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 #if KMP_OS_LINUX 135 long gCode; 136 unsigned char *buf; 137 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 138 139 // If the syscall returns a suggestion for the size, 140 // then we don't have to search for an appropriate size. 141 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_TRY_SIZE, buf); 142 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 143 "initial getaffinity call returned %ld errno = %d\n", 144 gCode, errno)); 145 146 if (gCode < 0 && errno != EINVAL) { 147 // System call not supported 148 if (__kmp_affinity_verbose || 149 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) && 150 (__kmp_affinity_type != affinity_default) && 151 (__kmp_affinity_type != affinity_disabled))) { 152 int error = errno; 153 kmp_msg_t err_code = KMP_ERR(error); 154 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 155 err_code, __kmp_msg_null); 156 if (__kmp_generate_warnings == kmp_warnings_off) { 157 __kmp_str_free(&err_code.str); 158 } 159 } 160 KMP_AFFINITY_DISABLE(); 161 KMP_INTERNAL_FREE(buf); 162 return; 163 } else if (gCode > 0) { 164 // The optimal situation: the OS returns the size of the buffer it expects. 165 KMP_AFFINITY_ENABLE(gCode); 166 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 167 "affinity supported (mask size %d)\n", 168 (int)__kmp_affin_mask_size)); 169 KMP_INTERNAL_FREE(buf); 170 return; 171 } 172 173 // Call the getaffinity system call repeatedly with increasing set sizes 174 // until we succeed, or reach an upper bound on the search. 175 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 176 "searching for proper set size\n")); 177 int size; 178 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) { 179 gCode = syscall(__NR_sched_getaffinity, 0, size, buf); 180 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 181 "getaffinity for mask size %ld returned %ld errno = %d\n", 182 size, gCode, errno)); 183 184 if (gCode < 0) { 185 if (errno == ENOSYS) { 186 // We shouldn't get here 187 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 188 "inconsistent OS call behavior: errno == ENOSYS for mask " 189 "size %d\n", 190 size)); 191 if (__kmp_affinity_verbose || 192 (__kmp_affinity_warnings && 193 (__kmp_affinity_type != affinity_none) && 194 (__kmp_affinity_type != affinity_default) && 195 (__kmp_affinity_type != affinity_disabled))) { 196 int error = errno; 197 kmp_msg_t err_code = KMP_ERR(error); 198 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 199 err_code, __kmp_msg_null); 200 if (__kmp_generate_warnings == kmp_warnings_off) { 201 __kmp_str_free(&err_code.str); 202 } 203 } 204 KMP_AFFINITY_DISABLE(); 205 KMP_INTERNAL_FREE(buf); 206 return; 207 } 208 continue; 209 } 210 211 KMP_AFFINITY_ENABLE(gCode); 212 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 213 "affinity supported (mask size %d)\n", 214 (int)__kmp_affin_mask_size)); 215 KMP_INTERNAL_FREE(buf); 216 return; 217 } 218 #elif KMP_OS_FREEBSD 219 long gCode; 220 unsigned char *buf; 221 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 222 gCode = pthread_getaffinity_np(pthread_self(), KMP_CPU_SET_SIZE_LIMIT, 223 reinterpret_cast<cpuset_t *>(buf)); 224 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 225 "initial getaffinity call returned %d errno = %d\n", 226 gCode, errno)); 227 if (gCode == 0) { 228 KMP_AFFINITY_ENABLE(KMP_CPU_SET_SIZE_LIMIT); 229 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 230 "affinity supported (mask size %d)\n", 231 (int)__kmp_affin_mask_size)); 232 KMP_INTERNAL_FREE(buf); 233 return; 234 } 235 #endif 236 KMP_INTERNAL_FREE(buf); 237 238 // Affinity is not supported 239 KMP_AFFINITY_DISABLE(); 240 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 241 "cannot determine mask size - affinity not supported\n")); 242 if (__kmp_affinity_verbose || 243 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) && 244 (__kmp_affinity_type != affinity_default) && 245 (__kmp_affinity_type != affinity_disabled))) { 246 KMP_WARNING(AffCantGetMaskSize, env_var); 247 } 248 } 249 250 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 251 252 #if KMP_USE_FUTEX 253 254 int __kmp_futex_determine_capable() { 255 int loc = 0; 256 long rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0); 257 int retval = (rc == 0) || (errno != ENOSYS); 258 259 KA_TRACE(10, 260 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno)); 261 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n", 262 retval ? "" : " not")); 263 264 return retval; 265 } 266 267 #endif // KMP_USE_FUTEX 268 269 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS) 270 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to 271 use compare_and_store for these routines */ 272 273 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) { 274 kmp_int8 old_value, new_value; 275 276 old_value = TCR_1(*p); 277 new_value = old_value | d; 278 279 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 280 KMP_CPU_PAUSE(); 281 old_value = TCR_1(*p); 282 new_value = old_value | d; 283 } 284 return old_value; 285 } 286 287 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) { 288 kmp_int8 old_value, new_value; 289 290 old_value = TCR_1(*p); 291 new_value = old_value & d; 292 293 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 294 KMP_CPU_PAUSE(); 295 old_value = TCR_1(*p); 296 new_value = old_value & d; 297 } 298 return old_value; 299 } 300 301 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) { 302 kmp_uint32 old_value, new_value; 303 304 old_value = TCR_4(*p); 305 new_value = old_value | d; 306 307 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 308 KMP_CPU_PAUSE(); 309 old_value = TCR_4(*p); 310 new_value = old_value | d; 311 } 312 return old_value; 313 } 314 315 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) { 316 kmp_uint32 old_value, new_value; 317 318 old_value = TCR_4(*p); 319 new_value = old_value & d; 320 321 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 322 KMP_CPU_PAUSE(); 323 old_value = TCR_4(*p); 324 new_value = old_value & d; 325 } 326 return old_value; 327 } 328 329 #if KMP_ARCH_X86 330 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) { 331 kmp_int8 old_value, new_value; 332 333 old_value = TCR_1(*p); 334 new_value = old_value + d; 335 336 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 337 KMP_CPU_PAUSE(); 338 old_value = TCR_1(*p); 339 new_value = old_value + d; 340 } 341 return old_value; 342 } 343 344 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) { 345 kmp_int64 old_value, new_value; 346 347 old_value = TCR_8(*p); 348 new_value = old_value + d; 349 350 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 351 KMP_CPU_PAUSE(); 352 old_value = TCR_8(*p); 353 new_value = old_value + d; 354 } 355 return old_value; 356 } 357 #endif /* KMP_ARCH_X86 */ 358 359 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) { 360 kmp_uint64 old_value, new_value; 361 362 old_value = TCR_8(*p); 363 new_value = old_value | d; 364 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 365 KMP_CPU_PAUSE(); 366 old_value = TCR_8(*p); 367 new_value = old_value | d; 368 } 369 return old_value; 370 } 371 372 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) { 373 kmp_uint64 old_value, new_value; 374 375 old_value = TCR_8(*p); 376 new_value = old_value & d; 377 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 378 KMP_CPU_PAUSE(); 379 old_value = TCR_8(*p); 380 new_value = old_value & d; 381 } 382 return old_value; 383 } 384 385 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */ 386 387 void __kmp_terminate_thread(int gtid) { 388 int status; 389 kmp_info_t *th = __kmp_threads[gtid]; 390 391 if (!th) 392 return; 393 394 #ifdef KMP_CANCEL_THREADS 395 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid)); 396 status = pthread_cancel(th->th.th_info.ds.ds_thread); 397 if (status != 0 && status != ESRCH) { 398 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status), 399 __kmp_msg_null); 400 } 401 #endif 402 KMP_YIELD(TRUE); 403 } // 404 405 /* Set thread stack info according to values returned by pthread_getattr_np(). 406 If values are unreasonable, assume call failed and use incremental stack 407 refinement method instead. Returns TRUE if the stack parameters could be 408 determined exactly, FALSE if incremental refinement is necessary. */ 409 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) { 410 int stack_data; 411 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 412 KMP_OS_HURD 413 pthread_attr_t attr; 414 int status; 415 size_t size = 0; 416 void *addr = 0; 417 418 /* Always do incremental stack refinement for ubermaster threads since the 419 initial thread stack range can be reduced by sibling thread creation so 420 pthread_attr_getstack may cause thread gtid aliasing */ 421 if (!KMP_UBER_GTID(gtid)) { 422 423 /* Fetch the real thread attributes */ 424 status = pthread_attr_init(&attr); 425 KMP_CHECK_SYSFAIL("pthread_attr_init", status); 426 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD 427 status = pthread_attr_get_np(pthread_self(), &attr); 428 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status); 429 #else 430 status = pthread_getattr_np(pthread_self(), &attr); 431 KMP_CHECK_SYSFAIL("pthread_getattr_np", status); 432 #endif 433 status = pthread_attr_getstack(&attr, &addr, &size); 434 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status); 435 KA_TRACE(60, 436 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:" 437 " %lu, low addr: %p\n", 438 gtid, size, addr)); 439 status = pthread_attr_destroy(&attr); 440 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status); 441 } 442 443 if (size != 0 && addr != 0) { // was stack parameter determination successful? 