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 #else 1055 (void)status; // unused variable 1056 #endif /* KMP_DEBUG */ 1057 1058 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n", 1059 th->th.th_info.ds.ds_gtid)); 1060 1061 KMP_MB(); /* Flush all pending memory write invalidates. */ 1062 } 1063 1064 #if KMP_HANDLE_SIGNALS 1065 1066 static void __kmp_null_handler(int signo) { 1067 // Do nothing, for doing SIG_IGN-type actions. 1068 } // __kmp_null_handler 1069 1070 static void __kmp_team_handler(int signo) { 1071 if (__kmp_global.g.g_abort == 0) { 1072 /* Stage 1 signal handler, let's shut down all of the threads */ 1073 #ifdef KMP_DEBUG 1074 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo); 1075 #endif 1076 switch (signo) { 1077 case SIGHUP: 1078 case SIGINT: 1079 case SIGQUIT: 1080 case SIGILL: 1081 case SIGABRT: 1082 case SIGFPE: 1083 case SIGBUS: 1084 case SIGSEGV: 1085 #ifdef SIGSYS 1086 case SIGSYS: 1087 #endif 1088 case SIGTERM: 1089 if (__kmp_debug_buf) { 1090 __kmp_dump_debug_buffer(); 1091 } 1092 __kmp_unregister_library(); // cleanup shared memory 1093 KMP_MB(); // Flush all pending memory write invalidates. 1094 TCW_4(__kmp_global.g.g_abort, signo); 1095 KMP_MB(); // Flush all pending memory write invalidates. 1096 TCW_4(__kmp_global.g.g_done, TRUE); 1097 KMP_MB(); // Flush all pending memory write invalidates. 1098 break; 1099 default: 1100 #ifdef KMP_DEBUG 1101 __kmp_debug_printf("__kmp_team_handler: unknown signal type"); 1102 #endif 1103 break; 1104 } 1105 } 1106 } // __kmp_team_handler 1107 1108 static void __kmp_sigaction(int signum, const struct sigaction *act, 1109 struct sigaction *oldact) { 1110 int rc = sigaction(signum, act, oldact); 1111 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc); 1112 } 1113 1114 static void __kmp_install_one_handler(int sig, sig_func_t handler_func, 1115 int parallel_init) { 1116 KMP_MB(); // Flush all pending memory write invalidates. 1117 KB_TRACE(60, 1118 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init)); 1119 if (parallel_init) { 1120 struct sigaction new_action; 1121 struct sigaction old_action; 1122 new_action.sa_handler = handler_func; 1123 new_action.sa_flags = 0; 1124 sigfillset(&new_action.sa_mask); 1125 __kmp_sigaction(sig, &new_action, &old_action); 1126 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) { 1127 sigaddset(&__kmp_sigset, sig); 1128 } else { 1129 // Restore/keep user's handler if one previously installed. 1130 __kmp_sigaction(sig, &old_action, NULL); 1131 } 1132 } else { 1133 // Save initial/system signal handlers to see if user handlers installed. 1134 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]); 1135 } 1136 KMP_MB(); // Flush all pending memory write invalidates. 1137 } // __kmp_install_one_handler 1138 1139 static void __kmp_remove_one_handler(int sig) { 1140 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig)); 1141 if (sigismember(&__kmp_sigset, sig)) { 1142 struct sigaction old; 1143 KMP_MB(); // Flush all pending memory write invalidates. 1144 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old); 1145 if ((old.sa_handler != __kmp_team_handler) && 1146 (old.sa_handler != __kmp_null_handler)) { 1147 // Restore the users signal handler. 1148 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, " 1149 "restoring: sig=%d\n", 1150 sig)); 1151 __kmp_sigaction(sig, &old, NULL); 1152 } 1153 sigdelset(&__kmp_sigset, sig); 1154 KMP_MB(); // Flush all pending memory write invalidates. 1155 } 1156 } // __kmp_remove_one_handler 1157 1158 void __kmp_install_signals(int parallel_init) { 1159 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init)); 1160 if (__kmp_handle_signals || !parallel_init) { 1161 // If ! parallel_init, we do not install handlers, just save original 1162 // handlers. Let us do it even __handle_signals is 0. 1163 sigemptyset(&__kmp_sigset); 1164 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init); 1165 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init); 1166 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init); 1167 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init); 1168 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init); 1169 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init); 1170 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init); 1171 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init); 1172 #ifdef SIGSYS 1173 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init); 1174 #endif // SIGSYS 1175 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init); 1176 #ifdef SIGPIPE 1177 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init); 1178 #endif // SIGPIPE 1179 } 1180 } // __kmp_install_signals 1181 1182 void __kmp_remove_signals(void) { 1183 int sig; 1184 KB_TRACE(10, ("__kmp_remove_signals()\n")); 1185 for (sig = 1; sig < NSIG; ++sig) { 