attr; 90 static pthread_mutexattr_t __kmp_suspend_mutex_attr; 91 92 static kmp_cond_align_t __kmp_wait_cv; 93 static kmp_mutex_align_t __kmp_wait_mx; 94 95 kmp_uint64 __kmp_ticks_per_msec = 1000000; 96 97 #ifdef DEBUG_SUSPEND 98 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) { 99 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))", 100 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock, 101 cond->c_cond.__c_waiting); 102 } 103 #endif 104 105 #if ((KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED) 106 107 /* Affinity support */ 108 109 void __kmp_affinity_bind_thread(int which) { 110 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(), 111 "Illegal set affinity operation when not capable"); 112 113 kmp_affin_mask_t *mask; 114 KMP_CPU_ALLOC_ON_STACK(mask); 115 KMP_CPU_ZERO(mask); 116 KMP_CPU_SET(which, mask); 117 __kmp_set_system_affinity(mask, TRUE); 118 KMP_CPU_FREE_FROM_STACK(mask); 119 } 120 121 /* Determine if we can access affinity functionality on this version of 122 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set 123 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */ 124 void __kmp_affinity_determine_capable(const char *env_var) { 125 // Check and see if the OS supports thread affinity. 126 127 #if KMP_OS_LINUX 128 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024) 129 #define KMP_CPU_SET_TRY_SIZE CACHE_LINE 130 #elif KMP_OS_FREEBSD 131 #define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t)) 132 #endif 133 134 int verbose = __kmp_affinity.flags.verbose; 135 int warnings = __kmp_affinity.flags.warnings; 136 enum affinity_type type = __kmp_affinity.type; 137 138 #if KMP_OS_LINUX 139 long gCode; 140 unsigned char *buf; 141 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 142 143 // If the syscall returns a suggestion for the size, 144 // then we don't have to search for an appropriate size. 145 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf); 146 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 147 "initial getaffinity call returned %ld errno = %d\n", 148 gCode, errno)); 149 150 if (gCode < 0 && errno != EINVAL) { 151 // System call not supported 152 if (verbose || 153 (warnings && (type != affinity_none) && (type != affinity_default) && 154 (type != affinity_disabled))) { 155 int error = errno; 156 kmp_msg_t err_code = KMP_ERR(error); 157 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 158 err_code, __kmp_msg_null); 159 if (__kmp_generate_warnings == kmp_warnings_off) { 160 __kmp_str_free(&err_code.str); 161 } 162 } 163 KMP_AFFINITY_DISABLE(); 164 KMP_INTERNAL_FREE(buf); 165 return; 166 } else if (gCode > 0) { 167 // The optimal situation: the OS returns the size of the buffer it expects. 168 KMP_AFFINITY_ENABLE(gCode); 169 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 170 "affinity supported (mask size %d)\n", 171 (int)__kmp_affin_mask_size)); 172 KMP_INTERNAL_FREE(buf); 173 return; 174 } 175 176 // Call the getaffinity system call repeatedly with increasing set sizes 177 // until we succeed, or reach an upper bound on the search. 178 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 179 "searching for proper set size\n")); 180 int size; 181 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) { 182 gCode = syscall(__NR_sched_getaffinity, 0, size, buf); 183 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 184 "getaffinity for mask size %ld returned %ld errno = %d\n", 185 size, gCode, errno)); 186 187 if (gCode < 0) { 188 if (errno == ENOSYS) { 189 // We shouldn't get here 190 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 191 "inconsistent OS call behavior: errno == ENOSYS for mask " 192 "size %d\n", 193 size)); 194 if (verbose || 195 (warnings && (type != affinity_none) && 196 (type != affinity_default) && (type != affinity_disabled))) { 197 int error = errno; 198 kmp_msg_t err_code = KMP_ERR(error); 199 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 200 err_code, __kmp_msg_null); 201 if (__kmp_generate_warnings == kmp_warnings_off) { 202 __kmp_str_free(&err_code.str); 203 } 204 } 205 KMP_AFFINITY_DISABLE(); 206 KMP_INTERNAL_FREE(buf); 207 return; 208 } 209 continue; 210 } 211 212 KMP_AFFINITY_ENABLE(gCode); 213 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 214 "affinity supported (mask size %d)\n", 215 (int)__kmp_affin_mask_size)); 216 KMP_INTERNAL_FREE(buf); 217 return; 218 } 219 #elif KMP_OS_FREEBSD 220 long gCode; 221 unsigned char *buf; 222 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 223 gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT, 224 reinterpret_cast<cpuset_t *>(buf)); 225 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 226 "initial getaffinity call returned %d errno = %d\n", 227 gCode, errno)); 228 if (gCode == 0) { 229 KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT); 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 #endif 237 KMP_INTERNAL_FREE(buf); 238 239 // Affinity is not supported 240 KMP_AFFINITY_DISABLE(); 241 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 242 "cannot determine mask size - affinity not supported\n")); 243 if (verbose || (warnings && (type != affinity_none) && 244 (type != affinity_default) && (type != affinity_disabled))) { 245 KMP_WARNING(AffCantGetMaskSize, env_var); 246 } 247 } 248 249 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 250 251 #if KMP_USE_FUTEX 252 253 int __kmp_futex_determine_capable() { 254 int loc = 0; 255 long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0); 256 int retval = (rc == 0) || (errno != ENOSYS); 257 258 KA_TRACE(10, 259 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno)); 260 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n", 261 retval ? "" : " not")); 262 263 return retval; 264 } 265 266 #endif // KMP_USE_FUTEX 267 268 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS) 269 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to 270 use compare_and_store for these routines */ 271 272 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) { 273 kmp_int8 old_value, new_value; 274 275 old_value = TCR_1(*p); 276 new_value = old_value | d; 277 278 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 279 KMP_CPU_PAUSE(); 280 old_value = TCR_1(*p); 281 new_value = old_value | d; 282 } 283 return old_value; 284 } 285 286 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) { 287 kmp_int8 old_value, new_value; 288 289 old_value = TCR_1(*p); 290 new_value = old_value & d; 291 292 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 293 KMP_CPU_PAUSE(); 294 old_value = TCR_1(*p); 295 new_value = old_value & d; 296 } 297 return old_value; 298 } 299 300 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) { 301 kmp_uint32 old_value, new_value; 302 303 old_value = TCR_4(*p); 304 new_value = old_value | d; 305 306 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 307 KMP_CPU_PAUSE(); 308 old_value = TCR_4(*p); 309 new_value = old_value | d; 310 } 311 return old_value; 312 } 313 314 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) { 315 kmp_uint32 old_value, new_value; 316 317 old_value = TCR_4(*p); 318 new_value = old_value & d; 319 320 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 321 KMP_CPU_PAUSE(); 322 old_value = TCR_4(*p); 323 new_value = old_value & d; 324 } 325 return old_value; 326 } 327 328 #if KMP_ARCH_X86 329 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) { 330 kmp_int8 old_value, new_value; 331 332 old_value = TCR_1(*p); 333 new_value = old_value + d; 334 335 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 336 KMP_CPU_PAUSE(); 337 old_value = TCR_1(*p); 338 new_value = old_value + d; 339 } 340 return old_value; 341 } 342 343 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) { 344 kmp_int64 old_value, new_value; 345 346 old_value = TCR_8(*p); 347 new_value = old_value + d; 348 349 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 350 KMP_CPU_PAUSE(); 351 old_value = TCR_8(*p); 352 new_value = old_value + d; 353 } 354 return old_value; 355 } 356 #endif /* KMP_ARCH_X86 */ 357 358 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) { 359 kmp_uint64 old_value, new_value; 360 361 old_value = TCR_8(*p); 362 new_value = old_value | d; 363 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 364 KMP_CPU_PAUSE(); 365 old_value = TCR_8(*p); 366 new_value = old_value | d; 367 } 368 return old_value; 369 } 370 371 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) { 372 kmp_uint64 old_value, new_value; 373 374 old_value = TCR_8(*p); 375 new_value = old_value & d; 376 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 377 KMP_CPU_PAUSE(); 378 old_value = TCR_8(*p); 379 new_value = old_value & d; 380 } 381 return old_value; 382 } 383 384 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */ 385 386 void __kmp_terminate_thread(int gtid) { 387 int status; 388 kmp_info_t *th = __kmp_threads[gtid]; 389 390 if (!th) 391 return; 392 393 #ifdef KMP_CANCEL_THREADS 394 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid)); 395 status = pthread_cancel(th->th.th_info.ds.ds_thread); 396 if (status != 0 && status != ESRCH) { 397 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status), 398 __kmp_msg_null); 399 } 400 #endif 401 KMP_YIELD(TRUE); 402 } // 403 404 /* Set thread stack info according to values returned by pthread_getattr_np(). 405 If values are unreasonable, assume call failed and use incremental stack 406 refinement method instead. Returns TRUE if the stack parameters could be 407 determined exactly, FALSE if incremental refinement is necessary. */ 408 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) { 409 int stack_data; 410 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 411 KMP_OS_HURD 412 pthread_attr_t attr; 413 int status; 414 size_t size = 0; 415 void *addr = 0; 416 417 /* Always do incremental stack refinement for ubermaster threads since the 418 initial thread stack range can be reduced by sibling thread creation so 419 pthread_attr_getstack may cause thread gtid aliasing */ 420 if (!KMP_UBER_GTID(gtid)) { 421 422 /* Fetch the real thread attributes */ 423 status = pthread_attr_init(&attr); 424 KMP_CHECK_SYSFAIL("pthread_attr_init", status); 425 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD 426 status = pthread_attr_get_np(pthread_self(), &attr); 427 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status); 428 #else 429 status = pthread_getattr_np(pthread_self(), &attr); 430 KMP_CHECK_SYSFAIL("pthread_getattr_np", status); 431 #endif 432 status = pthread_attr_getstack(&attr, &addr, &size); 433 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status); 434 KA_TRACE(60, 435 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:" 436 " %lu, low addr: %p\n", 437 gtid, size, addr)); 438 status = pthread_attr_destroy(&attr); 439 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status); 440 } 441 442 if (size != 0 && addr != 0) { // was stack parameter determination successful? 