444 /* Store the correct base and size */ 445 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size)); 446 TCW_PTR(th->th.th_info.ds.ds_stacksize, size); 447 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE); 448 return TRUE; 449 } 450 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD \ 451 || KMP_OS_HURD */ 452 /* Use incremental refinement starting from initial conservative estimate */ 453 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0); 454 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data); 455 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE); 456 return FALSE; 457 } 458 459 static void *__kmp_launch_worker(void *thr) { 460 int status, old_type, old_state; 461 #ifdef KMP_BLOCK_SIGNALS 462 sigset_t new_set, old_set; 463 #endif /* KMP_BLOCK_SIGNALS */ 464 void *exit_val; 465 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 466 KMP_OS_OPENBSD || KMP_OS_HURD 467 void *volatile padding = 0; 468 #endif 469 int gtid; 470 471 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid; 472 __kmp_gtid_set_specific(gtid); 473 #ifdef KMP_TDATA_GTID 474 __kmp_gtid = gtid; 475 #endif 476 #if KMP_STATS_ENABLED 477 // set thread local index to point to thread-specific stats 478 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats; 479 __kmp_stats_thread_ptr->startLife(); 480 KMP_SET_THREAD_STATE(IDLE); 481 KMP_INIT_PARTITIONED_TIMERS(OMP_idle); 482 #endif 483 484 #if USE_ITT_BUILD 485 __kmp_itt_thread_name(gtid); 486 #endif /* USE_ITT_BUILD */ 487 488 #if KMP_AFFINITY_SUPPORTED 489 __kmp_affinity_set_init_mask(gtid, FALSE); 490 #endif 491 492 #ifdef KMP_CANCEL_THREADS 493 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 494 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 495 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 496 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 497 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 498 #endif 499 500 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 501 // Set FP control regs to be a copy of the parallel initialization thread's. 502 __kmp_clear_x87_fpu_status_word(); 503 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); 504 __kmp_load_mxcsr(&__kmp_init_mxcsr); 505 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 506 507 #ifdef KMP_BLOCK_SIGNALS 508 status = sigfillset(&new_set); 509 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 510 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set); 511 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 512 #endif /* KMP_BLOCK_SIGNALS */ 513 514 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 515 KMP_OS_OPENBSD 516 if (__kmp_stkoffset > 0 && gtid > 0) { 517 padding = KMP_ALLOCA(gtid * __kmp_stkoffset); 518 (void)padding; 519 } 520 #endif 521 522 KMP_MB(); 523 __kmp_set_stack_info(gtid, (kmp_info_t *)thr); 524 525 __kmp_check_stack_overlap((kmp_info_t *)thr); 526 527 exit_val = __kmp_launch_thread((kmp_info_t *)thr); 528 529 #ifdef KMP_BLOCK_SIGNALS 530 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL); 531 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 532 #endif /* KMP_BLOCK_SIGNALS */ 533 534 return exit_val; 535 } 536 537 #if KMP_USE_MONITOR 538 /* The monitor thread controls all of the threads in the complex */ 539 540 static void *__kmp_launch_monitor(void *thr) { 541 int status, old_type, old_state; 542 #ifdef KMP_BLOCK_SIGNALS 543 sigset_t new_set; 544 #endif /* KMP_BLOCK_SIGNALS */ 545 struct timespec interval; 546 547 KMP_MB(); /* Flush all pending memory write invalidates. */ 548 549 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n")); 550 551 /* register us as the monitor thread */ 552 __kmp_gtid_set_specific(KMP_GTID_MONITOR); 553 #ifdef KMP_TDATA_GTID 554 __kmp_gtid = KMP_GTID_MONITOR; 555 #endif 556 557 KMP_MB(); 558 559 #if USE_ITT_BUILD 560 // Instruct Intel(R) Threading Tools to ignore monitor thread. 561 __kmp_itt_thread_ignore(); 562 #endif /* USE_ITT_BUILD */ 563 564 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid, 565 (kmp_info_t *)thr); 566 567 __kmp_check_stack_overlap((kmp_info_t *)thr); 568 569 #ifdef KMP_CANCEL_THREADS 570 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 571 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 572 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 573 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 574 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 575 #endif 576 577 #if KMP_REAL_TIME_FIX 578 // This is a potential fix which allows application with real-time scheduling 579 // policy work. However, decision about the fix is not made yet, so it is 580 // disabled by default. 581 { // Are program started with real-time scheduling policy? 582 int sched = sched_getscheduler(0); 583 if (sched == SCHED_FIFO || sched == SCHED_RR) { 584 // Yes, we are a part of real-time application. Try to increase the 585 // priority of the monitor. 586 struct sched_param param; 587 int max_priority = sched_get_priority_max(sched); 588 int rc; 589 KMP_WARNING(RealTimeSchedNotSupported); 590 sched_getparam(0, ¶m); 591 if (param.sched_priority < max_priority) { 592 param.sched_priority += 1; 593 rc = sched_setscheduler(0, sched, ¶m); 594 if (rc != 0) { 595 int error = errno; 596 kmp_msg_t err_code = KMP_ERR(error); 597 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority), 598 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null); 599 if (__kmp_generate_warnings == kmp_warnings_off) { 600 __kmp_str_free(&err_code.str); 601 } 602 } 603 } else { 604 // We cannot abort here, because number of CPUs may be enough for all 605 // the threads, including the monitor thread, so application could 606 // potentially work... 607 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority), 608 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority), 609 __kmp_msg_null); 610 } 611 } 612 // AC: free thread that waits for monitor started 613 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 614 } 615 #endif // KMP_REAL_TIME_FIX 616 617 KMP_MB(); /* Flush all pending memory write invalidates. */ 618 619 if (__kmp_monitor_wakeups == 1) { 620 interval.tv_sec = 1; 621 interval.tv_nsec = 0; 622 } else { 623 interval.tv_sec = 0; 624 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups); 625 } 626 627 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n")); 628 629 while (!TCR_4(__kmp_global.g.g_done)) { 630 struct timespec now; 631 struct timeval tval; 632 633 /* This thread monitors the state of the system */ 634 635 KA_TRACE(15, ("__kmp_launch_monitor: update\n")); 636 637 status = gettimeofday(&tval, NULL); 638 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 639 TIMEVAL_TO_TIMESPEC(&tval, &now); 640 641 now.tv_sec += interval.tv_sec; 642 now.tv_nsec += interval.tv_nsec; 643 644 if (now.tv_nsec >= KMP_NSEC_PER_SEC) { 645 now.tv_sec += 1; 646 now.tv_nsec -= KMP_NSEC_PER_SEC; 647 } 648 649 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 650 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 651 // AC: the monitor should not fall asleep if g_done has been set 652 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex 653 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond, 654 &__kmp_wait_mx.m_mutex, &now); 655 if (status != 0) { 656 if (status != ETIMEDOUT && status != EINTR) { 657 KMP_SYSFAIL("pthread_cond_timedwait", status); 658 } 659 } 660 } 661 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 662 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 663 664 TCW_4(__kmp_global.g.g_time.dt.t_value, 665 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1); 666 667 KMP_MB(); /* Flush all pending memory write invalidates. */ 668 } 669 670 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n")); 671 672 #ifdef KMP_BLOCK_SIGNALS 673 status = sigfillset(&new_set); 674 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 675 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL); 676 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 677 #endif /* KMP_BLOCK_SIGNALS */ 678 679 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n")); 680 681 if (__kmp_global.g.g_abort != 0) { 682 /* now we need to terminate the worker threads */ 683 /* the value of t_abort is the signal we caught */ 684 685 int gtid; 686 687 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n", 688 __kmp_global.g.g_abort)); 689 690 /* terminate the OpenMP worker threads */ 691 /* TODO this is not valid for sibling threads!! 692 * the uber master might not be 0 anymore.. */ 693 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid) 694 __kmp_terminate_thread(gtid); 695 696 __kmp_cleanup(); 697 698 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n", 699 __kmp_global.g.g_abort)); 700 701 if (__kmp_global.g.g_abort > 0) 702 raise(__kmp_global.g.g_abort); 703 } 704 705 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n")); 706 707 return thr; 708 } 709 #endif // KMP_USE_MONITOR 710 711 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) { 712 pthread_t handle; 713 pthread_attr_t thread_attr; 714 int status; 715 716 th->th.th_info.ds.ds_gtid = gtid; 717 718 #if KMP_STATS_ENABLED 719 // sets up worker thread stats 720 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid); 721 722 // th->th.th_stats is used to transfer thread-specific stats-pointer to 723 // __kmp_launch_worker. So when thread is created (goes into 724 // __kmp_launch_worker) it will set its thread local pointer to 725 // th->th.th_stats 726 if (!KMP_UBER_GTID(gtid)) { 727 th->th.th_stats = __kmp_stats_list->push_back(gtid); 728 } else { 729 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(), 730 // so set the th->th.th_stats field to it. 731 th->th.th_stats = __kmp_stats_thread_ptr; 732 } 733 __kmp_release_tas_lock(&__kmp_stats_lock, gtid); 734 735 #endif // KMP_STATS_ENABLED 736 737 if (KMP_UBER_GTID(gtid)) { 738 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid)); 739 th->th.th_info.ds.ds_thread = pthread_self(); 740 __kmp_set_stack_info(gtid, th); 741 __kmp_check_stack_overlap(th); 742 return; 743 } 744 745 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid)); 746 747 KMP_MB(); /* Flush all pending memory write invalidates. */ 748 749 #ifdef KMP_THREAD_ATTR 750 status = pthread_attr_init(&thread_attr); 751 if (status != 0) { 752 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 753 } 754 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 755 if (status != 0) { 756 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null); 757 } 758 759 /* Set stack size for this thread now. 760 The multiple of 2 is there because on some machines, requesting an unusual 761 stacksize causes the thread to have an offset before the dummy alloca() 762 takes place to create the offset. Since we want the user to have a 763 sufficient stacksize AND support a stack offset, we alloca() twice the 764 offset so that the upcoming alloca() does not eliminate any premade offset, 765 and also gives the user the stack space they requested for all threads */ 766 stack_size += gtid * __kmp_stkoffset * 2; 767 768 #if defined(__ANDROID__) && __ANDROID_API__ < 19 769 // Round the stack size to a multiple of the page size. Older versions of 770 // Android (until KitKat) would fail pthread_attr_setstacksize with EINVAL 771 // if the stack size was not a multiple of the page size. 772 stack_size = (stack_size + PAGE_SIZE - 1) & ~(PAGE_SIZE - 1); 773 #endif 774 775 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 776 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n", 777 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 778 779 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 780 status = pthread_attr_setstacksize(&thread_attr, stack_size); 781 #ifdef KMP_BACKUP_STKSIZE 782 if (status != 0) { 783 if (!