1186 __kmp_remove_one_handler(sig); 1187 } 1188 } // __kmp_remove_signals 1189 1190 #endif // KMP_HANDLE_SIGNALS 1191 1192 void __kmp_enable(int new_state) { 1193 #ifdef KMP_CANCEL_THREADS 1194 int status, old_state; 1195 status = pthread_setcancelstate(new_state, &old_state); 1196 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1197 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE); 1198 #endif 1199 } 1200 1201 void __kmp_disable(int *old_state) { 1202 #ifdef KMP_CANCEL_THREADS 1203 int status; 1204 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state); 1205 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1206 #endif 1207 } 1208 1209 static void __kmp_atfork_prepare(void) { 1210 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 1211 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1212 } 1213 1214 static void __kmp_atfork_parent(void) { 1215 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1216 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1217 } 1218 1219 /* Reset the library so execution in the child starts "all over again" with 1220 clean data structures in initial states. Don't worry about freeing memory 1221 allocated by parent, just abandon it to be safe. */ 1222 static void __kmp_atfork_child(void) { 1223 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1224 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1225 /* TODO make sure this is done right for nested/sibling */ 1226 // ATT: Memory leaks are here? TODO: Check it and fix. 1227 /* KMP_ASSERT( 0 ); */ 1228 1229 ++__kmp_fork_count; 1230 1231 #if KMP_AFFINITY_SUPPORTED 1232 #if KMP_OS_LINUX || KMP_OS_FREEBSD 1233 // reset the affinity in the child to the initial thread 1234 // affinity in the parent 1235 kmp_set_thread_affinity_mask_initial(); 1236 #endif 1237 // Set default not to bind threads tightly in the child (we're expecting 1238 // over-subscription after the fork and this can improve things for 1239 // scripting languages that use OpenMP inside process-parallel code). 1240 __kmp_affinity_type = affinity_none; 1241 if (__kmp_nested_proc_bind.bind_types != NULL) { 1242 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false; 1243 } 1244 __kmp_affinity_masks = NULL; 1245 __kmp_affinity_num_masks = 0; 1246 #endif // KMP_AFFINITY_SUPPORTED 1247 1248 #if KMP_USE_MONITOR 1249 __kmp_init_monitor = 0; 1250 #endif 1251 __kmp_init_parallel = FALSE; 1252 __kmp_init_middle = FALSE; 1253 __kmp_init_serial = FALSE; 1254 TCW_4(__kmp_init_gtid, FALSE); 1255 __kmp_init_common = FALSE; 1256 1257 TCW_4(__kmp_init_user_locks, FALSE); 1258 #if !KMP_USE_DYNAMIC_LOCK 1259 __kmp_user_lock_table.used = 1; 1260 __kmp_user_lock_table.allocated = 0; 1261 __kmp_user_lock_table.table = NULL; 1262 __kmp_lock_blocks = NULL; 1263 #endif 1264 1265 __kmp_all_nth = 0; 1266 TCW_4(__kmp_nth, 0); 1267 1268 __kmp_thread_pool = NULL; 1269 __kmp_thread_pool_insert_pt = NULL; 1270 __kmp_team_pool = NULL; 1271 1272 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate 1273 here so threadprivate doesn't use stale data */ 1274 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n", 1275 __kmp_threadpriv_cache_list)); 1276 1277 while (__kmp_threadpriv_cache_list != NULL) { 1278 1279 if (*__kmp_threadpriv_cache_list->addr != NULL) { 1280 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n", 1281 &(*__kmp_threadpriv_cache_list->addr))); 1282 1283 *__kmp_threadpriv_cache_list->addr = NULL; 1284 } 1285 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next; 1286 } 1287 1288 __kmp_init_runtime = FALSE; 1289 1290 /* reset statically initialized locks */ 1291 __kmp_init_bootstrap_lock(&__kmp_initz_lock); 1292 __kmp_init_bootstrap_lock(&__kmp_stdio_lock); 1293 __kmp_init_bootstrap_lock(&__kmp_console_lock); 1294 __kmp_init_bootstrap_lock(&__kmp_task_team_lock); 1295 1296 #if USE_ITT_BUILD 1297 __kmp_itt_reset(); // reset ITT's global state 1298 #endif /* USE_ITT_BUILD */ 1299 1300 __kmp_serial_initialize(); 1301 1302 /* This is necessary to make sure no stale data is left around */ 1303 /* AC: customers complain that we use unsafe routines in the atfork 1304 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen 1305 in dynamic_link when check the presence of shared tbbmalloc library. 1306 Suggestion is to make the library initialization lazier, similar 1307 to what done for __kmpc_begin(). */ 1308 // TODO: synchronize all static initializations with regular library 1309 // startup; look at kmp_global.cpp and etc. 1310 //__kmp_internal_begin (); 1311 } 1312 1313 void __kmp_register_atfork(void) { 1314 if (__kmp_need_register_atfork) { 1315 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent, 1316 __kmp_atfork_child); 1317 KMP_CHECK_SYSFAIL("pthread_atfork", status); 1318 __kmp_need_register_atfork = FALSE; 1319 } 1320 } 1321 1322 void __kmp_suspend_initialize(void) { 1323 int status; 1324 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr); 1325 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1326 status = pthread_condattr_init(&__kmp_suspend_cond_attr); 1327 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1328 } 1329 1330 void __kmp_suspend_initialize_thread(kmp_info_t *th) { 1331 int old_value = KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count); 1332 int new_value = __kmp_fork_count + 1; 1333 // Return if already initialized 1334 if (old_value == new_value) 1335 return; 1336 // Wait, then return if being initialized 1337 if (old_value == -1 || !