443 /* Store the correct base and size */ 444 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size)); 445 TCW_PTR(th->th.th_info.ds.ds_stacksize, size); 446 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE); 447 return TRUE; 448 } 449 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD \ 450 || KMP_OS_HURD */ 451 /* Use incremental refinement starting from initial conservative estimate */ 452 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0); 453 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data); 454 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE); 455 return FALSE; 456 } 457 458 static void *__kmp_launch_worker(void *thr) { 459 int status, old_type, old_state; 460 #ifdef KMP_BLOCK_SIGNALS 461 sigset_t new_set, old_set; 462 #endif /* KMP_BLOCK_SIGNALS */ 463 void *exit_val; 464 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 465 KMP_OS_OPENBSD || KMP_OS_HURD 466 void *volatile padding = 0; 467 #endif 468 int gtid; 469 470 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid; 471 __kmp_gtid_set_specific(gtid); 472 #ifdef KMP_TDATA_GTID 473 __kmp_gtid = gtid; 474 #endif 475 #if KMP_STATS_ENABLED 476 // set thread local index to point to thread-specific stats 477 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats; 478 __kmp_stats_thread_ptr->startLife(); 479 KMP_SET_THREAD_STATE(IDLE); 480 KMP_INIT_PARTITIONED_TIMERS(OMP_idle); 481 #endif 482 483 #if USE_ITT_BUILD 484 __kmp_itt_thread_name(gtid); 485 #endif /* USE_ITT_BUILD */ 486 487 #if KMP_AFFINITY_SUPPORTED 488 __kmp_affinity_set_init_mask(gtid, FALSE); 489 #endif 490 491 #ifdef KMP_CANCEL_THREADS 492 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 493 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 494 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 495 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 496 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 497 #endif 498 499 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 500 // Set FP control regs to be a copy of the parallel initialization thread's. 501 __kmp_clear_x87_fpu_status_word(); 502 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); 503 __kmp_load_mxcsr(&__kmp_init_mxcsr); 504 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 505 506 #ifdef KMP_BLOCK_SIGNALS 507 status = sigfillset(&new_set); 508 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 509 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set); 510 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 511 #endif /* KMP_BLOCK_SIGNALS */ 512 513 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 514 KMP_OS_OPENBSD 515 if (__kmp_stkoffset > 0 && gtid > 0) { 516 padding = KMP_ALLOCA(gtid * __kmp_stkoffset); 517 (void)padding; 518 } 519 #endif 520 521 KMP_MB(); 522 __kmp_set_stack_info(gtid, (kmp_info_t *)thr); 523 524 __kmp_check_stack_overlap((kmp_info_t *)thr); 525 526 exit_val = __kmp_launch_thread((kmp_info_t *)thr); 527 528 #ifdef KMP_BLOCK_SIGNALS 529 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL); 530 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 531 #endif /* KMP_BLOCK_SIGNALS */ 532 533 return exit_val; 534 } 535 536 #if KMP_USE_MONITOR 537 /* The monitor thread controls all of the threads in the complex */ 538 539 static void *__kmp_launch_monitor(void *thr) { 540 int status, old_type, old_state; 541 #ifdef KMP_BLOCK_SIGNALS 542 sigset_t new_set; 543 #endif /* KMP_BLOCK_SIGNALS */ 544 struct timespec interval; 545 546 KMP_MB(); /* Flush all pending memory write invalidates. */ 547 548 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n")); 549 550 /* register us as the monitor thread */ 551 __kmp_gtid_set_specific(KMP_GTID_MONITOR); 552 #ifdef KMP_TDATA_GTID 553 __kmp_gtid = KMP_GTID_MONITOR; 554 #endif 555 556 KMP_MB(); 557 558 #if USE_ITT_BUILD 559 // Instruct Intel(R) Threading Tools to ignore monitor thread. 560 __kmp_itt_thread_ignore(); 561 #endif /* USE_ITT_BUILD */ 562 563 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid, 564 (kmp_info_t *)thr); 565 566 __kmp_check_stack_overlap((kmp_info_t *)thr); 567 568 #ifdef KMP_CANCEL_THREADS 569 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 570 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 571 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 572 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 573 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 574 #endif 575 576 #if KMP_REAL_TIME_FIX 577 // This is a potential fix which allows application with real-time scheduling 578 // policy work. However, decision about the fix is not made yet, so it is 579 // disabled by default. 580 { // Are program started with real-time scheduling policy? 581 int sched = sched_getscheduler(0); 582 if (sched == SCHED_FIFO || sched == SCHED_RR) { 583 // Yes, we are a part of real-time application. Try to increase the 584 // priority of the monitor. 585 struct sched_param param; 586 int max_priority = sched_get_priority_max(sched); 587 int rc; 588 KMP_WARNING(RealTimeSchedNotSupported); 589 sched_getparam(0, ¶m); 590 if (param.sched_priority < max_priority) { 591 param.sched_priority += 1; 592 rc = sched_setscheduler(0, sched, ¶m); 593 if (rc != 0) { 594 int error = errno; 595 kmp_msg_t err_code = KMP_ERR(error); 596 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority), 597 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null); 598 if (__kmp_generate_warnings == kmp_warnings_off) { 599 __kmp_str_free(&err_code.str); 600 } 601 } 602 } else { 603 // We cannot abort here, because number of CPUs may be enough for all 604 // the threads, including the monitor thread, so application could 605 // potentially work... 606 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority), 607 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority), 608 __kmp_msg_null); 609 } 610 } 611 // AC: free thread that waits for monitor started 612 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 613 } 614 #endif // KMP_REAL_TIME_FIX 615 616 KMP_MB(); /* Flush all pending memory write invalidates. */ 617 618 if (__kmp_monitor_wakeups == 1) { 619 interval.tv_sec = 1; 620 interval.tv_nsec = 0; 621 } else { 622 interval.tv_sec = 0; 623 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups); 624 } 625 626 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n")); 627 628 while (!TCR_4(__kmp_global.g.g_done)) { 629 struct timespec now; 630 struct timeval tval; 631 632 /* This thread monitors the state of the system */ 633 634 KA_TRACE(15, ("__kmp_launch_monitor: update\n")); 635 636 status = gettimeofday(&tval, NULL); 637 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 638 TIMEVAL_TO_TIMESPEC(&tval, &now); 639 640 now.tv_sec += interval.tv_sec; 641 now.tv_nsec += interval.tv_nsec; 642 643 if (now.tv_nsec >= KMP_NSEC_PER_SEC) { 644 now.tv_sec += 1; 645 now.tv_nsec -= KMP_NSEC_PER_SEC; 646 } 647 648 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 649 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 650 // AC: the monitor should not fall asleep if g_done has been set 651 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex 652 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond, 653 &__kmp_wait_mx.m_mutex, &now); 654 if (status != 0) { 655 if (status != ETIMEDOUT && status != EINTR) { 656 KMP_SYSFAIL("pthread_cond_timedwait", status); 657 } 658 } 659 } 660 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 661 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 662 663 TCW_4(__kmp_global.g.g_time.dt.t_value, 664 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1); 665 666 KMP_MB(); /* Flush all pending memory write invalidates. */ 667 } 668 669 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n")); 670 671 #ifdef KMP_BLOCK_SIGNALS 672 status = sigfillset(&new_set); 673 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 674 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL); 675 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 676 #endif /* KMP_BLOCK_SIGNALS */ 677 678 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n")); 679 680 if (__kmp_global.g.g_abort != 0) { 681 /* now we need to terminate the worker threads */ 682 /* the value of t_abort is the signal we caught */ 683 684 int gtid; 685 686 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n", 687 __kmp_global.g.g_abort)); 688 689 /* terminate the OpenMP worker threads */ 690 /* TODO this is not valid for sibling threads!! 691 * the uber master might not be 0 anymore.. */ 692 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid) 693 __kmp_terminate_thread(gtid); 694 695 __kmp_cleanup(); 696 697 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n", 698 __kmp_global.g.g_abort)); 699 700 if (__kmp_global.g.g_abort > 0) 701 raise(__kmp_global.g.g_abort); 702 } 703 704 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n")); 705 706 return thr; 707 } 708 #endif // KMP_USE_MONITOR 709 710 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) { 711 pthread_t handle; 712 pthread_attr_t thread_attr; 713 int status; 714 715 th->th.th_info.ds.ds_gtid = gtid; 716 717 #if KMP_STATS_ENABLED 718 // sets up worker thread stats 719 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid); 720 721 // th->th.th_stats is used to transfer thread-specific stats-pointer to 722 // __kmp_launch_worker. So when thread is created (goes into 723 // __kmp_launch_worker) it will set its thread local pointer to 724 // th->th.th_stats 725 if (!KMP_UBER_GTID(gtid)) { 726 th->th.th_stats = __kmp_stats_list->push_back(gtid); 727 } else { 728 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(), 729 // so set the th->th.th_stats field to it. 730 th->th.th_stats = __kmp_stats_thread_ptr; 731 } 732 __kmp_release_tas_lock(&__kmp_stats_lock, gtid); 733 734 #endif // KMP_STATS_ENABLED 735 736 if (KMP_UBER_GTID(gtid)) { 737 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid)); 738 th->th.th_info.ds.ds_thread = pthread_self(); 739 __kmp_set_stack_info(gtid, th); 740 __kmp_check_stack_overlap(th); 741 return; 742 } 743 744 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid)); 745 746 KMP_MB(); /* Flush all pending memory write invalidates. */ 747 748 #ifdef KMP_THREAD_ATTR 749 status = pthread_attr_init(&thread_attr); 750 if (status != 0) { 751 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 752 } 753 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 754 if (status != 0) { 755 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null); 756 } 757 758 /* Set stack size for this thread now. 759 The multiple of 2 is there because on some machines, requesting an unusual 760 stacksize causes the thread to have an offset before the dummy alloca() 761 takes place to create the offset. Since we want the user to have a 762 sufficient stacksize AND support a stack offset, we alloca() twice the 763 offset so that the upcoming alloca() does not eliminate any premade offset, 764 and also gives the user the stack space they requested for all threads */ 765 stack_size += gtid * __kmp_stkoffset * 2; 766 767 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 768 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n", 769 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 770 771 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 772 status = pthread_attr_setstacksize(&thread_attr, stack_size); 773 #ifdef KMP_BACKUP_STKSIZE 774 if (status != 0) { 775 if (!