__kmp_env_stksize) { 784 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset; 785 __kmp_stksize = KMP_BACKUP_STKSIZE; 786 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 787 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu " 788 "bytes\n", 789 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 790 status = pthread_attr_setstacksize(&thread_attr, stack_size); 791 } 792 } 793 #endif /* KMP_BACKUP_STKSIZE */ 794 if (status != 0) { 795 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 796 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null); 797 } 798 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 799 800 #endif /* KMP_THREAD_ATTR */ 801 802 status = 803 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th); 804 if (status != 0 || !handle) { // ??? Why do we check handle?? 805 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 806 if (status == EINVAL) { 807 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 808 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null); 809 } 810 if (status == ENOMEM) { 811 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 812 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null); 813 } 814 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 815 if (status == EAGAIN) { 816 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status), 817 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null); 818 } 819 KMP_SYSFAIL("pthread_create", status); 820 } 821 822 th->th.th_info.ds.ds_thread = handle; 823 824 #ifdef KMP_THREAD_ATTR 825 status = pthread_attr_destroy(&thread_attr); 826 if (status) { 827 kmp_msg_t err_code = KMP_ERR(status); 828 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 829 __kmp_msg_null); 830 if (__kmp_generate_warnings == kmp_warnings_off) { 831 __kmp_str_free(&err_code.str); 832 } 833 } 834 #endif /* KMP_THREAD_ATTR */ 835 836 KMP_MB(); /* Flush all pending memory write invalidates. */ 837 838 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid)); 839 840 } // __kmp_create_worker 841 842 #if KMP_USE_MONITOR 843 void __kmp_create_monitor(kmp_info_t *th) { 844 pthread_t handle; 845 pthread_attr_t thread_attr; 846 size_t size; 847 int status; 848 int auto_adj_size = FALSE; 849 850 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { 851 // We don't need monitor thread in case of MAX_BLOCKTIME 852 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of " 853 "MAX blocktime\n")); 854 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op 855 th->th.th_info.ds.ds_gtid = 0; 856 return; 857 } 858 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n")); 859 860 KMP_MB(); /* Flush all pending memory write invalidates. */ 861 862 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR; 863 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR; 864 #if KMP_REAL_TIME_FIX 865 TCW_4(__kmp_global.g.g_time.dt.t_value, 866 -1); // Will use it for synchronization a bit later. 867 #else 868 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 869 #endif // KMP_REAL_TIME_FIX 870 871 #ifdef KMP_THREAD_ATTR 872 if (__kmp_monitor_stksize == 0) { 873 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 874 auto_adj_size = TRUE; 875 } 876 status = pthread_attr_init(&thread_attr); 877 if (status != 0) { 878 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 879 } 880 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 881 if (status != 0) { 882 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null); 883 } 884 885 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 886 status = pthread_attr_getstacksize(&thread_attr, &size); 887 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status); 888 #else 889 size = __kmp_sys_min_stksize; 890 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 891 #endif /* KMP_THREAD_ATTR */ 892 893 if (__kmp_monitor_stksize == 0) { 894 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 895 } 896 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) { 897 __kmp_monitor_stksize = __kmp_sys_min_stksize; 898 } 899 900 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes," 901 "requested stacksize = %lu bytes\n", 902 size, __kmp_monitor_stksize)); 903 904 retry: 905 906 /* Set stack size for this thread now. */ 907 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 908 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,", 909 __kmp_monitor_stksize)); 910 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize); 911 if (status != 0) { 912 if (auto_adj_size) { 913 __kmp_monitor_stksize *= 2; 914 goto retry; 915 } 916 kmp_msg_t err_code = KMP_ERR(status); 917 __kmp_msg(kmp_ms_warning, // should this be fatal? BB 918 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize), 919 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null); 920 if (__kmp_generate_warnings == kmp_warnings_off) { 921 __kmp_str_free(&err_code.str); 922 } 923 } 924 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 925 926 status = 927 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th); 928 929 if (status != 0) { 930 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 931 if (status == EINVAL) { 932 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) { 933 __kmp_monitor_stksize *= 2; 934 goto retry; 935 } 936 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 937 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize), 938 __kmp_msg_null); 939 } 940 if (status == ENOMEM) { 941 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 942 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize), 943 __kmp_msg_null); 944 } 945 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 946 if (status == EAGAIN) { 947 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status), 948 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null); 949 } 950 KMP_SYSFAIL("pthread_create", status); 951 } 952 953 th->th.th_info.ds.ds_thread = handle; 954 955 #if KMP_REAL_TIME_FIX 956 // Wait for the monitor thread is really started and set its *priority*. 957 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 958 sizeof(__kmp_global.g.g_time.dt.t_value)); 959 __kmp_wait_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, -1, 960 &__kmp_neq_4, NULL); 961 #endif // KMP_REAL_TIME_FIX 962 963 #ifdef KMP_THREAD_ATTR 964 status = pthread_attr_destroy(&thread_attr); 965 if (status != 0) { 966 kmp_msg_t err_code = KMP_ERR(status); 967 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 968 __kmp_msg_null); 969 if (__kmp_generate_warnings == kmp_warnings_off) { 970 __kmp_str_free(&err_code.str); 971 } 972 } 973 #endif 974 975 KMP_MB(); /* Flush all pending memory write invalidates. */ 976 977 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n", 978 th->th.th_info.ds.ds_thread)); 979 980 } // __kmp_create_monitor 981 #endif // KMP_USE_MONITOR 982 983 void __kmp_exit_thread(int exit_status) { 984 pthread_exit((void *)(intptr_t)exit_status); 985 } // __kmp_exit_thread 986 987 #if KMP_USE_MONITOR 988 void __kmp_resume_monitor(); 989 990 void __kmp_reap_monitor(kmp_info_t *th) { 991 int status; 992 void *exit_val; 993 994 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle" 995 " %#.8lx\n", 996 th->th.th_info.ds.ds_thread)); 997 998 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR. 999 // If both tid and gtid are 0, it means the monitor did not ever start. 