__kmp_atomic_compare_store( 1338 &th->th.th_suspend_init_count, old_value, -1)) { 1339 while (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) != new_value) { 1340 KMP_CPU_PAUSE(); 1341 } 1342 } else { 1343 // Claim to be the initializer and do initializations 1344 int status; 1345 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond, 1346 &__kmp_suspend_cond_attr); 1347 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1348 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex, 1349 &__kmp_suspend_mutex_attr); 1350 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1351 KMP_ATOMIC_ST_REL(&th->th.th_suspend_init_count, new_value); 1352 } 1353 } 1354 1355 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { 1356 if (KMP_ATOMIC_LD_ACQ(&th->th.th_suspend_init_count) > __kmp_fork_count) { 1357 /* this means we have initialize the suspension pthread objects for this 1358 thread in this instance of the process */ 1359 int status; 1360 1361 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond); 1362 if (status != 0 && status != EBUSY) { 1363 KMP_SYSFAIL("pthread_cond_destroy", status); 1364 } 1365 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex); 1366 if (status != 0 && status != EBUSY) { 1367 KMP_SYSFAIL("pthread_mutex_destroy", status); 1368 } 1369 --th->th.th_suspend_init_count; 1370 KMP_DEBUG_ASSERT(KMP_ATOMIC_LD_RLX(&th->th.th_suspend_init_count) == 1371 __kmp_fork_count); 1372 } 1373 } 1374 1375 // return true if lock obtained, false otherwise 1376 int __kmp_try_suspend_mx(kmp_info_t *th) { 1377 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0); 1378 } 1379 1380 void __kmp_lock_suspend_mx(kmp_info_t *th) { 1381 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1382 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1383 } 1384 1385 void __kmp_unlock_suspend_mx(kmp_info_t *th) { 1386 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1387 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1388 } 1389 1390 /* This routine puts the calling thread to sleep after setting the 1391 sleep bit for the indicated flag variable to true. */ 1392 template <class C> 1393 static inline void __kmp_suspend_template(int th_gtid, C *flag) { 1394 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend); 1395 kmp_info_t *th = __kmp_threads[th_gtid]; 1396 int status; 1397 typename C::flag_t old_spin; 1398 1399 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid, 1400 flag->get())); 1401 1402 __kmp_suspend_initialize_thread(th); 1403 1404 __kmp_lock_suspend_mx(th); 1405 1406 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n", 1407 th_gtid, flag->get())); 1408 1409 /* TODO: shouldn't this use release semantics to ensure that 1410 __kmp_suspend_initialize_thread gets called first? */ 1411 old_spin = flag->set_sleeping(); 1412 TCW_PTR(th->th.th_sleep_loc, (void *)flag); 1413 th->th.th_sleep_loc_type = flag->get_type(); 1414 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && 1415 __kmp_pause_status != kmp_soft_paused) { 1416 flag->unset_sleeping(); 1417 TCW_PTR(th->th.th_sleep_loc, NULL); 1418 th->th.th_sleep_loc_type = flag_unset; 1419 __kmp_unlock_suspend_mx(th); 1420 return; 1421 } 1422 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1423 " was %x\n", 1424 th_gtid, flag->get(), flag->load(), old_spin)); 1425 1426 if (flag->done_check_val(old_spin) || flag->done_check()) { 1427 flag->unset_sleeping(); 1428 TCW_PTR(th->th.th_sleep_loc, NULL); 1429 th->th.th_sleep_loc_type = flag_unset; 1430 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1431 "for spin(%p)\n", 1432 th_gtid, flag->get())); 1433 } else { 1434 /* Encapsulate in a loop as the documentation states that this may 1435 "with low probability" return when the condition variable has 1436 not been signaled or broadcast */ 1437 int deactivated = FALSE; 1438 1439 while (flag->is_sleeping()) { 1440 #ifdef DEBUG_SUSPEND 1441 char buffer[128]; 1442 __kmp_suspend_count++; 1443 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1444 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid, 1445 buffer); 1446 #endif 1447 // Mark the thread as no longer active (only in the first iteration of the 1448 // loop). 