__kmp_env_stksize) { 776 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset; 777 __kmp_stksize = KMP_BACKUP_STKSIZE; 778 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 779 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu " 780 "bytes\n", 781 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 782 status = pthread_attr_setstacksize(&thread_attr, stack_size); 783 } 784 } 785 #endif /* KMP_BACKUP_STKSIZE */ 786 if (status != 0) { 787 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 788 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null); 789 } 790 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 791 792 #endif /* KMP_THREAD_ATTR */ 793 794 status = 795 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th); 796 if (status != 0 || !handle) { // ??? Why do we check handle?? 797 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 798 if (status == EINVAL) { 799 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 800 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null); 801 } 802 if (status == ENOMEM) { 803 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 804 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null); 805 } 806 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 807 if (status == EAGAIN) { 808 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status), 809 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null); 810 } 811 KMP_SYSFAIL("pthread_create", status); 812 } 813 814 th->th.th_info.ds.ds_thread = handle; 815 816 #ifdef KMP_THREAD_ATTR 817 status = pthread_attr_destroy(&thread_attr); 818 if (status) { 819 kmp_msg_t err_code = KMP_ERR(status); 820 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 821 __kmp_msg_null); 822 if (__kmp_generate_warnings == kmp_warnings_off) { 823 __kmp_str_free(&err_code.str); 824 } 825 } 826 #endif /* KMP_THREAD_ATTR */ 827 828 KMP_MB(); /* Flush all pending memory write invalidates. */ 829 830 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid)); 831 832 } // __kmp_create_worker 833 834 #if KMP_USE_MONITOR 835 void __kmp_create_monitor(kmp_info_t *th) { 836 pthread_t handle; 837 pthread_attr_t thread_attr; 838 size_t size; 839 int status; 840 int auto_adj_size = FALSE; 841 842 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { 843 // We don't need monitor thread in case of MAX_BLOCKTIME 844 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of " 845 "MAX blocktime\n")); 846 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op 847 th->th.th_info.ds.ds_gtid = 0; 848 return; 849 } 850 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n")); 851 852 KMP_MB(); /* Flush all pending memory write invalidates. */ 853 854 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR; 855 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR; 856 #if KMP_REAL_TIME_FIX 857 TCW_4(__kmp_global.g.g_time.dt.t_value, 858 -1); // Will use it for synchronization a bit later. 859 #else 860 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 861 #endif // KMP_REAL_TIME_FIX 862 863 #ifdef KMP_THREAD_ATTR 864 if (__kmp_monitor_stksize == 0) { 865 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 866 auto_adj_size = TRUE; 867 } 868 status = pthread_attr_init(&thread_attr); 869 if (status != 0) { 870 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 871 } 872 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 873 if (status != 0) { 874 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null); 875 } 876 877 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 878 status = pthread_attr_getstacksize(&thread_attr, &size); 879 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status); 880 #else 881 size = __kmp_sys_min_stksize; 882 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 883 #endif /* KMP_THREAD_ATTR */ 884 885 if (__kmp_monitor_stksize == 0) { 886 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 887 } 888 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) { 889 __kmp_monitor_stksize = __kmp_sys_min_stksize; 890 } 891 892 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes," 893 "requested stacksize = %lu bytes\n", 894 size, __kmp_monitor_stksize)); 895 896 retry: 897 898 /* Set stack size for this thread now. */ 899 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 900 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,", 901 __kmp_monitor_stksize)); 902 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize); 903 if (status != 0) { 904 if (auto_adj_size) { 905 __kmp_monitor_stksize *= 2; 906 goto retry; 907 } 908 kmp_msg_t err_code = KMP_ERR(status); 909 __kmp_msg(kmp_ms_warning, // should this be fatal? BB 910 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize), 911 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null); 912 if (__kmp_generate_warnings == kmp_warnings_off) { 913 __kmp_str_free(&err_code.str); 914 } 915 } 916 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 917 918 status = 919 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th); 920 921 if (status != 0) { 922 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 923 if (status == EINVAL) { 924 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) { 925 __kmp_monitor_stksize *= 2; 926 goto retry; 927 } 928 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 929 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize), 930 __kmp_msg_null); 931 } 932 if (status == ENOMEM) { 933 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 934 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize), 935 __kmp_msg_null); 936 } 937 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 938 if (status == EAGAIN) { 939 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status), 940 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null); 941 } 942 KMP_SYSFAIL("pthread_create", status); 943 } 944 945 th->th.th_info.ds.ds_thread = handle; 946 947 #if KMP_REAL_TIME_FIX 948 // Wait for the monitor thread is really started and set its *priority*. 949 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 950 sizeof(__kmp_global.g.g_time.dt.t_value)); 951 __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1, 952 &__kmp_neq_4, NULL); 953 #endif // KMP_REAL_TIME_FIX 954 955 #ifdef KMP_THREAD_ATTR 956 status = pthread_attr_destroy(&thread_attr); 957 if (status != 0) { 958 kmp_msg_t err_code = KMP_ERR(status); 959 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 960 __kmp_msg_null); 961 if (__kmp_generate_warnings == kmp_warnings_off) { 962 __kmp_str_free(&err_code.str); 963 } 964 } 965 #endif 966 967 KMP_MB(); /* Flush all pending memory write invalidates. */ 968 969 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n", 970 th->th.th_info.ds.ds_thread)); 971 972 } // __kmp_create_monitor 973 #endif // KMP_USE_MONITOR 974 975 void __kmp_exit_thread(int exit_status) { 976 pthread_exit((void *)(intptr_t)exit_status); 977 } // __kmp_exit_thread 978 979 #if KMP_USE_MONITOR 980 void __kmp_resume_monitor(); 981 982 extern "C" void __kmp_reap_monitor(kmp_info_t *th) { 983 int status; 984 void *exit_val; 985 986 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle" 987 " %#.8lx\n", 988 th->th.th_info.ds.ds_thread)); 989 990 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR. 991 // If both tid and gtid are 0, it means the monitor did not ever start. 992 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down. 993 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid); 994 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) { 995 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n")); 996 return; 997 } 998 999 KMP_MB(); /* Flush all pending memory write invalidates. */ 1000 1001 /* First, check to see whether the monitor thread exists to wake it up. This 1002 is to avoid performance problem when the monitor sleeps during 1003 blocktime-size interval */ 1004 1005 status = pthread_kill(th->th.th_info.ds.ds_thread, 0); 1006 if (status != ESRCH) { 1007 __kmp_resume_monitor(); // Wake up the monitor thread 1008 } 1009 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n")); 1010 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1011 if (exit_val != th) { 1012 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null); 1013 } 1014 1015 th->th.th_info.ds.ds_tid = KMP_GTID_DNE; 1016 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE; 1017 1018 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle" 1019 " %#.8lx\n", 1020 th->th.th_info.ds.ds_thread)); 1021 1022 KMP_MB(); /* Flush all pending memory write invalidates. */ 1023 } 1024 #else 1025 // Empty symbol to export (see exports_so.txt) when 1026 // monitor thread feature is disabled 1027 extern "C" void __kmp_reap_monitor(kmp_info_t *th) { 1028 (void)th; 1029 } 1030 #endif // KMP_USE_MONITOR 1031 1032 void __kmp_reap_worker(kmp_info_t *th) { 1033 int status; 1034 void *exit_val; 1035 1036 KMP_MB(); /* Flush all pending memory write invalidates. */ 1037 1038 KA_TRACE( 1039 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid)); 1040 1041 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1042 #ifdef KMP_DEBUG 1043 /* Don't expose these to the user until we understand when they trigger */ 1044 if (status != 0) { 1045 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null); 1046 } 1047 if (exit_val != th) { 1048 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, " 1049 "exit_val = %p\n", 1050 th->th.th_info.ds.ds_gtid, exit_val)); 1051 } 1052 #else 1053 (void)status; // unused variable 1054 #endif /* KMP_DEBUG */ 1055 1056 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n", 1057 th->th.th_info.ds.ds_gtid)); 1058 1059 KMP_MB(); /* Flush all pending memory write invalidates. */ 1060 } 1061 1062 #if KMP_HANDLE_SIGNALS 1063 1064 static void __kmp_null_handler(int signo) { 1065 // Do nothing, for doing SIG_IGN-type actions. 1066 } // __kmp_null_handler 1067 1068 static void __kmp_team_handler(int signo) { 1069 if (__kmp_global.g.g_abort == 0) { 1070 /* Stage 1 signal handler, let's shut down all of the threads */ 1071 #ifdef KMP_DEBUG 1072 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo); 1073 #endif 1074 switch (signo) { 1075 case SIGHUP: 1076 case SIGINT: 1077 case SIGQUIT: 1078 case SIGILL: 1079 case SIGABRT: 1080 case SIGFPE: 1081 case SIGBUS: 1082 case SIGSEGV: 1083 #ifdef SIGSYS 1084 case SIGSYS: 1085 #endif 1086 case SIGTERM: 1087 if (__kmp_debug_buf) { 1088 __kmp_dump_debug_buffer(); 1089 } 1090 __kmp_unregister_library(); // cleanup shared memory 1091 KMP_MB(); // Flush all pending memory write invalidates. 1092 TCW_4(__kmp_global.g.g_abort, signo); 1093 KMP_MB(); // Flush all pending memory write invalidates. 1094 TCW_4(__kmp_global.g.g_done, TRUE); 1095 KMP_MB(); // Flush all pending memory write invalidates. 1096 break; 1097 default: 1098 #ifdef KMP_DEBUG 1099 __kmp_debug_printf("__kmp_team_handler: unknown signal type"); 1100 #endif 1101 break; 1102 } 1103 } 1104 } // __kmp_team_handler 1105 1106 static void __kmp_sigaction(int signum, const struct sigaction *act, 1107 struct sigaction *oldact) { 1108 int rc = sigaction(signum, act, oldact); 1109 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc); 1110 } 1111 1112 static void __kmp_install_one_handler(int sig, sig_func_t handler_func, 1113 int parallel_init) { 1114 KMP_MB(); // Flush all pending memory write invalidates. 