1000 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down. 1001 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid); 1002 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) { 1003 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n")); 1004 return; 1005 } 1006 1007 KMP_MB(); /* Flush all pending memory write invalidates. */ 1008 1009 /* First, check to see whether the monitor thread exists to wake it up. This 1010 is to avoid performance problem when the monitor sleeps during 1011 blocktime-size interval */ 1012 1013 status = pthread_kill(th->th.th_info.ds.ds_thread, 0); 1014 if (status != ESRCH) { 1015 __kmp_resume_monitor(); // Wake up the monitor thread 1016 } 1017 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n")); 1018 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1019 if (exit_val != th) { 1020 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null); 1021 } 1022 1023 th->th.th_info.ds.ds_tid = KMP_GTID_DNE; 1024 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE; 1025 1026 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle" 1027 " %#.8lx\n", 1028 th->th.th_info.ds.ds_thread)); 1029 1030 KMP_MB(); /* Flush all pending memory write invalidates. */ 1031 } 1032 #endif // KMP_USE_MONITOR 1033 1034 void __kmp_reap_worker(kmp_info_t *th) { 1035 int status; 1036 void *exit_val; 1037 1038 KMP_MB(); /* Flush all pending memory write invalidates. */ 1039 1040 KA_TRACE( 1041 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid)); 1042 1043 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1044 #ifdef KMP_DEBUG 1045 /* Don't expose these to the user until we understand when they trigger */ 1046 if (status != 0) { 1047 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null); 1048 } 1049 if (exit_val != th) { 1050 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, " 1051 "exit_val = %p\n", 1052 th->th.th_info.ds.ds_gtid, exit_val)); 1053 } 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 __kmp_affinity_type = affinity_none; 1239 if (__kmp_nested_proc_bind.bind_types != NULL) { 1240 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false; 1241 } 1242 __kmp_affinity_masks = NULL; 1243 __kmp_affinity_num_masks = 0; 1244 #endif // KMP_AFFINITY_SUPPORTED 1245 1246 #if KMP_USE_MONITOR 1247 __kmp_init_monitor = 0; 1248 #endif 1249 __kmp_init_parallel = FALSE; 1250 __kmp_init_middle = FALSE; 1251 __kmp_init_serial = FALSE; 1252 TCW_4(__kmp_init_gtid, FALSE); 1253 __kmp_init_common = FALSE; 1254 1255 TCW_4(__kmp_init_user_locks, FALSE); 1256 #if !KMP_USE_DYNAMIC_LOCK 1257 __kmp_user_lock_table.used = 1; 1258 __kmp_user_lock_table.allocated = 0; 1259 __kmp_user_lock_table.table = NULL; 1260 __kmp_lock_blocks = NULL; 1261 #endif 1262 1263 __kmp_all_nth = 0; 1264 TCW_4(__kmp_nth, 0); 1265 1266 __kmp_thread_pool = NULL; 1267 __kmp_thread_pool_insert_pt = NULL; 1268 __kmp_team_pool = NULL; 1269 1270 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate 1271 here so threadprivate doesn't use stale data */ 1272 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n", 1273 __kmp_threadpriv_cache_list)); 1274 1275 while (__kmp_threadpriv_cache_list != NULL) { 1276 1277 if (*__kmp_threadpriv_cache_list->addr != NULL) { 1278 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n", 1279 &(*__kmp_threadpriv_cache_list->addr))); 1280 1281 *__kmp_threadpriv_cache_list->addr = NULL; 1282 } 1283 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next; 1284 } 1285 1286 __kmp_init_runtime = FALSE; 1287 1288 /* reset statically initialized locks */ 1289 __kmp_init_bootstrap_lock(&__kmp_initz_lock); 1290 __kmp_init_bootstrap_lock(&__kmp_stdio_lock); 1291 __kmp_init_bootstrap_lock(&__kmp_console_lock); 1292 __kmp_init_bootstrap_lock(&__kmp_task_team_lock); 1293 1294 #if USE_ITT_BUILD 1295 __kmp_itt_reset(); // reset ITT's global state 1296 #endif /* USE_ITT_BUILD */ 1297 1298 __kmp_serial_initialize(); 1299 1300 /* This is necessary to make sure no stale data is left around */ 1301 /* AC: customers complain that we use unsafe routines in the atfork 1302 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen 1303 in dynamic_link when check the presence of shared tbbmalloc library. 1304 Suggestion is to make the library initialization lazier, similar 1305 to what done for __kmpc_begin(). */ 1306 // TODO: synchronize all static initializations with regular library 1307 // startup; look at kmp_global.cpp and etc. 1308 //__kmp_internal_begin (); 1309 } 1310 1311 void __kmp_register_atfork(void) { 1312 if (__kmp_need_register_atfork) { 1313 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent, 1314 __kmp_atfork_child); 1315 KMP_CHECK_SYSFAIL("pthread_atfork", status); 1316 __kmp_need_register_atfork = FALSE; 1317 } 1318 } 1319 1320 void __kmp_suspend_initialize(void) { 1321 int status; 1322 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr); 1323 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1324 status = pthread_condattr_init(&__kmp_suspend_cond_attr); 1325 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1326 } 1327 1328 void __kmp_suspend_initialize_thread(kmp_info_t *th) { 1329 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count); 1330 int new_value = __kmp_fork_count + 1; 1331 // Return if already initialized 1332 if (old_value == new_value) 1333 return; 1334 // Wait, then return if being initialized 1335 if (old_value == -1 || !__kmp_atomic_compare_store( 1336 &th->th.th_suspend_init_count, old_value, -1)) { 1337 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) { 1338 KMP_CPU_PAUSE(); 1339 } 1340 } else { 1341 // Claim to be the initializer and do initializations 1342 int status; 1343 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond, 1344 &__kmp_suspend_cond_attr); 1345 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1346 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex, 1347 &__kmp_suspend_mutex_attr); 1348 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1349 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value); 1350 } 1351 } 1352 1353 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { 1354 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) { 1355 /* this means we have initialize the suspension pthread objects for this 1356 thread in this instance of the process */ 1357 int status; 1358 1359 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond); 1360 if (status != 0 && status != EBUSY) { 1361 KMP_SYSFAIL("pthread_cond_destroy", status); 1362 } 1363 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex); 1364 if (status != 0 && status != EBUSY) { 1365 KMP_SYSFAIL("pthread_mutex_destroy", status); 1366 } 1367 --th->th.th_suspend_init_count; 1368 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) == 1369 __kmp_fork_count); 1370 } 1371 } 1372 1373 // return true if lock obtained, false otherwise 1374 int __kmp_try_suspend_mx(kmp_info_t *th) { 1375 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0); 1376 } 1377 1378 void __kmp_lock_suspend_mx(kmp_info_t *th) { 1379 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1380 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1381 } 1382 1383 void __kmp_unlock_suspend_mx(kmp_info_t *th) { 1384 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1385 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1386 } 1387 1388 /* This routine puts the calling thread to sleep after setting the 1389 sleep bit for the indicated flag variable to true. */ 1390 template <class C> 1391 static inline void __kmp_suspend_template(int th_gtid, C *flag) { 1392 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend); 1393 kmp_info_t *th = __kmp_threads[th_gtid]; 1394 int status; 1395 typename C::flag_t old_spin; 1396 1397 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid, 1398 flag->get())); 1399 1400 __kmp_suspend_initialize_thread(th); 1401 1402 __kmp_lock_suspend_mx(th); 1403 1404 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n", 1405 th_gtid, flag->get())); 1406 1407 /* TODO: shouldn't this use release semantics to ensure that 1408 __kmp_suspend_initialize_thread gets called first? */ 1409 old_spin = flag->set_sleeping(); 1410 TCW_PTR(th->th.th_sleep_loc, (void *)flag); 1411 th->th.th_sleep_loc_type = flag->get_type(); 1412 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && 1413 __kmp_pause_status != kmp_soft_paused) { 1414 flag->unset_sleeping(); 1415 TCW_PTR(th->th.th_sleep_loc, NULL); 1416 th->th.th_sleep_loc_type = flag_unset; 1417 __kmp_unlock_suspend_mx(th); 1418 return; 1419 } 1420 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1421 " was %x\n", 1422 th_gtid, flag->get(), flag->load(), old_spin)); 1423 1424 if (flag->done_check_val(old_spin) || flag->done_check()) { 1425 flag->unset_sleeping(); 1426 TCW_PTR(th->th.th_sleep_loc, NULL); 1427 th->th.th_sleep_loc_type = flag_unset; 1428 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1429 "for spin(%p)\n", 1430 th_gtid, flag->get())); 1431 } else { 1432 /* Encapsulate in a loop as the documentation states that this may 1433 "with low probability" return when the condition variable has 1434 not been signaled or broadcast */ 1435 int deactivated = FALSE; 1436 1437 while (flag->is_sleeping()) { 1438 #ifdef DEBUG_SUSPEND 1439 char buffer[128]; 1440 __kmp_suspend_count++; 1441 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1442 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid, 1443 buffer); 1444 #endif 1445 // Mark the thread as no longer active (only in the first iteration of the 1446 // loop). 1447 if (!deactivated) { 1448 th->th.th_active = FALSE; 1449 if (th->th.th_active_in_pool) { 1450 th->th.th_active_in_pool = FALSE; 1451 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); 1452 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 1453 } 1454 deactivated = TRUE; 1455 } 1456 1457 KMP_DEBUG_ASSERT(th->th.th_sleep_loc); 1458 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type); 1459 1460 #if USE_SUSPEND_TIMEOUT 1461 struct timespec now; 1462 struct timeval tval; 1463 int msecs; 1464 1465 status = gettimeofday(&tval, NULL); 1466 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1467 TIMEVAL_TO_TIMESPEC(&tval, &now); 1468 1469 msecs = (4 * __kmp_dflt_blocktime) + 200; 1470 now.tv_sec += msecs / 1000; 1471 now.tv_nsec += (msecs % 1000) * 1000; 1472 1473 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " 1474 "pthread_cond_timedwait\n", 1475 th_gtid)); 1476 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond, 1477 &th->th.th_suspend_mx.m_mutex, &now); 1478 #else 1479 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform" 1480 " pthread_cond_wait\n", 1481 th_gtid)); 1482 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond, 1483 &th->th.th_suspend_mx.m_mutex); 1484 #endif // USE_SUSPEND_TIMEOUT 1485 1486 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) { 1487 KMP_SYSFAIL("pthread_cond_wait", status); 1488 } 1489 1490 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type()); 1491 1492 if (!flag->is_sleeping() && 1493 ((status == EINTR) || (status == ETIMEDOUT))) { 1494 // if interrupt or timeout, and thread is no longer sleeping, we need to 1495 // make sure sleep_loc gets reset; however, this shouldn't be needed if 1496 // we woke up with resume 1497 flag->unset_sleeping(); 1498 TCW_PTR(th->th.th_sleep_loc, NULL); 1499 th->th.th_sleep_loc_type = flag_unset; 1500 } 1501 #ifdef KMP_DEBUG 1502 if (status == ETIMEDOUT) { 1503 if (flag->is_sleeping()) { 1504 KF_TRACE(100, 1505 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid)); 1506 } else { 1507 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1508 "not set!