1449 if (!deactivated) { 1450 th->th.th_active = FALSE; 1451 if (th->th.th_active_in_pool) { 1452 th->th.th_active_in_pool = FALSE; 1453 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); 1454 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 1455 } 1456 deactivated = TRUE; 1457 } 1458 1459 KMP_DEBUG_ASSERT(th->th.th_sleep_loc); 1460 KMP_DEBUG_ASSERT(flag->get_type() == th->th.th_sleep_loc_type); 1461 1462 #if USE_SUSPEND_TIMEOUT 1463 struct timespec now; 1464 struct timeval tval; 1465 int msecs; 1466 1467 status = gettimeofday(&tval, NULL); 1468 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1469 TIMEVAL_TO_TIMESPEC(&tval, &now); 1470 1471 msecs = (4 * __kmp_dflt_blocktime) + 200; 1472 now.tv_sec += msecs / 1000; 1473 now.tv_nsec += (msecs % 1000) * 1000; 1474 1475 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " 1476 "pthread_cond_timedwait\n", 1477 th_gtid)); 1478 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond, 1479 &th->th.th_suspend_mx.m_mutex, &now); 1480 #else 1481 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform" 1482 " pthread_cond_wait\n", 1483 th_gtid)); 1484 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond, 1485 &th->th.th_suspend_mx.m_mutex); 1486 #endif // USE_SUSPEND_TIMEOUT 1487 1488 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) { 1489 KMP_SYSFAIL("pthread_cond_wait", status); 1490 } 1491 1492 KMP_DEBUG_ASSERT(flag->get_type() == flag->get_ptr_type()); 1493 1494 if (!flag->is_sleeping() && 1495 ((status == EINTR) || (status == ETIMEDOUT))) { 1496 // if interrupt or timeout, and thread is no longer sleeping, we need to 1497 // make sure sleep_loc gets reset; however, this shouldn't be needed if 1498 // we woke up with resume 1499 flag->unset_sleeping(); 1500 TCW_PTR(th->th.th_sleep_loc, NULL); 1501 th->th.th_sleep_loc_type = flag_unset; 1502 } 1503 #ifdef KMP_DEBUG 1504 if (status == ETIMEDOUT) { 1505 if (flag->is_sleeping()) { 1506 KF_TRACE(100, 1507 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid)); 1508 } else { 1509 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1510 "not set!\n", 1511 th_gtid)); 1512 TCW_PTR(th->th.th_sleep_loc, NULL); 1513 th->th.th_sleep_loc_type = flag_unset; 1514 } 1515 } else if (flag->is_sleeping()) { 1516 KF_TRACE(100, 1517 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid)); 1518 } 1519 #endif 1520 } // while 1521 1522 // Mark the thread as active again (if it was previous marked as inactive) 1523 if (deactivated) { 1524 th->th.th_active = TRUE; 1525 if (TCR_4(th->th.th_in_pool)) { 1526 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); 1527 th->th.th_active_in_pool = TRUE; 1528 } 1529 } 1530 } 1531 // We may have had the loop variable set before entering the loop body; 1532 // so we need to reset sleep_loc. 1533 TCW_PTR(th->th.th_sleep_loc, NULL); 1534 th->th.th_sleep_loc_type = flag_unset; 1535 1536 KMP_DEBUG_ASSERT(!flag->is_sleeping()); 1537 KMP_DEBUG_ASSERT(!th->th.th_sleep_loc); 1538 #ifdef DEBUG_SUSPEND 1539 { 1540 char buffer[128]; 1541 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1542 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid, 1543 buffer); 1544 } 1545 #endif 1546 1547 __kmp_unlock_suspend_mx(th); 1548 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid)); 1549 } 1550 1551 template <bool C, bool S> 1552 void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag) { 1553 __kmp_suspend_template(th_gtid, flag); 1554 } 1555 template <bool C, bool S> 1556 void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag) { 1557 __kmp_suspend_template(th_gtid, flag); 1558 } 1559 template <bool C, bool S> 1560 void __kmp_atomic_suspend_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag) { 1561 __kmp_suspend_template(th_gtid, flag); 1562 } 1563 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { 1564 __kmp_suspend_template(th_gtid, flag); 1565 } 1566 1567 template void __kmp_suspend_32<false, false>(int, kmp_flag_32<false, false> *); 1568 template void __kmp_suspend_64<false, true>(int, kmp_flag_64<false, true> *); 1569 template void __kmp_suspend_64<true, false>(int, kmp_flag_64<true, false> *); 1570 template void 1571 __kmp_atomic_suspend_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1572 template void 1573 __kmp_atomic_suspend_64<true, false>(int, kmp_atomic_flag_64<true, false> *); 1574 1575 /* This routine signals the thread specified by target_gtid to wake up 1576 after setting the sleep bit indicated by the flag argument to FALSE. 