1115 KB_TRACE(60, 1116 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init)); 1117 if (parallel_init) { 1118 struct sigaction new_action; 1119 struct sigaction old_action; 1120 new_action.sa_handler = handler_func; 1121 new_action.sa_flags = 0; 1122 sigfillset(&new_action.sa_mask); 1123 __kmp_sigaction(sig, &new_action, &old_action); 1124 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) { 1125 sigaddset(&__kmp_sigset, sig); 1126 } else { 1127 // Restore/keep user's handler if one previously installed. 1128 __kmp_sigaction(sig, &old_action, NULL); 1129 } 1130 } else { 1131 // Save initial/system signal handlers to see if user handlers installed. 1132 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]); 1133 } 1134 KMP_MB(); // Flush all pending memory write invalidates. 1135 } // __kmp_install_one_handler 1136 1137 static void __kmp_remove_one_handler(int sig) { 1138 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig)); 1139 if (sigismember(&__kmp_sigset, sig)) { 1140 struct sigaction old; 1141 KMP_MB(); // Flush all pending memory write invalidates. 1142 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old); 1143 if ((old.sa_handler != __kmp_team_handler) && 1144 (old.sa_handler != __kmp_null_handler)) { 1145 // Restore the users signal handler. 1146 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, " 1147 "restoring: sig=%d\n", 1148 sig)); 1149 __kmp_sigaction(sig, &old, NULL); 1150 } 1151 sigdelset(&__kmp_sigset, sig); 1152 KMP_MB(); // Flush all pending memory write invalidates. 1153 } 1154 } // __kmp_remove_one_handler 1155 1156 void __kmp_install_signals(int parallel_init) { 1157 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init)); 1158 if (__kmp_handle_signals || !parallel_init) { 1159 // If ! parallel_init, we do not install handlers, just save original 1160 // handlers. Let us do it even __handle_signals is 0. 1161 sigemptyset(&__kmp_sigset); 1162 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init); 1163 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init); 1164 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init); 1165 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init); 1166 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init); 1167 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init); 1168 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init); 1169 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init); 1170 #ifdef SIGSYS 1171 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init); 1172 #endif // SIGSYS 1173 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init); 1174 #ifdef SIGPIPE 1175 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init); 1176 #endif // SIGPIPE 1177 } 1178 } // __kmp_install_signals 1179 1180 void __kmp_remove_signals(void) { 1181 int sig; 1182 KB_TRACE(10, ("__kmp_remove_signals()\n")); 1183 for (sig = 1; sig < NSIG; ++sig) { 1184 __kmp_remove_one_handler(sig); 1185 } 1186 } // __kmp_remove_signals 1187 1188 #endif // KMP_HANDLE_SIGNALS 1189 1190 void __kmp_enable(int new_state) { 1191 #ifdef KMP_CANCEL_THREADS 1192 int status, old_state; 1193 status = pthread_setcancelstate(new_state, &old_state); 1194 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1195 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE); 1196 #endif 1197 } 1198 1199 void __kmp_disable(int *old_state) { 1200 #ifdef KMP_CANCEL_THREADS 1201 int status; 1202 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state); 1203 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1204 #endif 1205 } 1206 1207 static void __kmp_atfork_prepare(void) { 1208 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 1209 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1210 } 1211 1212 static void __kmp_atfork_parent(void) { 1213 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1214 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1215 } 1216 1217 /* Reset the library so execution in the child starts "all over again" with 1218 clean data structures in initial states. Don't worry about freeing memory 1219 allocated by parent, just abandon it to be safe. */ 1220 static void __kmp_atfork_child(void) { 1221 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1222 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1223 /* TODO make sure this is done right for nested/sibling */ 1224 // ATT: Memory leaks are here? TODO: Check it and fix. 1225 /* KMP_ASSERT( 0 ); */ 1226 1227 ++__kmp_fork_count; 1228 1229 #if KMP_AFFINITY_SUPPORTED 1230 #if KMP_OS_LINUX || KMP_OS_FREEBSD 1231 // reset the affinity in the child to the initial thread 1232 // affinity in the parent 1233 kmp_set_thread_affinity_mask_initial(); 1234 #endif 1235 // Set default not to bind threads tightly in the child (we're expecting 1236 // over-subscription after the fork and this can improve things for 1237 // scripting languages that use OpenMP inside process-parallel code). 1238 if (__kmp_nested_proc_bind.bind_types != NULL) { 1239 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false; 1240 } 1241 for (kmp_affinity_t *affinity : __kmp_affinities) 1242 *affinity = KMP_AFFINITY_INIT(affinity->env_var); 1243 __kmp_affin_fullMask = nullptr; 1244 __kmp_affin_origMask = nullptr; 1245 #endif // KMP_AFFINITY_SUPPORTED 1246 1247 #if KMP_USE_MONITOR 1248 __kmp_init_monitor = 0; 1249 #endif 1250 __kmp_init_parallel = FALSE; 1251 __kmp_init_middle = FALSE; 1252 __kmp_init_serial = FALSE; 1253 TCW_4(__kmp_init_gtid, FALSE); 1254 __kmp_init_common = FALSE; 1255 1256 TCW_4(__kmp_init_user_locks, FALSE); 1257 #if !KMP_USE_DYNAMIC_LOCK 1258 __kmp_user_lock_table.used = 1; 1259 __kmp_user_lock_table.allocated = 0; 1260 __kmp_user_lock_table.table = NULL; 1261 __kmp_lock_blocks = NULL; 1262 #endif 1263 1264 __kmp_all_nth = 0; 1265 TCW_4(__kmp_nth, 0); 1266 1267 __kmp_thread_pool = NULL; 1268 __kmp_thread_pool_insert_pt = NULL; 1269 __kmp_team_pool = NULL; 1270 1271 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate 1272 here so threadprivate doesn't use stale data */ 1273 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n", 1274 __kmp_threadpriv_cache_list)); 1275 1276 while (__kmp_threadpriv_cache_list != NULL) { 1277 1278 if (*__kmp_threadpriv_cache_list->addr != NULL) { 1279 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n", 1280 &(*__kmp_threadpriv_cache_list->addr))); 1281 1282 *__kmp_threadpriv_cache_list->addr = NULL; 1283 } 1284 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next; 1285 } 1286 1287 __kmp_init_runtime = FALSE; 1288 1289 /* reset statically initialized locks */ 1290 __kmp_init_bootstrap_lock(&__kmp_initz_lock); 1291 __kmp_init_bootstrap_lock(&__kmp_stdio_lock); 1292 __kmp_init_bootstrap_lock(&__kmp_console_lock); 1293 __kmp_init_bootstrap_lock(&__kmp_task_team_lock); 1294 1295 #if USE_ITT_BUILD 1296 __kmp_itt_reset(); // reset ITT's global state 1297 #endif /* USE_ITT_BUILD */ 1298 1299 { 1300 // Child process often get terminated without any use of OpenMP. That might 1301 // cause mapped shared memory file to be left unattended. Thus we postpone 1302 // library registration till middle initialization in the child process. 1303 __kmp_need_register_serial = FALSE; 1304 __kmp_serial_initialize(); 1305 } 1306 1307 /* This is necessary to make sure no stale data is left around */ 1308 /* AC: customers complain that we use unsafe routines in the atfork 1309 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen 1310 in dynamic_link when check the presence of shared tbbmalloc library. 1311 Suggestion is to make the library initialization lazier, similar 1312 to what done for __kmpc_begin(). */ 1313 // TODO: synchronize all static initializations with regular library 1314 // startup; look at kmp_global.cpp and etc. 1315 //__kmp_internal_begin (); 1316 } 1317 1318 void __kmp_register_atfork(void) { 1319 if (__kmp_need_register_atfork) { 1320 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent, 1321 __kmp_atfork_child); 1322 KMP_CHECK_SYSFAIL("pthread_atfork", status); 1323 __kmp_need_register_atfork = FALSE; 1324 } 1325 } 1326 1327 void __kmp_suspend_initialize(void) { 1328 int status; 1329 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr); 1330 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1331 status = pthread_condattr_init(&__kmp_suspend_cond_attr); 1332 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1333 } 1334 1335 void __kmp_suspend_initialize_thread(kmp_info_t *th) { 1336 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count); 1337 int new_value = __kmp_fork_count + 1; 1338 // Return if already initialized 1339 if (old_value == new_value) 1340 return; 1341 // Wait, then return if being initialized 1342 if (old_value == -1 || !__kmp_atomic_compare_store( 1343 &th->th.th_suspend_init_count, old_value, -1)) { 1344 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) { 1345 KMP_CPU_PAUSE(); 1346 } 1347 } else { 1348 // Claim to be the initializer and do initializations 1349 int status; 1350 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond, 1351 &__kmp_suspend_cond_attr); 1352 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1353 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex, 1354 &__kmp_suspend_mutex_attr); 1355 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1356 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value); 1357 } 1358 } 1359 1360 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { 1361 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) { 1362 /* this means we have initialize the suspension pthread objects for this 1363 thread in this instance of the process */ 1364 int status; 1365 1366 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond); 1367 if (status != 0 && status != EBUSY) { 1368 KMP_SYSFAIL("pthread_cond_destroy", status); 1369 } 1370 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex); 1371 if (status != 0 && status != EBUSY) { 1372 KMP_SYSFAIL("pthread_mutex_destroy", status); 1373 } 1374 --th->th.th_suspend_init_count; 1375 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) == 1376 __kmp_fork_count); 1377 } 1378 } 1379 1380 // return true if lock obtained, false otherwise 1381 int __kmp_try_suspend_mx(kmp_info_t *th) { 1382 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0); 1383 } 1384 1385 void __kmp_lock_suspend_mx(kmp_info_t *th) { 1386 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1387 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1388 } 1389 1390 void __kmp_unlock_suspend_mx(kmp_info_t *th) { 1391 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1392 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1393 } 1394 1395 /* This routine puts the calling thread to sleep after setting the 1396 sleep bit for the indicated flag variable to true. */ 1397 template <class C> 1398 static inline void __kmp_suspend_template(int th_gtid, C *flag) { 1399 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend); 