\n", 1509 th_gtid)); 1510 TCW_PTR(th->th.th_sleep_loc, NULL); 1511 th->th.th_sleep_loc_type = flag_unset; 1512 } 1513 } else if (flag->is_sleeping()) { 1514 KF_TRACE(100, 1515 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid)); 1516 } 1517 #endif 1518 } // while 1519 1520 // Mark the thread as active again (if it was previous marked as inactive) 1521 if (deactivated) { 1522 th->th.th_active = TRUE; 1523 if (TCR_4(th->th.th_in_pool)) { 1524 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); 1525 th->th.th_active_in_pool = TRUE; 1526 } 1527 } 1528 } 1529 // We may have had the loop variable set before entering the loop body; 1530 // so we need to reset sleep_loc. 1531 TCW_PTR(th->th.th_sleep_loc, NULL); 1532 th->th.th_sleep_loc_type = flag_unset; 1533 1534 KMP_DEBUG_ASSERT(!flag->is_sleeping()); 1535 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc); 1536 #ifdef DEBUG_SUSPEND 1537 { 1538 char buffer[128]; 1539 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1540 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid, 1541 buffer); 1542 } 1543 #endif 1544 1545 __kmp_unlock_suspend_mx(th); 1546 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid)); 1547 } 1548 1549 template <bool C, bool S> 1550 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) { 1551 __kmp_suspend_template(th_gtid, flag); 1552 } 1553 template <bool C, bool S> 1554 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) { 1555 __kmp_suspend_template(th_gtid, flag); 1556 } 1557 template <bool C, bool S> 1558 void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) { 1559 __kmp_suspend_template(th_gtid, flag); 1560 } 1561 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { 1562 __kmp_suspend_template(th_gtid, flag); 1563 } 1564 1565 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *); 1566 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *); 1567 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *); 1568 template void 1569 __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1570 template void 1571 __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *); 1572 1573 /* This routine signals the thread specified by target_gtid to wake up 1574 after setting the sleep bit indicated by the flag argument to FALSE. 1575 The target thread must already have called __kmp_suspend_template() */ 1576 template <class C> 1577 static inline void __kmp_resume_template(int target_gtid, C *flag) { 1578 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1579 kmp_info_t *th = __kmp_threads[target_gtid]; 1580 int status; 1581 1582 #ifdef KMP_DEBUG 1583 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1584 #endif 1585 1586 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n", 1587 gtid, target_gtid)); 1588 KMP_DEBUG_ASSERT(gtid != target_gtid); 1589 1590 __kmp_suspend_initialize_thread(th); 1591 1592 __kmp_lock_suspend_mx(th); 1593 1594 if (!flag || flag != th->th.th_sleep_loc) { 1595 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a 1596 // different location; wake up at new location 1597 flag = (C *)CCAST(void *, th->th.th_sleep_loc); 1598 } 1599 1600 // First, check if the flag is null or its type has changed. If so, someone 1601 // else woke it up. 1602 if (!flag) { // Thread doesn't appear to be sleeping on anything 1603 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1604 "awake: flag(%p)\n", 1605 gtid, target_gtid, (void *)NULL)); 1606 __kmp_unlock_suspend_mx(th); 1607 return; 1608 } else if (flag->get_type() != th->th.th_sleep_loc_type) { 1609 // Flag type does not appear to match this function template; possibly the 1610 // thread is sleeping on something else. Try null resume again. 1611 KF_TRACE( 1612 5, 1613 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), " 1614 "spin(%p) type=%d ptr_type=%d\n", 1615 gtid, target_gtid, flag, flag->get(), flag->get_type(), 1616 th->th.th_sleep_loc_type)); 1617 __kmp_unlock_suspend_mx(th); 1618 __kmp_null_resume_wrapper(th); 1619 return; 1620 } else { // if multiple threads are sleeping, flag should be internally 1621 // referring to a specific thread here 1622 if (!flag->is_sleeping()) { 1623 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1624 "awake: flag(%p): %u\n", 1625 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1626 __kmp_unlock_suspend_mx(th); 1627 return; 1628 } 1629 } 1630 KMP_DEBUG_ASSERT(flag); 1631 flag->unset_sleeping(); 1632 TCW_PTR(th->th.th_sleep_loc, NULL); 1633 th->th.th_sleep_loc_type = flag_unset; 1634 1635 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1636 "sleep bit for flag's loc(%p): %u\n", 1637 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1638 1639 #ifdef DEBUG_SUSPEND 1640 { 1641 char buffer[128]; 1642 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1643 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid, 1644 target_gtid, buffer); 1645 } 1646 #endif 1647 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond); 1648 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1649 __kmp_unlock_suspend_mx(th); 1650 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" 1651 " for T#%d\n", 1652 gtid, target_gtid)); 1653 } 1654 1655 template <bool C, bool S> 1656 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) { 1657 __kmp_resume_template(target_gtid, flag); 1658 } 1659 template <bool C, bool S> 1660 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) { 1661 __kmp_resume_template(target_gtid, flag); 1662 } 1663 template <bool C, bool S> 1664 void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) { 1665 __kmp_resume_template(target_gtid, flag); 1666 } 1667 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { 1668 __kmp_resume_template(target_gtid, flag); 1669 } 1670 1671 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *); 1672 template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *); 1673 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *); 1674 template void 1675 __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1676 1677 #if KMP_USE_MONITOR 1678 void __kmp_resume_monitor() { 1679 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1680 int status; 1681 #ifdef KMP_DEBUG 1682 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1683 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid, 1684 KMP_GTID_MONITOR)); 1685 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR); 1686 #endif 1687 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 1688 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1689 #ifdef DEBUG_SUSPEND 1690 { 1691 char buffer[128]; 1692 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond); 1693 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid, 1694 KMP_GTID_MONITOR, buffer); 1695 } 1696 #endif 1697 status = pthread_cond_signal(&__kmp_wait_cv.c_cond); 1698 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1699 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 1700 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1701 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up" 1702 " for T#%d\n", 1703 gtid, KMP_GTID_MONITOR)); 1704 } 1705 #endif // KMP_USE_MONITOR 1706 1707 void __kmp_yield() { sched_yield(); } 1708 1709 void __kmp_gtid_set_specific(int gtid) { 1710 if (__kmp_init_gtid) { 1711 int status; 1712 status = pthread_setspecific(__kmp_gtid_threadprivate_key, 1713 (void *)(intptr_t)(gtid + 1)); 1714 KMP_CHECK_SYSFAIL("pthread_setspecific", status); 1715 } else { 1716 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n")); 1717 } 1718 } 1719 1720 int __kmp_gtid_get_specific() { 1721 int gtid; 1722 if (!__kmp_init_gtid) { 1723 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " 1724 "KMP_GTID_SHUTDOWN\n")); 1725 return KMP_GTID_SHUTDOWN; 1726 } 1727 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key); 1728 if (gtid == 0) { 1729 gtid = KMP_GTID_DNE; 1730 } else { 1731 gtid--; 1732 } 1733 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n", 1734 __kmp_gtid_threadprivate_key, gtid)); 1735 return gtid; 1736 } 1737 1738 double __kmp_read_cpu_time(void) { 1739 /*clock_t t;*/ 1740 struct tms buffer; 1741 1742 /*t =*/times(&buffer); 1743 1744 return (double)(buffer.tms_utime + buffer.tms_cutime) / 1745 (double)CLOCKS_PER_SEC; 1746 } 1747 1748 int __kmp_read_system_info(struct kmp_sys_info *info) { 1749 int status; 1750 struct rusage r_usage; 1751 1752 memset(info, 0, sizeof(*info)); 1753 1754 status = getrusage(RUSAGE_SELF, &r_usage); 1755 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status); 1756 1757 // The maximum resident set size utilized (in kilobytes) 1758 info->maxrss = r_usage.ru_maxrss; 1759 // The number of page faults serviced without any I/O 1760 info->minflt = r_usage.ru_minflt; 1761 // The number of page faults serviced that required I/O 1762 info->majflt = r_usage.ru_majflt; 1763 // The number of times a process was "swapped" out of memory 1764 info->nswap = r_usage.ru_nswap; 1765 // The number of times the file system had to perform input 1766 info->inblock = r_usage.ru_inblock; 1767 // The number of times the file system had to perform output 1768 info->oublock = r_usage.ru_oublock; 1769 // The number of times a context switch was voluntarily 1770 info->nvcsw = r_usage.ru_nvcsw; 1771 // The number of times a context switch was forced 1772 info->nivcsw = r_usage.ru_nivcsw; 1773 1774 return (status != 0); 1775 } 1776 1777 void __kmp_read_system_time(double *delta) { 1778 double t_ns; 1779 struct timeval tval; 1780 struct timespec stop; 1781 int status; 1782 1783 status = gettimeofday(&tval, NULL); 1784 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1785 TIMEVAL_TO_TIMESPEC(&tval, &stop); 1786 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start)); 1787 *delta = (t_ns * 1e-9); 1788 } 1789 1790 void __kmp_clear_system_time(void) { 1791 struct timeval tval; 1792 int status; 1793 status = gettimeofday(&tval, NULL); 1794 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1795 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start); 1796 } 1797 1798 static int __kmp_get_xproc(void) { 1799 1800 int r = 0; 1801 1802 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 1803 KMP_OS_OPENBSD || KMP_OS_HURD 1804 1805 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r)); 1806 1807 #elif KMP_OS_DARWIN 1808 1809 // Bug C77011 High "OpenMP Threads and number of active cores". 