1577 The target thread must already have called __kmp_suspend_template() */ 1578 template <class C> 1579 static inline void __kmp_resume_template(int target_gtid, C *flag) { 1580 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1581 kmp_info_t *th = __kmp_threads[target_gtid]; 1582 int status; 1583 1584 #ifdef KMP_DEBUG 1585 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1586 #endif 1587 1588 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n", 1589 gtid, target_gtid)); 1590 KMP_DEBUG_ASSERT(gtid != target_gtid); 1591 1592 __kmp_suspend_initialize_thread(th); 1593 1594 __kmp_lock_suspend_mx(th); 1595 1596 if (!flag || flag != th->th.th_sleep_loc) { 1597 // coming from __kmp_null_resume_wrapper, or thread is now sleeping on a 1598 // different location; wake up at new location 1599 flag = (C *)CCAST(void *, th->th.th_sleep_loc); 1600 } 1601 1602 // First, check if the flag is null or its type has changed. If so, someone 1603 // else woke it up. 1604 if (!flag) { // Thread doesn't appear to be sleeping on anything 1605 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1606 "awake: flag(%p)\n", 1607 gtid, target_gtid, (void *)NULL)); 1608 __kmp_unlock_suspend_mx(th); 1609 return; 1610 } else if (flag->get_type() != th->th.th_sleep_loc_type) { 1611 // Flag type does not appear to match this function template; possibly the 1612 // thread is sleeping on something else. Try null resume again. 1613 KF_TRACE( 1614 5, 1615 ("__kmp_resume_template: T#%d retrying, thread T#%d Mismatch flag(%p), " 1616 "spin(%p) type=%d ptr_type=%d\n", 1617 gtid, target_gtid, flag, flag->get(), flag->get_type(), 1618 th->th.th_sleep_loc_type)); 1619 __kmp_unlock_suspend_mx(th); 1620 __kmp_null_resume_wrapper(th); 1621 return; 1622 } else { // if multiple threads are sleeping, flag should be internally 1623 // referring to a specific thread here 1624 if (!flag->is_sleeping()) { 1625 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1626 "awake: flag(%p): %u\n", 1627 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1628 __kmp_unlock_suspend_mx(th); 1629 return; 1630 } 1631 } 1632 KMP_DEBUG_ASSERT(flag); 1633 flag->unset_sleeping(); 1634 TCW_PTR(th->th.th_sleep_loc, NULL); 1635 th->th.th_sleep_loc_type = flag_unset; 1636 1637 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1638 "sleep bit for flag's loc(%p): %u\n", 1639 gtid, target_gtid, flag->get(), (unsigned int)flag->load())); 1640 1641 #ifdef DEBUG_SUSPEND 1642 { 1643 char buffer[128]; 1644 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1645 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid, 1646 target_gtid, buffer); 1647 } 1648 #endif 1649 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond); 1650 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1651 __kmp_unlock_suspend_mx(th); 1652 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" 1653 " for T#%d\n", 1654 gtid, target_gtid)); 1655 } 1656 1657 template <bool C, bool S> 1658 void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag) { 1659 __kmp_resume_template(target_gtid, flag); 1660 } 1661 template <bool C, bool S> 1662 void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag) { 1663 __kmp_resume_template(target_gtid, flag); 1664 } 1665 template <bool C, bool S> 1666 void __kmp_atomic_resume_64(int target_gtid, kmp_atomic_flag_64<C, S> *flag) { 1667 __kmp_resume_template(target_gtid, flag); 1668 } 1669 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { 1670 __kmp_resume_template(target_gtid, flag); 1671 } 1672 1673 template void __kmp_resume_32<false, true>(int, kmp_flag_32<false, true> *); 1674 template void __kmp_resume_32<false, false>(int, kmp_flag_32<false, false> *); 1675 template void __kmp_resume_64<false, true>(int, kmp_flag_64<false, true> *); 1676 template void 1677 __kmp_atomic_resume_64<false, true>(int, kmp_atomic_flag_64<false, true> *); 1678 1679 #if KMP_USE_MONITOR 1680 void __kmp_resume_monitor() { 1681 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1682 int status; 1683 #ifdef KMP_DEBUG 1684 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1685 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid, 1686 KMP_GTID_MONITOR)); 1687 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR); 1688 #endif 1689 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 1690 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1691 #ifdef DEBUG_SUSPEND 1692 { 1693 char buffer[128]; 1694 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond); 1695 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid, 1696 KMP_GTID_MONITOR, buffer); 1697 } 1698 #endif 1699 status = pthread_cond_signal(&__kmp_wait_cv.c_cond); 1700 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1701 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 1702 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1703 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up" 1704 " for T#%d\n", 1705 gtid, KMP_GTID_MONITOR)); 1706 } 1707 #endif // KMP_USE_MONITOR 1708 1709 void __kmp_yield() { sched_yield(); } 1710 1711 void __kmp_gtid_set_specific(int gtid) { 1712 if (__kmp_init_gtid) { 1713 int status; 1714 status = pthread_setspecific(__kmp_gtid_threadprivate_key, 1715 (void *)(intptr_t)(gtid + 1)); 1716 KMP_CHECK_SYSFAIL("pthread_setspecific", status); 1717 } else { 1718 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n")); 1719 } 1720 } 1721 1722 int __kmp_gtid_get_specific() { 1723 int gtid; 1724 if (!