1400 kmp_info_t *th = __kmp_threads[th_gtid]; 1401 int status; 1402 typename C::flag_t old_spin; 1403 1404 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid, 1405 flag->get())); 1406 1407 __kmp_suspend_initialize_thread(th); 1408 1409 __kmp_lock_suspend_mx(th); 1410 1411 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n", 1412 th_gtid, flag->get())); 1413 1414 /* TODO: shouldn't this use release semantics to ensure that 1415 __kmp_suspend_initialize_thread gets called first? */ 1416 old_spin = flag->set_sleeping(); 1417 TCW_PTR(th->th.th_sleep_loc, (void *)flag); 1418 th->th.th_sleep_loc_type = flag->get_type(); 1419 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && 1420 __kmp_pause_status != kmp_soft_paused) { 1421 flag->unset_sleeping(); 1422 TCW_PTR(th->th.th_sleep_loc, NULL); 1423 th->th.th_sleep_loc_type = flag_unset; 1424 __kmp_unlock_suspend_mx(th); 1425 return; 1426 } 1427 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1428 " was %x\n", 1429 th_gtid, flag->get(), flag->load(), old_spin)); 1430 1431 if (flag->done_check_val(old_spin) || flag->done_check()) { 1432 flag->unset_sleeping(); 1433 TCW_PTR(th->th.th_sleep_loc, NULL); 1434 th->th.th_sleep_loc_type = flag_unset; 1435 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1436 "for spin(%p)\n", 1437 th_gtid, flag->get())); 1438 } else { 1439 /* Encapsulate in a loop as the documentation states that this may 1440 "with low probability" return when the condition variable has 1441 not been signaled or broadcast */ 1442 int deactivated = FALSE; 1443 1444 while (flag->is_sleeping()) { 1445 #ifdef DEBUG_SUSPEND 1446 char buffer[128]; 1447 __kmp_suspend_count++; 1448 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1449 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid, 1450 buffer); 1451 #endif 1452 // Mark the thread as no longer active (only in the first iteration of the 1453 // loop). 1454 if (!deactivated) { 1455 th->th.th_active = FALSE; 1456 if (th->th.th_active_in_pool) { 1457 th->th.th_active_in_pool = FALSE; 1458 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); 1459 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 1460 } 1461 deactivated = TRUE; 1462 } 1463 1464 KMP_DEBUG_ASSERT(th->th.th_sleep_loc); 1465 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type); 1466 1467 #if USE_SUSPEND_TIMEOUT 1468 struct timespec now; 1469 struct timeval tval; 1470 int msecs; 1471 1472 status = gettimeofday(&tval, NULL); 1473 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1474 TIMEVAL_TO_TIMESPEC(&tval, &now); 1475 1476 msecs = (4 * __kmp_dflt_blocktime) + 200; 1477 now.tv_sec += msecs / 1000; 1478 now.tv_nsec += (msecs % 1000) * 1000; 1479 1480 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " 1481 "pthread_cond_timedwait\n", 1482 th_gtid)); 1483 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond, 1484 &th->th.th_suspend_mx.m_mutex, &now); 1485 #else 1486 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform" 1487 " pthread_cond_wait\n", 1488 th_gtid)); 1489 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond, 1490 &th->th.th_suspend_mx.m_mutex); 1491 #endif // USE_SUSPEND_TIMEOUT 1492 1493 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) { 1494 KMP_SYSFAIL("pthread_cond_wait", status); 1495 } 1496 1497 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type()); 1498 1499 if (!flag->is_sleeping() && 1500 ((status == EINTR) || (status == ETIMEDOUT))) { 1501 // if interrupt or timeout, and thread is no longer sleeping, we need to 1502 // make sure sleep_loc gets reset; however, this shouldn't be needed if 1503 // we woke up with resume 1504 flag->unset_sleeping(); 1505 TCW_PTR(th->th.th_sleep_loc, NULL); 1506 th->th.th_sleep_loc_type = flag_unset; 1507 } 1508 #ifdef KMP_DEBUG 1509 if (status == ETIMEDOUT) { 1510 if (flag->is_sleeping()) { 1511 KF_TRACE(100, 1512 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid)); 1513 } else { 1514 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1515 "not set!\n", 1516 th_gtid)); 1517 TCW_PTR(th->th.th_sleep_loc, NULL); 1518 th->th.th_sleep_loc_type = flag_unset; 1519 } 1520 } else if (flag->is_sleeping()) { 1521 KF_TRACE(100, 1522 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid)); 1523 } 1524 #endif 1525 } // while 1526 1527 // Mark the thread as active again (if it was previous marked as inactive) 1528 if (deactivated) { 1529 th->th.th_active = TRUE; 1530 if (TCR_4(th->th.th_in_pool)) { 1531 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); 1532 th->th.th_active_in_pool = TRUE; 1533 } 1534 } 1535 } 1536 // We may have had the loop variable set before entering the loop body; 1537 // so we need to reset sleep_loc. 1538 TCW_PTR(th->th.th_sleep_loc, NULL); 1539 th->th.th_sleep_loc_type = flag_unset; 1540 1541 KMP_DEBUG_ASSERT(!flag->is_sleeping()); 1542 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc); 1543 #ifdef DEBUG_SUSPEND 1544 { 1545 char buffer[128]; 1546 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1547 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid, 1548 buffer); 1549 } 1550 #endif 1551 1552 __kmp_unlock_suspend_mx(th); 1553 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid)); 1554 } 1555 1556 template <bool C, bool S> 1557 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) { 1558 __kmp_suspend_template(th_gtid, flag); 1559 } 1560 template <bool C, bool S> 1561 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) { 1562 __kmp_suspend_template(th_gtid, flag); 1563 } 1564 template <bool C, bool S> 1565 void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) { 1566 __kmp_suspend_template(th_gtid, flag); 1567 } 1568 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { 1569 __kmp_suspend_template(th_gtid, flag); 1570 } 1571 1572 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *); 1573 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *); 1574 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *); 1575 template void 1576 __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1577 template void 1578 __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *); 1579 1580 /* This routine signals the thread specified by target_gtid to wake up 1581 after setting the sleep bit indicated by the flag argument to FALSE. 1582 The target thread must already have called __kmp_suspend_template() */ 1583 template <class C> 1584 static inline void __kmp_resume_template(int target_gtid, C *flag) { 1585 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1586 kmp_info_t *th = __kmp_threads[target_gtid]; 1587 int status; 1588 1589 #ifdef KMP_DEBUG 1590 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1591 #endif 1592 1593 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n", 1594 gtid, target_gtid)); 1595 KMP_DEBUG_ASSERT(gtid != target_gtid); 1596 1597 __kmp_suspend_initialize_thread(th); 1598 1599 __kmp_lock_suspend_mx(th); 1600 1601 if (!flag || flag != th->th.th_sleep_loc) { 1602 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a 1603 // different location; wake up at new location 1604 flag = (C *)CCAST(void *, th->th.th_sleep_loc); 1605 } 1606 1607 // First, check if the flag is null or its type has changed. If so, someone 1608 // else woke it up. 1609 if (!flag) { // Thread doesn't appear to be sleeping on anything 1610 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1611 "awake: flag(%p)\n", 1612 gtid, target_gtid, (void *)NULL)); 1613 __kmp_unlock_suspend_mx(th); 1614 return; 1615 } else if (flag->get_type() != th->th.th_sleep_loc_type) { 1616 // Flag type does not appear to match this function template; possibly the 1617 // thread is sleeping on something else. Try null resume again. 1618 KF_TRACE( 1619 5, 1620 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), " 1621 "spin(%p) type=%d ptr_type=%d\n", 1622 gtid, target_gtid, flag, flag->get(), flag->get_type(), 1623 th->th.th_sleep_loc_type)); 1624 __kmp_unlock_suspend_mx(th); 1625 __kmp_null_resume_wrapper(th); 1626 return; 1627 } else { // if multiple threads are sleeping, flag should be internally 1628 // referring to a specific thread here 1629 if (!flag->is_sleeping()) { 1630 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1631 "awake: flag(%p): %u\n", 1632 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1633 __kmp_unlock_suspend_mx(th); 1634 return; 1635 } 1636 } 1637 KMP_DEBUG_ASSERT(flag); 1638 flag->unset_sleeping(); 1639 TCW_PTR(th->th.th_sleep_loc, NULL); 1640 th->th.th_sleep_loc_type = flag_unset; 1641 1642 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1643 "sleep bit for flag's loc(%p): %u\n", 1644 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1645 1646 #ifdef DEBUG_SUSPEND 1647 { 1648 char buffer[128]; 1649 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1650 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid, 1651 target_gtid, buffer); 1652 } 1653 #endif 1654 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond); 1655 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1656 __kmp_unlock_suspend_mx(th); 1657 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" 1658 " for T#%d\n", 1659 gtid, target_gtid)); 1660 } 1661 1662 template <bool C, bool S> 1663 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) { 1664 __kmp_resume_template(target_gtid, flag); 1665 } 1666 template <bool C, bool S> 1667 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) { 1668 __kmp_resume_template(target_gtid, flag); 1669 } 1670 template <bool C, bool S> 1671 void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) { 1672 __kmp_resume_template(target_gtid, flag); 1673 } 1674 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { 1675 __kmp_resume_template(target_gtid, flag); 1676 } 1677 1678 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *); 1679 template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *); 1680 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *); 1681 template void 1682 __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1683 1684 #if KMP_USE_MONITOR 1685 void __kmp_resume_monitor() { 1686 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1687 int status; 1688 #ifdef KMP_DEBUG 1689 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1690 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid, 1691 KMP_GTID_MONITOR)); 1692 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR); 1693 #endif 1694 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 1695 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1696 #ifdef DEBUG_SUSPEND 1697 { 1698 char buffer[128]; 1699 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond); 1700 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid, 1701 KMP_GTID_MONITOR, buffer); 1702 } 1703 #endif 1704 status = pthread_cond_signal(&__kmp_wait_cv.c_cond); 1705 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1706 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 1707 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1708 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up" 1709 " for T#%d\n", 1710 gtid, KMP_GTID_MONITOR)); 1711 } 1712 #endif // KMP_USE_MONITOR 1713 1714 void __kmp_yield() { sched_yield(); } 1715 1716 void __kmp_gtid_set_specific(int gtid) { 1717 if (__kmp_init_gtid) { 1718 int status; 1719 status = pthread_setspecific(__kmp_gtid_threadprivate_key, 1720 (void *)(intptr_t)(gtid + 1)); 1721 KMP_CHECK_SYSFAIL("pthread_setspecific", status); 1722 } else { 1723 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n")); 1724 } 1725 } 1726 1727 int __kmp_gtid_get_specific() { 1728 int gtid; 1729 if (!