1810 1811 // Find the number of available CPUs. 1812 kern_return_t rc; 1813 host_basic_info_data_t info; 1814 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT; 1815 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num); 1816 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) { 1817 // Cannot use KA_TRACE() here because this code works before trace support 1818 // is initialized. 1819 r = info.avail_cpus; 1820 } else { 1821 KMP_WARNING(CantGetNumAvailCPU); 1822 KMP_INFORM(AssumedNumCPU); 1823 } 1824 1825 #else 1826 1827 #error "Unknown or unsupported OS." 1828 1829 #endif 1830 1831 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */ 1832 1833 } // __kmp_get_xproc 1834 1835 int __kmp_read_from_file(char const *path, char const *format, ...) { 1836 int result; 1837 va_list args; 1838 1839 va_start(args, format); 1840 FILE *f = fopen(path, "rb"); 1841 if (f == NULL) 1842 return 0; 1843 result = vfscanf(f, format, args); 1844 fclose(f); 1845 1846 return result; 1847 } 1848 1849 void __kmp_runtime_initialize(void) { 1850 int status; 1851 pthread_mutexattr_t mutex_attr; 1852 pthread_condattr_t cond_attr; 1853 1854 if (__kmp_init_runtime) { 1855 return; 1856 } 1857 1858 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1859 if (!__kmp_cpuinfo.initialized) { 1860 __kmp_query_cpuid(&__kmp_cpuinfo); 1861 } 1862 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1863 1864 __kmp_xproc = __kmp_get_xproc(); 1865 1866 #if !KMP_32_BIT_ARCH 1867 struct rlimit rlim; 1868 // read stack size of calling thread, save it as default for worker threads; 1869 // this should be done before reading environment variables 1870 status = getrlimit(RLIMIT_STACK, &rlim); 1871 if (status == 0) { // success? 1872 __kmp_stksize = rlim.rlim_cur; 1873 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed 1874 } 1875 #endif /* KMP_32_BIT_ARCH */ 1876 1877 if (sysconf(_SC_THREADS)) { 1878 1879 /* Query the maximum number of threads */ 1880 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth)); 1881 if (__kmp_sys_max_nth == -1) { 1882 /* Unlimited threads for NPTL */ 1883 __kmp_sys_max_nth = INT_MAX; 1884 } else if (__kmp_sys_max_nth <= 1) { 1885 /* Can't tell, just use PTHREAD_THREADS_MAX */ 1886 __kmp_sys_max_nth = KMP_MAX_NTH; 1887 } 1888 1889 /* Query the minimum stack size */ 1890 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN); 1891 if (__kmp_sys_min_stksize <= 1) { 1892 __kmp_sys_min_stksize = KMP_MIN_STKSIZE; 1893 } 1894 } 1895 1896 /* Set up minimum number of threads to switch to TLS gtid */ 1897 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; 1898 1899 status = pthread_key_create(&__kmp_gtid_threadprivate_key, 1900 __kmp_internal_end_dest); 1901 KMP_CHECK_SYSFAIL("pthread_key_create", status); 1902 status = pthread_mutexattr_init(&mutex_attr); 1903 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1904 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr); 1905 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1906 status = pthread_mutexattr_destroy(&mutex_attr); 1907 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status); 1908 status = pthread_condattr_init(&cond_attr); 1909 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1910 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr); 1911 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1912 status = pthread_condattr_destroy(&cond_attr); 1913 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status); 1914 #if USE_ITT_BUILD 1915 __kmp_itt_initialize(); 1916 #endif /* USE_ITT_BUILD */ 1917 1918 __kmp_init_runtime = TRUE; 1919 } 1920 1921 void __kmp_runtime_destroy(void) { 1922 int status; 1923 1924 if (!__kmp_init_runtime) { 1925 return; // Nothing to do. 1926 } 1927 1928 #if USE_ITT_BUILD 1929 __kmp_itt_destroy(); 1930 #endif /* USE_ITT_BUILD */ 1931 1932 status = pthread_key_delete(__kmp_gtid_threadprivate_key); 1933 KMP_CHECK_SYSFAIL("pthread_key_delete", status); 1934 1935 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex); 1936 if (status != 0 && status != EBUSY) { 1937 KMP_SYSFAIL("pthread_mutex_destroy", status); 1938 } 1939 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond); 1940 if (status != 0 && status != EBUSY) { 1941 KMP_SYSFAIL("pthread_cond_destroy", status); 1942 } 1943 #if KMP_AFFINITY_SUPPORTED 1944 __kmp_affinity_uninitialize(); 1945 #endif 1946 1947 __kmp_init_runtime = FALSE; 1948 } 1949 1950 /* Put the thread to sleep for a time period */ 1951 /* NOTE: not currently used anywhere */ 1952 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); } 1953 1954 /* Calculate the elapsed wall clock time for the user */ 1955 void __kmp_elapsed(double *t) { 1956 int status; 1957 #ifdef FIX_SGI_CLOCK 1958 struct timespec ts; 1959 1960 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); 1961 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status); 1962 *t = 1963 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec; 1964 #else 1965 struct timeval tv; 1966 1967 status = gettimeofday(&tv, NULL); 1968 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1969 *t = 1970 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec; 1971 #endif 1972 } 1973 1974 /* Calculate the elapsed wall clock tick for the user */ 1975 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; } 1976 1977 /* Return the current time stamp in nsec */ 1978 kmp_uint64 __kmp_now_nsec() { 1979 struct timeval t; 1980 gettimeofday(&t, NULL); 1981 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec + 1982 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec; 1983 return nsec; 1984 } 1985 1986 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1987 /* Measure clock ticks per millisecond */ 1988 void __kmp_initialize_system_tick() { 1989 kmp_uint64 now, nsec2, diff; 1990 kmp_uint64 delay = 100000; // 50~100 usec on most machines. 1991 kmp_uint64 nsec = __kmp_now_nsec(); 1992 kmp_uint64 goal = __kmp_hardware_timestamp() + delay; 1993 while ((now = __kmp_hardware_timestamp()) < goal) 1994 ; 1995 nsec2 = __kmp_now_nsec(); 1996 diff = nsec2 - nsec; 1997 if (diff > 0) { 1998 kmp_uint64 tpms = ((kmp_uint64)1e6 * (delay + (now - goal)) / diff); 1999 if (tpms > 0) 2000 __kmp_ticks_per_msec = tpms; 2001 } 2002 } 2003 #endif 2004 2005 /* Determine whether the given address is mapped into the current address 2006 space. */ 2007 2008 int __kmp_is_address_mapped(void *addr) { 2009 2010 int found = 0; 2011 int rc; 2012 2013 #if KMP_OS_LINUX || KMP_OS_HURD 2014 2015 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the 2016 address ranges mapped into the address space. */ 2017 2018 char *name = __kmp_str_format("/proc/%d/maps", getpid()); 2019 FILE *file = NULL; 2020 2021 file = fopen(name, "r"); 2022 KMP_ASSERT(file != NULL); 2023 2024 for (;;) { 2025 2026 void *beginning = NULL; 2027 void *ending = NULL; 2028 char perms[5]; 2029 2030 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms); 2031 if (rc == EOF) { 2032 break; 2033 } 2034 KMP_ASSERT(rc == 3 && 2035 KMP_STRLEN(perms) == 4); // Make sure all fields are read. 2036 2037 // Ending address is not included in the region, but beginning is. 2038 if ((addr >= beginning) && (addr < ending)) { 2039 perms[2] = 0; // 3th and 4th character does not matter. 2040 if (strcmp(perms, "rw") == 0) { 2041 // Memory we are looking for should be readable and writable. 2042 found = 1; 2043 } 2044 break; 2045 } 2046 } 2047 2048 // Free resources. 2049 fclose(file); 2050 KMP_INTERNAL_FREE(name); 2051 #elif KMP_OS_FREEBSD 2052 char *buf; 2053 size_t lstsz; 2054 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()}; 2055 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0); 2056 if (rc < 0) 2057 return 0; 2058 // We pass from number of vm entry's semantic 2059 // to size of whole entry map list. 2060 lstsz = lstsz * 4 / 3; 2061 buf = reinterpret_cast<char *>(kmpc_malloc(lstsz)); 2062 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0); 2063 if (rc < 0) { 2064 kmpc_free(buf); 2065 return 0; 2066 } 2067 2068 char *lw = buf; 2069 char *up = buf + lstsz; 2070 2071 while (lw < up) { 2072 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw); 2073 size_t cursz = cur->kve_structsize; 2074 if (cursz == 0) 2075 break; 2076 void *start = reinterpret_cast<void *>(cur->kve_start); 2077 void *end = reinterpret_cast<void *>(cur->kve_end); 2078 // Readable/Writable addresses within current map entry 2079 if ((addr >= start) && (addr < end)) { 2080 if ((cur->kve_protection & KVME_PROT_READ) != 0 && 2081 (cur->kve_protection & KVME_PROT_WRITE) != 0) { 2082 found = 1; 2083 break; 2084 } 2085 } 2086 lw += cursz; 2087 } 2088 kmpc_free(buf); 2089 2090 #elif KMP_OS_DARWIN 2091 2092 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory 2093 using vm interface. */ 2094 2095 int buffer; 2096 vm_size_t count; 2097 rc = vm_read_overwrite( 2098 mach_task_self(), // Task to read memory of. 2099 (vm_address_t)(addr), // Address to read from. 2100 1, // Number of bytes to be read. 2101 (vm_address_t)(&buffer), // Address of buffer to save read bytes in. 2102 &count // Address of var to save number of read bytes in. 2103 ); 2104 if (rc == 0) { 2105 // Memory successfully read. 2106 found = 1; 2107 } 2108 2109 #elif KMP_OS_NETBSD 2110 2111 int mib[5]; 2112 mib[0] = CTL_VM; 2113 mib[1] = VM_PROC; 2114 mib[2] = VM_PROC_MAP; 2115 mib[3] = getpid(); 2116 mib[4] = sizeof(struct kinfo_vmentry); 2117 2118 size_t size; 2119 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0); 2120 KMP_ASSERT(!rc); 2121 KMP_ASSERT(size); 2122 2123 size = size * 4 / 3; 2124 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size); 2125 KMP_ASSERT(kiv); 2126 2127 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0); 2128 KMP_ASSERT(!rc); 2129 KMP_ASSERT(size); 2130 2131 for (size_t i = 0; i < size; i++) { 2132 if (kiv[i].kve_start >= (uint64_t)addr && 2133 kiv[i].kve_end <= (uint64_t)addr) { 2134 found = 1; 2135 break; 2136 } 2137 } 2138 KMP_INTERNAL_FREE(kiv); 2139 #elif KMP_OS_OPENBSD 2140 2141 int mib[3]; 2142 mib[0] = CTL_KERN; 2143 mib[1] = KERN_PROC_VMMAP; 2144 mib[2] = getpid(); 2145 2146 size_t size; 2147 uint64_t end; 2148 rc = sysctl(mib, 3, NULL, &size, NULL, 0); 2149 KMP_ASSERT(!rc); 2150 KMP_ASSERT(size); 2151 end = size; 2152 2153 struct kinfo_vmentry kiv = {.kve_start = 0}; 2154 2155 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) { 2156 KMP_ASSERT(size); 2157 if (kiv.kve_end == end) 2158 break; 2159 2160 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) { 2161 found = 1; 2162 break; 2163 } 2164 kiv.kve_start += 1; 2165 } 2166 #elif KMP_OS_DRAGONFLY 2167 2168 // FIXME(DragonFly): Implement this 2169 found = 1; 2170 2171 #else 2172 2173 #error "Unknown or unsupported OS" 2174 2175 #endif 2176 2177 return found; 2178 2179 } // __kmp_is_address_mapped 2180 2181 #ifdef USE_LOAD_BALANCE 2182 2183 #if KMP_OS_DARWIN || KMP_OS_NETBSD 2184 2185 // The function returns the rounded value of the system load average 2186 // during given time interval which depends on the value of 2187 // __kmp_load_balance_interval variable (default is 60 sec, other values 2188 // may be 300 sec or 900 sec). 