__kmp_init_gtid) { 1725 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " 1726 "KMP_GTID_SHUTDOWN\n")); 1727 return KMP_GTID_SHUTDOWN; 1728 } 1729 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key); 1730 if (gtid == 0) { 1731 gtid = KMP_GTID_DNE; 1732 } else { 1733 gtid--; 1734 } 1735 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n", 1736 __kmp_gtid_threadprivate_key, gtid)); 1737 return gtid; 1738 } 1739 1740 double __kmp_read_cpu_time(void) { 1741 /*clock_t t;*/ 1742 struct tms buffer; 1743 1744 /*t =*/times(&buffer); 1745 1746 return (double)(buffer.tms_utime + buffer.tms_cutime) / 1747 (double)CLOCKS_PER_SEC; 1748 } 1749 1750 int __kmp_read_system_info(struct kmp_sys_info *info) { 1751 int status; 1752 struct rusage r_usage; 1753 1754 memset(info, 0, sizeof(*info)); 1755 1756 status = getrusage(RUSAGE_SELF, &r_usage); 1757 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status); 1758 1759 // The maximum resident set size utilized (in kilobytes) 1760 info->maxrss = r_usage.ru_maxrss; 1761 // The number of page faults serviced without any I/O 1762 info->minflt = r_usage.ru_minflt; 1763 // The number of page faults serviced that required I/O 1764 info->majflt = r_usage.ru_majflt; 1765 // The number of times a process was "swapped" out of memory 1766 info->nswap = r_usage.ru_nswap; 1767 // The number of times the file system had to perform input 1768 info->inblock = r_usage.ru_inblock; 1769 // The number of times the file system had to perform output 1770 info->oublock = r_usage.ru_oublock; 1771 // The number of times a context switch was voluntarily 1772 info->nvcsw = r_usage.ru_nvcsw; 1773 // The number of times a context switch was forced 1774 info->nivcsw = r_usage.ru_nivcsw; 1775 1776 return (status != 0); 1777 } 1778 1779 void __kmp_read_system_time(double *delta) { 1780 double t_ns; 1781 struct timeval tval; 1782 struct timespec stop; 1783 int status; 1784 1785 status = gettimeofday(&tval, NULL); 1786 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1787 TIMEVAL_TO_TIMESPEC(&tval, &stop); 1788 t_ns = (double)(TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start)); 1789 *delta = (t_ns * 1e-9); 1790 } 1791 1792 void __kmp_clear_system_time(void) { 1793 struct timeval tval; 1794 int status; 1795 status = gettimeofday(&tval, NULL); 1796 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1797 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start); 1798 } 1799 1800 static int __kmp_get_xproc(void) { 1801 1802 int r = 0; 1803 1804 #if KMP_OS_LINUX 1805 1806 __kmp_type_convert(sysconf(_SC_NPROCESSORS_CONF), &(r)); 1807 1808 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_OPENBSD || \ 1809 KMP_OS_HURD 1810 1811 __kmp_type_convert(sysconf(_SC_NPROCESSORS_ONLN), &(r)); 1812 1813 #elif KMP_OS_DARWIN 1814 1815 // Bug C77011 High "OpenMP Threads and number of active cores". 1816 1817 // Find the number of available CPUs. 1818 kern_return_t rc; 1819 host_basic_info_data_t info; 1820 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT; 1821 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num); 1822 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) { 1823 // Cannot use KA_TRACE() here because this code works before trace support 1824 // is initialized. 1825 r = info.avail_cpus; 1826 } else { 1827 KMP_WARNING(CantGetNumAvailCPU); 1828 KMP_INFORM(AssumedNumCPU); 1829 } 1830 1831 #else 1832 1833 #error "Unknown or unsupported OS." 1834 1835 #endif 1836 1837 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */ 1838 1839 } // __kmp_get_xproc 1840 1841 int __kmp_read_from_file(char const *path, char const *format, ...) { 1842 int result; 1843 va_list args; 1844 1845 va_start(args, format); 1846 FILE *f = fopen(path, "rb"); 1847 if (f == NULL) 1848 return 0; 1849 result = vfscanf(f, format, args); 1850 fclose(f); 1851 1852 return result; 1853 } 1854 1855 void __kmp_runtime_initialize(void) { 1856 int status; 1857 pthread_mutexattr_t mutex_attr; 1858 pthread_condattr_t cond_attr; 1859 1860 if (__kmp_init_runtime) { 1861 return; 1862 } 1863 1864 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1865 if (!__kmp_cpuinfo.initialized) { 1866 __kmp_query_cpuid(&__kmp_cpuinfo); 1867 } 1868 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1869 1870 __kmp_xproc = __kmp_get_xproc(); 1871 1872 #if !KMP_32_BIT_ARCH 1873 struct rlimit rlim; 1874 // read stack size of calling thread, save it as default for worker threads; 1875 // this should be done before reading environment variables 1876 status = getrlimit(RLIMIT_STACK, &rlim); 1877 if (status == 0) { // success? 