__kmp_init_gtid) { 1730 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " 1731 "KMP_GTID_SHUTDOWN\n")); 1732 return KMP_GTID_SHUTDOWN; 1733 } 1734 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key); 1735 if (gtid == 0) { 1736 gtid = KMP_GTID_DNE; 1737 } else { 1738 gtid--; 1739 } 1740 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n", 1741 __kmp_gtid_threadprivate_key, gtid)); 1742 return gtid; 1743 } 1744 1745 double __kmp_read_cpu_time(void) { 1746 /*clock_t t;*/ 1747 struct tms buffer; 1748 1749 /*t =*/times(&buffer); 1750 1751 return (double)(buffer.tms_utime + buffer.tms_cutime) / 1752 (double)CLOCKS_PER_SEC; 1753 } 1754 1755 int __kmp_read_system_info(struct kmp_sys_info *info) { 1756 int status; 1757 struct rusage r_usage; 1758 1759 memset(info, 0, sizeof(*info)); 1760 1761 status = getrusage(RUSAGE_SELF, &r_usage); 1762 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status); 1763 1764 // The maximum resident set size utilized (in kilobytes) 1765 info->maxrss = r_usage.ru_maxrss; 1766 // The number of page faults serviced without any I/O 1767 info->minflt = r_usage.ru_minflt; 1768 // The number of page faults serviced that required I/O 1769 info->majflt = r_usage.ru_majflt; 1770 // The number of times a process was "swapped" out of memory 1771 info->nswap = r_usage.ru_nswap; 1772 // The number of times the file system had to perform input 1773 info->inblock = r_usage.ru_inblock; 1774 // The number of times the file system had to perform output 1775 info->oublock = r_usage.ru_oublock; 1776 // The number of times a context switch was voluntarily 1777 info->nvcsw = r_usage.ru_nvcsw; 1778 // The number of times a context switch was forced 1779 info->nivcsw = r_usage.ru_nivcsw; 1780 1781 return (status != 0); 1782 } 1783 1784 void __kmp_read_system_time(double *delta) { 1785 double t_ns; 1786 struct timeval tval; 1787 struct timespec stop; 1788 int status; 1789 1790 status = gettimeofday(&tval, NULL); 1791 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1792 TIMEVAL_TO_TIMESPEC(&tval, &stop); 1793 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start)); 1794 *delta = (t_ns * 1e-9); 1795 } 1796 1797 void __kmp_clear_system_time(void) { 1798 struct timeval tval; 1799 int status; 1800 status = gettimeofday(&tval, NULL); 1801 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1802 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start); 1803 } 1804 1805 static int __kmp_get_xproc(void) { 1806 1807 int r = 0; 1808 1809 #if KMP_OS_LINUX 1810 1811 __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r)); 1812 1813 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \ 1814 KMP_OS_HURD 1815 1816 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r)); 1817 1818 #elif KMP_OS_DARWIN 1819 1820 // Bug C77011 High "OpenMP Threads and number of active cores". 1821 1822 // Find the number of available CPUs. 1823 kern_return_t rc; 1824 host_basic_info_data_t info; 1825 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT; 1826 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num); 1827 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) { 1828 // Cannot use KA_TRACE() here because this code works before trace support 1829 // is initialized. 1830 r = info.avail_cpus; 1831 } else { 1832 KMP_WARNING(CantGetNumAvailCPU); 1833 KMP_INFORM(AssumedNumCPU); 1834 } 1835 1836 #else 1837 1838 #error "Unknown or unsupported OS." 1839 1840 #endif 1841 1842 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */ 1843 1844 } // __kmp_get_xproc 1845 1846 int __kmp_read_from_file(char const *path, char const *format, ...) { 1847 int result; 1848 va_list args; 1849 1850 va_start(args, format); 1851 FILE *f = fopen(path, "rb"); 1852 if (f == NULL) 1853 return 0; 1854 result = vfscanf(f, format, args); 1855 fclose(f); 1856 1857 return result; 1858 } 1859 1860 void __kmp_runtime_initialize(void) { 1861 int status; 1862 pthread_mutexattr_t mutex_attr; 1863 pthread_condattr_t cond_attr; 1864 1865 if (__kmp_init_runtime) { 1866 return; 1867 } 1868 1869 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1870 if (!__kmp_cpuinfo.initialized) { 1871 __kmp_query_cpuid(&__kmp_cpuinfo); 1872 } 1873 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1874 1875 __kmp_xproc = __kmp_get_xproc(); 1876 1877 #if !KMP_32_BIT_ARCH 1878 struct rlimit rlim; 1879 // read stack size of calling thread, save it as default for worker threads; 1880 // this should be done before reading environment variables 1881 status = getrlimit(RLIMIT_STACK, &rlim); 1882 if (status == 0) { // success? 1883 __kmp_stksize = rlim.rlim_cur; 1884 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed 1885 } 1886 #endif /* KMP_32_BIT_ARCH */ 1887 1888 if (sysconf(_SC_THREADS)) { 1889 1890 /* Query the maximum number of threads */ 1891 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth)); 1892 if (__kmp_sys_max_nth == -1) { 1893 /* Unlimited threads for NPTL */ 1894 __kmp_sys_max_nth = INT_MAX; 1895 } else if (__kmp_sys_max_nth <= 1) { 1896 /* Can't tell, just use PTHREAD_THREADS_MAX */ 1897 __kmp_sys_max_nth = KMP_MAX_NTH; 1898 } 1899 1900 /* Query the minimum stack size */ 1901 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN); 1902 if (__kmp_sys_min_stksize <= 1) { 1903 __kmp_sys_min_stksize = KMP_MIN_STKSIZE; 1904 } 1905 } 1906 1907 /* Set up minimum number of threads to switch to TLS gtid */ 1908 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; 1909 1910 status = pthread_key_create(&__kmp_gtid_threadprivate_key, 1911 __kmp_internal_end_dest); 1912 KMP_CHECK_SYSFAIL("pthread_key_create", status); 1913 status = pthread_mutexattr_init(&mutex_attr); 1914 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1915 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr); 1916 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1917 status = pthread_mutexattr_destroy(&mutex_attr); 1918 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status); 1919 status = pthread_condattr_init(&cond_attr); 1920 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1921 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr); 1922 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1923 status = pthread_condattr_destroy(&cond_attr); 1924 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status); 1925 #if USE_ITT_BUILD 1926 __kmp_itt_initialize(); 1927 #endif /* USE_ITT_BUILD */ 1928 1929 __kmp_init_runtime = TRUE; 1930 } 1931 1932 void __kmp_runtime_destroy(void) { 1933 int status; 1934 1935 if (!__kmp_init_runtime) { 1936 return; // Nothing to do. 1937 } 1938 1939 #if USE_ITT_BUILD 1940 __kmp_itt_destroy(); 1941 #endif /* USE_ITT_BUILD */ 1942 1943 status = pthread_key_delete(__kmp_gtid_threadprivate_key); 1944 KMP_CHECK_SYSFAIL("pthread_key_delete", status); 1945 1946 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex); 1947 if (status != 0 && status != EBUSY) { 1948 KMP_SYSFAIL("pthread_mutex_destroy", status); 1949 } 1950 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond); 1951 if (status != 0 && status != EBUSY) { 1952 KMP_SYSFAIL("pthread_cond_destroy", status); 1953 } 1954 #if KMP_AFFINITY_SUPPORTED 1955 __kmp_affinity_uninitialize(); 1956 #endif 1957 1958 __kmp_init_runtime = FALSE; 1959 } 1960 1961 /* Put the thread to sleep for a time period */ 1962 /* NOTE: not currently used anywhere */ 1963 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); } 1964 1965 /* Calculate the elapsed wall clock time for the user */ 1966 void __kmp_elapsed(double *t) { 1967 int status; 1968 #ifdef FIX_SGI_CLOCK 1969 struct timespec ts; 1970 1971 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); 1972 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status); 1973 *t = 1974 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec; 1975 #else 1976 struct timeval tv; 1977 1978 status = gettimeofday(&tv, NULL); 1979 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1980 *t = 1981 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec; 1982 #endif 1983 } 1984 1985 /* Calculate the elapsed wall clock tick for the user */ 1986 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; } 1987 1988 /* Return the current time stamp in nsec */ 1989 kmp_uint64 __kmp_now_nsec() { 1990 struct timeval t; 1991 gettimeofday(&t, NULL); 1992 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec + 1993 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec; 1994 return nsec; 1995 } 1996 1997 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1998 /* Measure clock ticks per millisecond */ 1999 void __kmp_initialize_system_tick() { 2000 kmp_uint64 now, nsec2, diff; 2001 kmp_uint64 delay = 100000; // 50~100 usec on most machines. 2002 kmp_uint64 nsec = __kmp_now_nsec(); 2003 kmp_uint64 goal = __kmp_hardware_timestamp() + delay; 2004 while ((now = __kmp_hardware_timestamp()) < goal) 2005 ; 2006 nsec2 = __kmp_now_nsec(); 2007 diff = nsec2 - nsec; 2008 if (diff > 0) { 2009 kmp_uint64 tpms = ((kmp_uint64)1e6 * (delay + (now - goal)) / diff); 2010 if (tpms > 0) 2011 __kmp_ticks_per_msec = tpms; 2012 } 2013 } 2014 #endif 2015 2016 /* Determine whether the given address is mapped into the current address 2017 space. */ 2018 2019 int __kmp_is_address_mapped(void *addr) { 2020 2021 int found = 0; 2022 int rc; 2023 2024 #if KMP_OS_LINUX || KMP_OS_HURD 2025 2026 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the 2027 address ranges mapped into the address space. */ 2028 2029 char *name = __kmp_str_format("/proc/%d/maps", getpid()); 2030 FILE *file = NULL; 2031 2032 file = fopen(name, "r"); 2033 KMP_ASSERT(file != NULL); 2034 2035 for (;;) { 2036 2037 void *beginning = NULL; 2038 void *ending = NULL; 2039 char perms[5]; 2040 2041 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms); 2042 if (rc == EOF) { 2043 break; 2044 } 2045 KMP_ASSERT(rc == 3 && 2046 KMP_STRLEN(perms) == 4); // Make sure all fields are read. 2047 2048 // Ending address is not included in the region, but beginning is. 2049 if ((addr >= beginning) && (addr < ending)) { 2050 perms[2] = 0; // 3th and 4th character does not matter. 2051 if (strcmp(perms, "rw") == 0) { 2052 // Memory we are looking for should be readable and writable. 2053 found = 1; 2054 } 2055 break; 2056 } 2057 } 2058 2059 // Free resources. 