2189 // It returns -1 in case of error. 2190 int __kmp_get_load_balance(int max) { 2191 double averages[3]; 2192 int ret_avg = 0; 2193 2194 int res = getloadavg(averages, 3); 2195 2196 // Check __kmp_load_balance_interval to determine which of averages to use. 2197 // getloadavg() may return the number of samples less than requested that is 2198 // less than 3. 2199 if (__kmp_load_balance_interval < 180 && (res >= 1)) { 2200 ret_avg = (int)averages[0]; // 1 min 2201 } else if ((__kmp_load_balance_interval >= 180 && 2202 __kmp_load_balance_interval < 600) && 2203 (res >= 2)) { 2204 ret_avg = (int)averages[1]; // 5 min 2205 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) { 2206 ret_avg = (int)averages[2]; // 15 min 2207 } else { // Error occurred 2208 return -1; 2209 } 2210 2211 return ret_avg; 2212 } 2213 2214 #else // Linux* OS 2215 2216 // The function returns number of running (not sleeping) threads, or -1 in case 2217 // of error. Error could be reported if Linux* OS kernel too old (without 2218 // "/proc" support). Counting running threads stops if max running threads 2219 // encountered. 2220 int __kmp_get_load_balance(int max) { 2221 static int permanent_error = 0; 2222 static int glb_running_threads = 0; // Saved count of the running threads for 2223 // the thread balance algorithm 2224 static double glb_call_time = 0; /* Thread balance algorithm call time */ 2225 2226 int running_threads = 0; // Number of running threads in the system. 2227 2228 DIR *proc_dir = NULL; // Handle of "/proc/" directory. 2229 struct dirent *proc_entry = NULL; 2230 2231 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path. 2232 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory. 2233 struct dirent *task_entry = NULL; 2234 int task_path_fixed_len; 2235 2236 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path. 2237 int stat_file = -1; 2238 int stat_path_fixed_len; 2239 2240 int total_processes = 0; // Total number of processes in system. 2241 int total_threads = 0; // Total number of threads in system. 2242 2243 double call_time = 0.0; 2244 2245 __kmp_str_buf_init(&task_path); 2246 __kmp_str_buf_init(&stat_path); 2247 2248 __kmp_elapsed(&call_time); 2249 2250 if (glb_call_time && 2251 (call_time - glb_call_time < __kmp_load_balance_interval)) { 2252 running_threads = glb_running_threads; 2253 goto finish; 2254 } 2255 2256 glb_call_time = call_time; 2257 2258 // Do not spend time on scanning "/proc/" if we have a permanent error. 2259 if (permanent_error) { 2260 running_threads = -1; 2261 goto finish; 2262 } 2263 2264 if (max <= 0) { 2265 max = INT_MAX; 2266 } 2267 2268 // Open "/proc/" directory. 2269 proc_dir = opendir("/proc"); 2270 if (proc_dir == NULL) { 2271 // Cannot open "/prroc/". Probably the kernel does not support it. Return an 2272 // error now and in subsequent calls. 2273 running_threads = -1; 2274 permanent_error = 1; 2275 goto finish; 2276 } 2277 2278 // Initialize fixed part of task_path. This part will not change. 2279 __kmp_str_buf_cat(&task_path, "/proc/", 6); 2280 task_path_fixed_len = task_path.used; // Remember number of used characters. 2281 2282 proc_entry = readdir(proc_dir); 2283 while (proc_entry != NULL) { 2284 // Proc entry is a directory and name starts with a digit. Assume it is a 2285 // process' directory. 2286 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) { 2287 2288 ++total_processes; 2289 // Make sure init process is the very first in "/proc", so we can replace 2290 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes == 2291 // 1. We are going to check that total_processes == 1 => d_name == "1" is 2292 // true (where "=>" is implication). Since C++ does not have => operator, 2293 // let us replace it with its equivalent: a => b == ! a || b. 2294 KMP_DEBUG_ASSERT(total_processes != 1 || 2295 strcmp(proc_entry->d_name, "1") == 0); 2296 2297 // Construct task_path. 2298 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/". 2299 __kmp_str_buf_cat(&task_path, proc_entry->d_name, 2300 KMP_STRLEN(proc_entry->d_name)); 2301 __kmp_str_buf_cat(&task_path, "/task", 5); 2302 2303 task_dir = opendir(task_path.str); 2304 if (task_dir == NULL) { 2305 // Process can finish between reading "/proc/" directory entry and 2306 // opening process' "task/" directory. So, in general case we should not 2307 // complain, but have to skip this process and read the next one. But on 2308 // systems with no "task/" support we will spend lot of time to scan 2309 // "/proc/" tree again and again without any benefit. "init" process 2310 // (its pid is 1) should exist always, so, if we cannot open 2311 // "/proc/1/task/" directory, it means "task/" is not supported by 2312 // kernel. Report an error now and in the future. 2313 if (strcmp(proc_entry->d_name, "1") == 0) { 2314 running_threads = -1; 2315 permanent_error = 1; 2316 goto finish; 2317 } 2318 } else { 2319 // Construct fixed part of stat file path. 2320 __kmp_str_buf_clear(&stat_path); 2321 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used); 2322 __kmp_str_buf_cat(&stat_path, "/", 1); 2323 stat_path_fixed_len = stat_path.used; 2324 2325 task_entry = readdir(task_dir); 2326 while (task_entry != NULL) { 2327 // It is a directory and name starts with a digit. 2328 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) { 2329 ++total_threads; 2330 2331 // Construct complete stat file path. Easiest way would be: 2332 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str, 2333 // task_entry->d_name ); 2334 // but seriae of __kmp_str_buf_cat works a bit faster. 2335 stat_path.used = 2336 stat_path_fixed_len; // Reset stat path to its fixed part. 2337 __kmp_str_buf_cat(&stat_path, task_entry->d_name, 2338 KMP_STRLEN(task_entry->d_name)); 2339 __kmp_str_buf_cat(&stat_path, "/stat", 5); 2340 2341 // Note: Low-level API (open/read/close) is used. High-level API 2342 // (fopen/fclose) works ~ 30 % slower. 2343 stat_file = open(stat_path.str, O_RDONLY); 2344 if (stat_file == -1) { 2345 // We cannot report an error because task (thread) can terminate 2346 // just before reading this file. 2347 } else { 2348 /* Content of "stat" file looks like: 2349 24285 (program) S ... 2350 2351 It is a single line (if program name does not include funny 2352 symbols). First number is a thread id, then name of executable 2353 file name in paretheses, then state of the thread. We need just 2354 thread state. 2355 2356 Good news: Length of program name is 15 characters max. Longer 2357 names are truncated. 2358 2359 Thus, we need rather short buffer: 15 chars for program name + 2360 2 parenthesis, + 3 spaces + ~7 digits of pid = 37. 2361 2362 Bad news: Program name may contain special symbols like space, 2363 closing parenthesis, or even new line. This makes parsing 2364 "stat" file not 100 % reliable. In case of fanny program names 2365 parsing may fail (report incorrect thread state). 2366 2367 Parsing "status" file looks more promissing (due to different 2368 file structure and escaping special symbols) but reading and 2369 parsing of "status" file works slower. 2370 -- ln 2371 */ 2372 char buffer[65]; 2373 ssize_t len; 2374 len = read(stat_file, buffer, sizeof(buffer) - 1); 2375 if (len >= 0) { 2376 buffer[len] = 0; 2377 // Using scanf: 2378 // sscanf( buffer, "%*d (%*s) %c ", & state ); 2379 // looks very nice, but searching for a closing parenthesis 2380 // works a bit faster. 2381 char *close_parent = strstr(buffer, ") "); 2382 if (close_parent != NULL) { 2383 char state = *(close_parent + 2); 2384 if (state == 'R') { 2385 ++running_threads; 2386 if (running_threads >= max) { 2387 goto finish; 2388 } 2389 } 2390 } 2391 } 2392 close(stat_file); 2393 stat_file = -1; 2394 } 2395 } 2396 task_entry = readdir(task_dir); 2397 } 2398 closedir(task_dir); 2399 task_dir = NULL; 2400 } 2401 } 2402 proc_entry = readdir(proc_dir); 2403 } 2404 2405 // There _might_ be a timing hole where the thread executing this 2406 // code get skipped in the load balance, and running_threads is 0. 2407 // Assert in the debug builds only!!! 2408 KMP_DEBUG_ASSERT(running_threads > 0); 2409 if (running_threads <= 0) { 2410 running_threads = 1; 2411 } 2412 2413 finish: // Clean up and exit. 2414 if (proc_dir != NULL) { 2415 closedir(proc_dir); 2416 } 2417 __kmp_str_buf_free(&task_path); 2418 if (task_dir != NULL) { 2419 closedir(task_dir); 2420 } 2421 __kmp_str_buf_free(&stat_path); 2422 if (stat_file != -1) { 2423 close(stat_file); 2424 } 2425 2426 glb_running_threads = running_threads; 2427 2428 return running_threads; 2429 2430 } // __kmp_get_load_balance 2431 2432 #endif // KMP_OS_DARWIN 2433 2434 #endif // USE_LOAD_BALANCE 2435 2436 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2437 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \ 2438 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64) 2439 2440 // we really only need the case with 1 argument, because CLANG always build 2441 // a struct of pointers to shared variables referenced in the outlined function 2442 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc, 2443 void *p_argv[] 2444 #if OMPT_SUPPORT 2445 , 2446 void **exit_frame_ptr 2447 #endif 2448 ) { 2449 #if OMPT_SUPPORT 2450 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0); 2451 #endif 2452 2453 switch (argc) { 2454 default: 2455 fprintf(stderr, "Too many args to microtask: %d!