1878 __kmp_stksize = rlim.rlim_cur; 1879 __kmp_check_stksize(&__kmp_stksize); // check value and adjust if needed 1880 } 1881 #endif /* KMP_32_BIT_ARCH */ 1882 1883 if (sysconf(_SC_THREADS)) { 1884 1885 /* Query the maximum number of threads */ 1886 __kmp_type_convert(sysconf(_SC_THREAD_THREADS_MAX), &(__kmp_sys_max_nth)); 1887 if (__kmp_sys_max_nth == -1) { 1888 /* Unlimited threads for NPTL */ 1889 __kmp_sys_max_nth = INT_MAX; 1890 } else if (__kmp_sys_max_nth <= 1) { 1891 /* Can't tell, just use PTHREAD_THREADS_MAX */ 1892 __kmp_sys_max_nth = KMP_MAX_NTH; 1893 } 1894 1895 /* Query the minimum stack size */ 1896 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN); 1897 if (__kmp_sys_min_stksize <= 1) { 1898 __kmp_sys_min_stksize = KMP_MIN_STKSIZE; 1899 } 1900 } 1901 1902 /* Set up minimum number of threads to switch to TLS gtid */ 1903 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; 1904 1905 status = pthread_key_create(&__kmp_gtid_threadprivate_key, 1906 __kmp_internal_end_dest); 1907 KMP_CHECK_SYSFAIL("pthread_key_create", status); 1908 status = pthread_mutexattr_init(&mutex_attr); 1909 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1910 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr); 1911 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1912 status = pthread_mutexattr_destroy(&mutex_attr); 1913 KMP_CHECK_SYSFAIL("pthread_mutexattr_destroy", status); 1914 status = pthread_condattr_init(&cond_attr); 1915 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1916 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr); 1917 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1918 status = pthread_condattr_destroy(&cond_attr); 1919 KMP_CHECK_SYSFAIL("pthread_condattr_destroy", status); 1920 #if USE_ITT_BUILD 1921 __kmp_itt_initialize(); 1922 #endif /* USE_ITT_BUILD */ 1923 1924 __kmp_init_runtime = TRUE; 1925 } 1926 1927 void __kmp_runtime_destroy(void) { 1928 int status; 1929 1930 if (!__kmp_init_runtime) { 1931 return; // Nothing to do. 1932 } 1933 1934 #if USE_ITT_BUILD 1935 __kmp_itt_destroy(); 1936 #endif /* USE_ITT_BUILD */ 1937 1938 status = pthread_key_delete(__kmp_gtid_threadprivate_key); 1939 KMP_CHECK_SYSFAIL("pthread_key_delete", status); 1940 1941 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex); 1942 if (status != 0 && status != EBUSY) { 1943 KMP_SYSFAIL("pthread_mutex_destroy", status); 1944 } 1945 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond); 1946 if (status != 0 && status != EBUSY) { 1947 KMP_SYSFAIL("pthread_cond_destroy", status); 1948 } 1949 #if KMP_AFFINITY_SUPPORTED 1950 __kmp_affinity_uninitialize(); 1951 #endif 1952 1953 __kmp_init_runtime = FALSE; 1954 } 1955 1956 /* Put the thread to sleep for a time period */ 1957 /* NOTE: not currently used anywhere */ 1958 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); } 1959 1960 /* Calculate the elapsed wall clock time for the user */ 1961 void __kmp_elapsed(double *t) { 1962 int status; 1963 #ifdef FIX_SGI_CLOCK 1964 struct timespec ts; 1965 1966 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); 1967 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status); 1968 *t = 1969 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec; 1970 #else 1971 struct timeval tv; 1972 1973 status = gettimeofday(&tv, NULL); 1974 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1975 *t = 1976 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec; 1977 #endif 1978 } 1979 1980 /* Calculate the elapsed wall clock tick for the user */ 1981 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; } 1982 1983 /* Return the current time stamp in nsec */ 1984 kmp_uint64 __kmp_now_nsec() { 1985 struct timeval t; 1986 gettimeofday(&t, NULL); 1987 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec + 1988 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec; 1989 return nsec; 1990 } 1991 1992 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1993 /* Measure clock ticks per millisecond */ 1994 void __kmp_initialize_system_tick() { 1995 kmp_uint64 now, nsec2, diff; 1996 kmp_uint64 delay = 100000; // 50~100 usec on most machines. 1997 kmp_uint64 nsec = __kmp_now_nsec(); 1998 kmp_uint64 goal = __kmp_hardware_timestamp() + delay; 1999 while ((now = __kmp_hardware_timestamp()) < goal) 2000 ; 2001 nsec2 = __kmp_now_nsec(); 2002 diff = nsec2 - nsec; 2003 if (diff > 0) { 2004 kmp_uint64 tpms = ((kmp_uint64)1e6 * (delay + (now - goal)) / diff); 2005 if (tpms > 0) 2006 __kmp_ticks_per_msec = tpms; 2007 } 2008 } 2009 #endif 2010 2011 /* Determine whether the given address is mapped into the current address 2012 space. */ 2013 2014 int __kmp_is_address_mapped(void *addr) { 2015 2016 int found = 0; 2017 int rc; 2018 2019 #if KMP_OS_LINUX || KMP_OS_HURD 2020 2021 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the 2022 address ranges mapped into the address space. */ 2023 2024 char *name = __kmp_str_format("/proc/%d/maps", getpid()); 2025 FILE *file = NULL; 2026 2027 file = fopen(name, "r"); 2028 KMP_ASSERT(file != NULL); 2029 2030 for (;;) { 2031 2032 void *beginning = NULL; 2033 void *ending = NULL; 2034 char perms[5]; 2035 2036 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms); 2037 if (rc == EOF) { 2038 break; 2039 } 2040 KMP_ASSERT(rc == 3 && 2041 KMP_STRLEN(perms) == 4); // Make sure all fields are read. 2042 2043 // Ending address is not included in the region, but beginning is. 2044 if ((addr >= beginning) && (addr < ending)) { 2045 perms[2] = 0; // 3th and 4th character does not matter. 