2060 fclose(file); 2061 KMP_INTERNAL_FREE(name); 2062 #elif KMP_OS_FREEBSD 2063 char *buf; 2064 size_t lstsz; 2065 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()}; 2066 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0); 2067 if (rc < 0) 2068 return 0; 2069 // We pass from number of vm entry's semantic 2070 // to size of whole entry map list. 2071 lstsz = lstsz * 4 / 3; 2072 buf = reinterpret_cast<char *>(kmpc_malloc(lstsz)); 2073 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0); 2074 if (rc < 0) { 2075 kmpc_free(buf); 2076 return 0; 2077 } 2078 2079 char *lw = buf; 2080 char *up = buf + lstsz; 2081 2082 while (lw < up) { 2083 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw); 2084 size_t cursz = cur->kve_structsize; 2085 if (cursz == 0) 2086 break; 2087 void *start = reinterpret_cast<void *>(cur->kve_start); 2088 void *end = reinterpret_cast<void *>(cur->kve_end); 2089 // Readable/Writable addresses within current map entry 2090 if ((addr >= start) && (addr < end)) { 2091 if ((cur->kve_protection & KVME_PROT_READ) != 0 && 2092 (cur->kve_protection & KVME_PROT_WRITE) != 0) { 2093 found = 1; 2094 break; 2095 } 2096 } 2097 lw += cursz; 2098 } 2099 kmpc_free(buf); 2100 2101 #elif KMP_OS_DARWIN 2102 2103 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory 2104 using vm interface. */ 2105 2106 int buffer; 2107 vm_size_t count; 2108 rc = vm_read_overwrite( 2109 mach_task_self(), // Task to read memory of. 2110 (vm_address_t)(addr), // Address to read from. 2111 1, // Number of bytes to be read. 2112 (vm_address_t)(&buffer), // Address of buffer to save read bytes in. 2113 &count // Address of var to save number of read bytes in. 2114 ); 2115 if (rc == 0) { 2116 // Memory successfully read. 2117 found = 1; 2118 } 2119 2120 #elif KMP_OS_NETBSD 2121 2122 int mib[5]; 2123 mib[0] = CTL_VM; 2124 mib[1] = VM_PROC; 2125 mib[2] = VM_PROC_MAP; 2126 mib[3] = getpid(); 2127 mib[4] = sizeof(struct kinfo_vmentry); 2128 2129 size_t size; 2130 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0); 2131 KMP_ASSERT(!rc); 2132 KMP_ASSERT(size); 2133 2134 size = size * 4 / 3; 2135 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size); 2136 KMP_ASSERT(kiv); 2137 2138 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0); 2139 KMP_ASSERT(!rc); 2140 KMP_ASSERT(size); 2141 2142 for (size_t i = 0; i < size; i++) { 2143 if (kiv[i].kve_start >= (uint64_t)addr && 2144 kiv[i].kve_end <= (uint64_t)addr) { 2145 found = 1; 2146 break; 2147 } 2148 } 2149 KMP_INTERNAL_FREE(kiv); 2150 #elif KMP_OS_OPENBSD 2151 2152 int mib[3]; 2153 mib[0] = CTL_KERN; 2154 mib[1] = KERN_PROC_VMMAP; 2155 mib[2] = getpid(); 2156 2157 size_t size; 2158 uint64_t end; 2159 rc = sysctl(mib, 3, NULL, &size, NULL, 0); 2160 KMP_ASSERT(!rc); 2161 KMP_ASSERT(size); 2162 end = size; 2163 2164 struct kinfo_vmentry kiv = {.kve_start = 0}; 2165 2166 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) { 2167 KMP_ASSERT(size); 2168 if (kiv.kve_end == end) 2169 break; 2170 2171 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) { 2172 found = 1; 2173 break; 2174 } 2175 kiv.kve_start += 1; 2176 } 2177 #elif KMP_OS_DRAGONFLY 2178 2179 // FIXME(DragonFly): Implement this 2180 found = 1; 2181 2182 #else 2183 2184 #error "Unknown or unsupported OS" 2185 2186 #endif 2187 2188 return found; 2189 2190 } // __kmp_is_address_mapped 2191 2192 #ifdef USE_LOAD_BALANCE 2193 2194 #if KMP_OS_DARWIN || KMP_OS_NETBSD 2195 2196 // The function returns the rounded value of the system load average 2197 // during given time interval which depends on the value of 2198 // __kmp_load_balance_interval variable (default is 60 sec, other values 2199 // may be 300 sec or 900 sec). 2200 // It returns -1 in case of error. 2201 int __kmp_get_load_balance(int max) { 2202 double averages[3]; 2203 int ret_avg = 0; 2204 2205 int res = getloadavg(averages, 3); 2206 2207 // Check __kmp_load_balance_interval to determine which of averages to use. 2208 // getloadavg() may return the number of samples less than requested that is 2209 // less than 3. 2210 if (__kmp_load_balance_interval < 180 && (res >= 1)) { 2211 ret_avg = (int)averages[0]; // 1 min 2212 } else if ((__kmp_load_balance_interval >= 180 && 2213 __kmp_load_balance_interval < 600) && 2214 (res >= 2)) { 2215 ret_avg = (int)averages[1]; // 5 min 2216 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) { 2217 ret_avg = (int)averages[2]; // 15 min 2218 } else { // Error occurred 2219 return -1; 2220 } 2221 2222 return ret_avg; 2223 } 2224 2225 #else // Linux* OS 2226 2227 // The function returns number of running (not sleeping) threads, or -1 in case 2228 // of error. Error could be reported if Linux* OS kernel too old (without 2229 // "/proc" support). Counting running threads stops if max running threads 2230 // encountered. 2231 int __kmp_get_load_balance(int max) { 2232 static int permanent_error = 0; 2233 static int glb_running_threads = 0; // Saved count of the running threads for 2234 // the thread balance algorithm 2235 static double glb_call_time = 0; /* Thread balance algorithm call time */ 2236 2237 int running_threads = 0; // Number of running threads in the system. 2238 2239 DIR *proc_dir = NULL; // Handle of "/proc/" directory. 2240 struct dirent *proc_entry = NULL; 2241 2242 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path. 2243 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory. 2244 struct dirent *task_entry = NULL; 2245 int task_path_fixed_len; 2246 2247 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path. 2248 int stat_file = -1; 2249 int stat_path_fixed_len; 2250 2251 #ifdef KMP_DEBUG 2252 int total_processes = 0; // Total number of processes in system. 2253 #endif 2254 2255 double call_time = 0.0; 2256 2257 __kmp_str_buf_init(&task_path); 2258 __kmp_str_buf_init(&stat_path); 2259 2260 __kmp_elapsed(&call_time); 2261 2262 if (glb_call_time && 2263 (call_time - glb_call_time < __kmp_load_balance_interval)) { 2264 running_threads = glb_running_threads; 2265 goto finish; 2266 } 2267 2268 glb_call_time = call_time; 2269 2270 // Do not spend time on scanning "/proc/" if we have a permanent error. 2271 if (permanent_error) { 2272 running_threads = -1; 2273 goto finish; 2274 } 2275 2276 if (max <= 0) { 2277 max = INT_MAX; 2278 } 2279 2280 // Open "/proc/" directory. 2281 proc_dir = opendir("/proc"); 2282 if (proc_dir == NULL) { 2283 // Cannot open "/prroc/". Probably the kernel does not support it. Return an 2284 // error now and in subsequent calls. 2285 running_threads = -1; 2286 permanent_error = 1; 2287 goto finish; 2288 } 2289 2290 // Initialize fixed part of task_path. This part will not change. 2291 __kmp_str_buf_cat(&task_path, "/proc/", 6); 2292 task_path_fixed_len = task_path.used; // Remember number of used characters. 2293 2294 proc_entry = readdir(proc_dir); 2295 while (proc_entry != NULL) { 2296 // Proc entry is a directory and name starts with a digit. Assume it is a 2297 // process' directory. 2298 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) { 2299 2300 #ifdef KMP_DEBUG 2301 ++total_processes; 2302 #endif 2303 // Make sure init process is the very first in "/proc", so we can replace 2304 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes == 2305 // 1. We are going to check that total_processes == 1 => d_name == "1" is 2306 // true (where "=>" is implication). Since C++ does not have => operator, 2307 // let us replace it with its equivalent: a => b == ! a || b. 2308 KMP_DEBUG_ASSERT(total_processes != 1 || 2309 strcmp(proc_entry->d_name, "1") == 0); 2310 2311 // Construct task_path. 2312 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/". 2313 __kmp_str_buf_cat(&task_path, proc_entry->d_name, 2314 KMP_STRLEN(proc_entry->d_name)); 2315 __kmp_str_buf_cat(&task_path, "/task", 5); 2316 2317 task_dir = opendir(task_path.str); 2318 if (task_dir == NULL) { 2319 // Process can finish between reading "/proc/" directory entry and 2320 // opening process' "task/" directory. So, in general case we should not 2321 // complain, but have to skip this process and read the next one. But on 2322 // systems with no "task/" support we will spend lot of time to scan 2323 // "/proc/" tree again and again without any benefit. "init" process 2324 // (its pid is 1) should exist always, so, if we cannot open 2325 // "/proc/1/task/" directory, it means "task/" is not supported by 2326 // kernel. Report an error now and in the future. 2327 if (strcmp(proc_entry->d_name, "1") == 0) { 2328 running_threads = -1; 2329 permanent_error = 1; 2330 goto finish; 2331 } 2332 } else { 2333 // Construct fixed part of stat file path. 2334 __kmp_str_buf_clear(&stat_path); 2335 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used); 2336 __kmp_str_buf_cat(&stat_path, "/", 1); 2337 stat_path_fixed_len = stat_path.used; 2338 2339 task_entry = readdir(task_dir); 2340 while (task_entry != NULL) { 2341 // It is a directory and name starts with a digit. 2342 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) { 2343 2344 // Construct complete stat file path. Easiest way would be: 2345 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str, 2346 // task_entry->d_name ); 2347 // but seriae of __kmp_str_buf_cat works a bit faster. 2348 stat_path.used = 2349 stat_path_fixed_len; // Reset stat path to its fixed part. 2350 __kmp_str_buf_cat(&stat_path, task_entry->d_name, 2351 KMP_STRLEN(task_entry->d_name)); 2352 __kmp_str_buf_cat(&stat_path, "/stat", 5); 2353 2354 // Note: Low-level API (open/read/close) is used. High-level API 2355 // (fopen/fclose) works ~ 30 % slower. 2356 stat_file = open(stat_path.str, O_RDONLY); 2357 if (stat_file == -1) { 2358 // We cannot report an error because task (thread) can terminate 2359 // just before reading this file. 2360 } else { 2361 /* Content of "stat" file looks like: 2362 24285 (program) S ... 2363 2364 It is a single line (if program name does not include funny 2365 symbols). First number is a thread id, then name of executable 2366 file name in paretheses, then state of the thread. We need just 2367 thread state. 2368 2369 Good news: Length of program name is 15 characters max. Longer 2370 names are truncated. 2371 2372 Thus, we need rather short buffer: 15 chars for program name + 2373 2 parenthesis, + 3 spaces + ~7 digits of pid = 37. 2374 2375 Bad news: Program name may contain special symbols like space, 2376 closing parenthesis, or even new line. This makes parsing 2377 "stat" file not 100 % reliable. In case of fanny program names 2378 parsing may fail (report incorrect thread state). 2379 2380 Parsing "status" file looks more promissing (due to different 2381 file structure and escaping special symbols) but reading and 2382 parsing of "status" file works slower. 2383 -- ln 2384 */ 2385 char buffer[65]; 2386 ssize_t len; 2387 len = read(stat_file, buffer, sizeof(buffer) - 1); 2388 if (len >= 0) { 2389 buffer[len] = 0; 2390 // Using scanf: 2391 // sscanf( buffer, "%*d (%*s) %c ", & state ); 2392 // looks very nice, but searching for a closing parenthesis 2393 // works a bit faster. 2394 char *close_parent = strstr(buffer, ") "); 2395 if (close_parent != NULL) { 2396 char state = *(close_parent + 2); 2397 if (state == 'R') { 2398 ++running_threads; 2399 if (running_threads >= max) { 2400 goto finish; 2401 } 2402 } 2403 } 2404 } 2405 close(stat_file); 2406 stat_file = -1; 2407 } 2408 } 2409 task_entry = readdir(task_dir); 2410 } 2411 closedir(task_dir); 2412 task_dir = NULL; 2413 } 2414 } 2415 proc_entry = readdir(proc_dir); 2416 } 2417 2418 // There _might_ be a timing hole where the thread executing this 2419 // code get skipped in the load balance, and running_threads is 0. 