\n", argc); 2456 fflush(stderr); 2457 exit(-1); 2458 case 0: 2459 (*pkfn)(>id, &tid); 2460 break; 2461 case 1: 2462 (*pkfn)(>id, &tid, p_argv[0]); 2463 break; 2464 case 2: 2465 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]); 2466 break; 2467 case 3: 2468 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]); 2469 break; 2470 case 4: 2471 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]); 2472 break; 2473 case 5: 2474 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]); 2475 break; 2476 case 6: 2477 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2478 p_argv[5]); 2479 break; 2480 case 7: 2481 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2482 p_argv[5], p_argv[6]); 2483 break; 2484 case 8: 2485 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2486 p_argv[5], p_argv[6], p_argv[7]); 2487 break; 2488 case 9: 2489 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2490 p_argv[5], p_argv[6], p_argv[7], p_argv[8]); 2491 break; 2492 case 10: 2493 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2494 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]); 2495 break; 2496 case 11: 2497 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2498 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]); 2499 break; 2500 case 12: 2501 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2502 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2503 p_argv[11]); 2504 break; 2505 case 13: 2506 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2507 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2508 p_argv[11], p_argv[12]); 2509 break; 2510 case 14: 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 p_argv[11], p_argv[12], p_argv[13]); 2514 break; 2515 case 15: 2516 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2517 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2518 p_argv[11], p_argv[12], p_argv[13], p_argv[14]); 2519 break; 2520 } 2521 2522 return 1; 2523 } 2524 2525 #endif 2526 2527 #if KMP_OS_LINUX 2528 // Functions for hidden helper task 2529 namespace { 2530 // Condition variable for initializing hidden helper team 2531 pthread_cond_t hidden_helper_threads_initz_cond_var; 2532 pthread_mutex_t hidden_helper_threads_initz_lock; 2533 volatile int hidden_helper_initz_signaled = FALSE; 2534 2535 // Condition variable for deinitializing hidden helper team 2536 pthread_cond_t hidden_helper_threads_deinitz_cond_var; 2537 pthread_mutex_t hidden_helper_threads_deinitz_lock; 2538 volatile int hidden_helper_deinitz_signaled = FALSE; 2539 2540 // Condition variable for the wrapper function of main thread 2541 pthread_cond_t hidden_helper_main_thread_cond_var; 2542 pthread_mutex_t hidden_helper_main_thread_lock; 2543 volatile int hidden_helper_main_thread_signaled = FALSE; 2544 2545 // Semaphore for worker threads. We don't use condition variable here in case 2546 // that when multiple signals are sent at the same time, only one thread might 2547 // be waken. 2548 sem_t hidden_helper_task_sem; 2549 } // namespace 2550 2551 void __kmp_hidden_helper_worker_thread_wait() { 2552 int status = sem_wait(&hidden_helper_task_sem); 2553 KMP_CHECK_SYSFAIL("sem_wait", status); 2554 } 2555 2556 void __kmp_do_initialize_hidden_helper_threads() { 2557 // Initialize condition variable 2558 int status = 2559 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr); 2560 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2561 2562 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr); 2563 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2564 2565 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr); 2566 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2567 2568 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr); 2569 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2570 2571 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr); 2572 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2573 2574 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr); 2575 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2576 2577 // Initialize the semaphore 2578 status = sem_init(&hidden_helper_task_sem, 0, 0); 2579 KMP_CHECK_SYSFAIL("sem_init", status); 2580 2581 // Create a new thread to finish initialization 2582 pthread_t handle; 2583 status = pthread_create( 2584 &handle, nullptr, 2585 [](void *) -> void * { 2586 __kmp_hidden_helper_threads_initz_routine(); 2587 return nullptr; 2588 }, 2589 nullptr); 2590 KMP_CHECK_SYSFAIL("pthread_create", status); 2591 } 2592 2593 void __kmp_hidden_helper_threads_initz_wait() { 2594 // Initial thread waits here for the completion of the initialization. The 2595 // condition variable will be notified by main thread of hidden helper teams. 2596 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2597 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2598 2599 if (!TCR_4(hidden_helper_initz_signaled)) { 2600 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var, 2601 &hidden_helper_threads_initz_lock); 2602 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2603 } 2604 2605 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2606 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2607 } 2608 2609 void __kmp_hidden_helper_initz_release() { 2610 // After all initialization, reset __kmp_init_hidden_helper_threads to false. 2611 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2612 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2613 2614 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var); 2615 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2616 2617 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE); 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_main_thread_wait() { 2624 // The main thread of hidden helper team will be blocked here. The 2625 // condition variable can only be signal in the destructor of RTL. 2626 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2627 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2628 2629 if (!TCR_4(hidden_helper_main_thread_signaled)) { 2630 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var, 2631 &hidden_helper_main_thread_lock); 2632 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2633 } 2634 2635 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2636 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2637 } 2638 2639 void __kmp_hidden_helper_main_thread_release() { 2640 // The initial thread of OpenMP RTL should call this function to wake up the 2641 // main thread of hidden helper team. 2642 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2643 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2644 2645 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var); 2646 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 2647 2648 // The hidden helper team is done here 2649 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE); 2650 2651 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2652 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2653 } 2654 2655 void __kmp_hidden_helper_worker_thread_signal() { 2656 int status = sem_post(&hidden_helper_task_sem); 2657 KMP_CHECK_SYSFAIL("sem_post", status); 2658 } 2659 2660 void __kmp_hidden_helper_threads_deinitz_wait() { 2661 // Initial thread waits here for the completion of the deinitialization. The 2662 // condition variable will be notified by main thread of hidden helper teams. 2663 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2664 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2665 2666 if (!TCR_4(hidden_helper_deinitz_signaled)) { 2667 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var, 2668 &hidden_helper_threads_deinitz_lock); 2669 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2670 } 2671 2672 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2673 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2674 } 2675 2676 void __kmp_hidden_helper_threads_deinitz_release() { 2677 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2678 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2679 2680 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var); 2681 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2682 2683 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE); 2684 2685 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2686 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2687 } 2688 #else // KMP_OS_LINUX 2689 void __kmp_hidden_helper_worker_thread_wait() { 2690 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2691 } 2692 2693 void __kmp_do_initialize_hidden_helper_threads() { 2694 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2695 } 2696 2697 void __kmp_hidden_helper_threads_initz_wait() { 2698 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2699 } 2700 2701 void __kmp_hidden_helper_initz_release() { 2702 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2703 } 2704 2705 void __kmp_hidden_helper_main_thread_wait() { 2706 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2707 } 2708 2709 void __kmp_hidden_helper_main_thread_release() { 2710 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2711 } 2712 2713 void __kmp_hidden_helper_worker_thread_signal() { 2714 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2715 } 2716 2717 void __kmp_hidden_helper_threads_deinitz_wait() { 2718 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2719 } 2720 2721 void __kmp_hidden_helper_threads_deinitz_release() { 2722 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2723 } 2724 #endif // KMP_OS_LINUX 2725 2726 // end of file // 2727