2420 // Assert in the debug builds only!!! 2421 KMP_DEBUG_ASSERT(running_threads > 0); 2422 if (running_threads <= 0) { 2423 running_threads = 1; 2424 } 2425 2426 finish: // Clean up and exit. 2427 if (proc_dir != NULL) { 2428 closedir(proc_dir); 2429 } 2430 __kmp_str_buf_free(&task_path); 2431 if (task_dir != NULL) { 2432 closedir(task_dir); 2433 } 2434 __kmp_str_buf_free(&stat_path); 2435 if (stat_file != -1) { 2436 close(stat_file); 2437 } 2438 2439 glb_running_threads = running_threads; 2440 2441 return running_threads; 2442 2443 } // __kmp_get_load_balance 2444 2445 #endif // KMP_OS_DARWIN 2446 2447 #endif // USE_LOAD_BALANCE 2448 2449 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2450 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \ 2451 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64 || KMP_ARCH_LOONGARCH64 || \ 2452 KMP_ARCH_ARM) 2453 2454 // we really only need the case with 1 argument, because CLANG always build 2455 // a struct of pointers to shared variables referenced in the outlined function 2456 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc, 2457 void *p_argv[] 2458 #if OMPT_SUPPORT 2459 , 2460 void **exit_frame_ptr 2461 #endif 2462 ) { 2463 #if OMPT_SUPPORT 2464 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0); 2465 #endif 2466 2467 switch (argc) { 2468 default: 2469 fprintf(stderr, "Too many args to microtask: %d!\n", argc); 2470 fflush(stderr); 2471 exit(-1); 2472 case 0: 2473 (*pkfn)(>id, &tid); 2474 break; 2475 case 1: 2476 (*pkfn)(>id, &tid, p_argv[0]); 2477 break; 2478 case 2: 2479 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]); 2480 break; 2481 case 3: 2482 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]); 2483 break; 2484 case 4: 2485 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]); 2486 break; 2487 case 5: 2488 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]); 2489 break; 2490 case 6: 2491 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2492 p_argv[5]); 2493 break; 2494 case 7: 2495 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2496 p_argv[5], p_argv[6]); 2497 break; 2498 case 8: 2499 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2500 p_argv[5], p_argv[6], p_argv[7]); 2501 break; 2502 case 9: 2503 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2504 p_argv[5], p_argv[6], p_argv[7], p_argv[8]); 2505 break; 2506 case 10: 2507 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2508 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]); 2509 break; 2510 case 11: 2511 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2512 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]); 2513 break; 2514 case 12: 2515 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2516 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2517 p_argv[11]); 2518 break; 2519 case 13: 2520 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2521 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2522 p_argv[11], p_argv[12]); 2523 break; 2524 case 14: 2525 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2526 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2527 p_argv[11], p_argv[12], p_argv[13]); 2528 break; 2529 case 15: 2530 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2531 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2532 p_argv[11], p_argv[12], p_argv[13], p_argv[14]); 2533 break; 2534 } 2535 2536 return 1; 2537 } 2538 2539 #endif 2540 2541 #if KMP_OS_LINUX 2542 // Functions for hidden helper task 2543 namespace { 2544 // Condition variable for initializing hidden helper team 2545 pthread_cond_t hidden_helper_threads_initz_cond_var; 2546 pthread_mutex_t hidden_helper_threads_initz_lock; 2547 volatile int hidden_helper_initz_signaled = FALSE; 2548 2549 // Condition variable for deinitializing hidden helper team 2550 pthread_cond_t hidden_helper_threads_deinitz_cond_var; 2551 pthread_mutex_t hidden_helper_threads_deinitz_lock; 2552 volatile int hidden_helper_deinitz_signaled = FALSE; 2553 2554 // Condition variable for the wrapper function of main thread 2555 pthread_cond_t hidden_helper_main_thread_cond_var; 2556 pthread_mutex_t hidden_helper_main_thread_lock; 2557 volatile int hidden_helper_main_thread_signaled = FALSE; 2558 2559 // Semaphore for worker threads. We don't use condition variable here in case 2560 // that when multiple signals are sent at the same time, only one thread might 2561 // be waken. 2562 sem_t hidden_helper_task_sem; 2563 } // namespace 2564 2565 void __kmp_hidden_helper_worker_thread_wait() { 2566 int status = sem_wait(&hidden_helper_task_sem); 2567 KMP_CHECK_SYSFAIL("sem_wait", status); 2568 } 2569 2570 void __kmp_do_initialize_hidden_helper_threads() { 2571 // Initialize condition variable 2572 int status = 2573 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr); 2574 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2575 2576 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr); 2577 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2578 2579 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr); 2580 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2581 2582 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr); 2583 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2584 2585 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr); 2586 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2587 2588 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr); 2589 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2590 2591 // Initialize the semaphore 2592 status = sem_init(&hidden_helper_task_sem, 0, 0); 2593 KMP_CHECK_SYSFAIL("sem_init", status); 2594 2595 // Create a new thread to finish initialization 2596 pthread_t handle; 2597 status = pthread_create( 2598 &handle, nullptr, 2599 [](void *) -> void * { 2600 __kmp_hidden_helper_threads_initz_routine(); 2601 return nullptr; 2602 }, 2603 nullptr); 2604 KMP_CHECK_SYSFAIL("pthread_create", status); 2605 } 2606 2607 void __kmp_hidden_helper_threads_initz_wait() { 2608 // Initial thread waits here for the completion of the initialization. The 2609 // condition variable will be notified by main thread of hidden helper teams. 2610 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2611 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2612 2613 if (!TCR_4(hidden_helper_initz_signaled)) { 2614 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var, 2615 &hidden_helper_threads_initz_lock); 2616 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2617 } 2618 2619 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2620 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2621 } 2622 2623 void __kmp_hidden_helper_initz_release() { 2624 // After all initialization, reset __kmp_init_hidden_helper_threads to false. 2625 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2626 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2627 2628 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var); 2629 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2630 2631 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE); 2632 2633 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2634 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2635 } 2636 2637 void __kmp_hidden_helper_main_thread_wait() { 2638 // The main thread of hidden helper team will be blocked here. The 2639 // condition variable can only be signal in the destructor of RTL. 2640 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2641 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2642 2643 if (!TCR_4(hidden_helper_main_thread_signaled)) { 2644 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var, 2645 &hidden_helper_main_thread_lock); 2646 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2647 } 2648 2649 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2650 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2651 } 2652 2653 void __kmp_hidden_helper_main_thread_release() { 2654 // The initial thread of OpenMP RTL should call this function to wake up the 2655 // main thread of hidden helper team. 2656 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2657 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2658 2659 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var); 2660 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 2661 2662 // The hidden helper team is done here 2663 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE); 2664 2665 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2666 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2667 } 2668 2669 void __kmp_hidden_helper_worker_thread_signal() { 2670 int status = sem_post(&hidden_helper_task_sem); 2671 KMP_CHECK_SYSFAIL("sem_post", status); 2672 } 2673 2674 void __kmp_hidden_helper_threads_deinitz_wait() { 2675 // Initial thread waits here for the completion of the deinitialization. The 2676 // condition variable will be notified by main thread of hidden helper teams. 2677 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2678 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2679 2680 if (!TCR_4(hidden_helper_deinitz_signaled)) { 2681 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var, 2682 &hidden_helper_threads_deinitz_lock); 2683 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2684 } 2685 2686 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2687 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2688 } 2689 2690 void __kmp_hidden_helper_threads_deinitz_release() { 2691 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2692 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2693 2694 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var); 2695 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2696 2697 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE); 2698 2699 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2700 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2701 } 2702 #else // KMP_OS_LINUX 2703 void __kmp_hidden_helper_worker_thread_wait() { 2704 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2705 } 2706 2707 void __kmp_do_initialize_hidden_helper_threads() { 2708 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2709 } 2710 2711 void __kmp_hidden_helper_threads_initz_wait() { 2712 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2713 } 2714 2715 void __kmp_hidden_helper_initz_release() { 2716 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2717 } 2718 2719 void __kmp_hidden_helper_main_thread_wait() { 2720 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2721 } 2722 2723 void __kmp_hidden_helper_main_thread_release() { 2724 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2725 } 2726 2727 void __kmp_hidden_helper_worker_thread_signal() { 2728 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2729 } 2730 2731 void __kmp_hidden_helper_threads_deinitz_wait() { 2732 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2733 } 2734 2735 void __kmp_hidden_helper_threads_deinitz_release() { 2736 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2737 } 2738 #endif // KMP_OS_LINUX 2739 2740 // end of file // 2741