2046 if (strcmp(perms, "rw") == 0) { 2047 // Memory we are looking for should be readable and writable. 2048 found = 1; 2049 } 2050 break; 2051 } 2052 } 2053 2054 // Free resources. 2055 fclose(file); 2056 KMP_INTERNAL_FREE(name); 2057 #elif KMP_OS_FREEBSD 2058 char *buf; 2059 size_t lstsz; 2060 int mib[] = {CTL_KERN, KERN_PROC, KERN_PROC_VMMAP, getpid()}; 2061 rc = sysctl(mib, 4, NULL, &lstsz, NULL, 0); 2062 if (rc < 0) 2063 return 0; 2064 // We pass from number of vm entry's semantic 2065 // to size of whole entry map list. 2066 lstsz = lstsz * 4 / 3; 2067 buf = reinterpret_cast<char *>(kmpc_malloc(lstsz)); 2068 rc = sysctl(mib, 4, buf, &lstsz, NULL, 0); 2069 if (rc < 0) { 2070 kmpc_free(buf); 2071 return 0; 2072 } 2073 2074 char *lw = buf; 2075 char *up = buf + lstsz; 2076 2077 while (lw < up) { 2078 struct kinfo_vmentry *cur = reinterpret_cast<struct kinfo_vmentry *>(lw); 2079 size_t cursz = cur->kve_structsize; 2080 if (cursz == 0) 2081 break; 2082 void *start = reinterpret_cast<void *>(cur->kve_start); 2083 void *end = reinterpret_cast<void *>(cur->kve_end); 2084 // Readable/Writable addresses within current map entry 2085 if ((addr >= start) && (addr < end)) { 2086 if ((cur->kve_protection & KVME_PROT_READ) != 0 && 2087 (cur->kve_protection & KVME_PROT_WRITE) != 0) { 2088 found = 1; 2089 break; 2090 } 2091 } 2092 lw += cursz; 2093 } 2094 kmpc_free(buf); 2095 2096 #elif KMP_OS_DARWIN 2097 2098 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory 2099 using vm interface. */ 2100 2101 int buffer; 2102 vm_size_t count; 2103 rc = vm_read_overwrite( 2104 mach_task_self(), // Task to read memory of. 2105 (vm_address_t)(addr), // Address to read from. 2106 1, // Number of bytes to be read. 2107 (vm_address_t)(&buffer), // Address of buffer to save read bytes in. 2108 &count // Address of var to save number of read bytes in. 2109 ); 2110 if (rc == 0) { 2111 // Memory successfully read. 2112 found = 1; 2113 } 2114 2115 #elif KMP_OS_NETBSD 2116 2117 int mib[5]; 2118 mib[0] = CTL_VM; 2119 mib[1] = VM_PROC; 2120 mib[2] = VM_PROC_MAP; 2121 mib[3] = getpid(); 2122 mib[4] = sizeof(struct kinfo_vmentry); 2123 2124 size_t size; 2125 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0); 2126 KMP_ASSERT(!rc); 2127 KMP_ASSERT(size); 2128 2129 size = size * 4 / 3; 2130 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size); 2131 KMP_ASSERT(kiv); 2132 2133 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0); 2134 KMP_ASSERT(!rc); 2135 KMP_ASSERT(size); 2136 2137 for (size_t i = 0; i < size; i++) { 2138 if (kiv[i].kve_start >= (uint64_t)addr && 2139 kiv[i].kve_end <= (uint64_t)addr) { 2140 found = 1; 2141 break; 2142 } 2143 } 2144 KMP_INTERNAL_FREE(kiv); 2145 #elif KMP_OS_OPENBSD 2146 2147 int mib[3]; 2148 mib[0] = CTL_KERN; 2149 mib[1] = KERN_PROC_VMMAP; 2150 mib[2] = getpid(); 2151 2152 size_t size; 2153 uint64_t end; 2154 rc = sysctl(mib, 3, NULL, &size, NULL, 0); 2155 KMP_ASSERT(!rc); 2156 KMP_ASSERT(size); 2157 end = size; 2158 2159 struct kinfo_vmentry kiv = {.kve_start = 0}; 2160 2161 while ((rc = sysctl(mib, 3, &kiv, &size, NULL, 0)) == 0) { 2162 KMP_ASSERT(size); 2163 if (kiv.kve_end == end) 2164 break; 2165 2166 if (kiv.kve_start >= (uint64_t)addr && kiv.kve_end <= (uint64_t)addr) { 2167 found = 1; 2168 break; 2169 } 2170 kiv.kve_start += 1; 2171 } 2172 #elif KMP_OS_DRAGONFLY 2173 2174 // FIXME(DragonFly): Implement this 2175 found = 1; 2176 2177 #else 2178 2179 #error "Unknown or unsupported OS" 2180 2181 #endif 2182 2183 return found; 2184 2185 } // __kmp_is_address_mapped 2186 2187 #ifdef USE_LOAD_BALANCE 2188 2189 #if KMP_OS_DARWIN || KMP_OS_NETBSD 2190 2191 // The function returns the rounded value of the system load average 2192 // during given time interval which depends on the value of 2193 // __kmp_load_balance_interval variable (default is 60 sec, other values 2194 // may be 300 sec or 900 sec). 2195 // It returns -1 in case of error. 2196 int __kmp_get_load_balance(int max) { 2197 double averages[3]; 2198 int ret_avg = 0; 2199 2200 int res = getloadavg(averages, 3); 2201 2202 // Check __kmp_load_balance_interval to determine which of averages to use. 2203 // getloadavg() may return the number of samples less than requested that is 2204 // less than 3. 2205 if (__kmp_load_balance_interval < 180 && (res >= 1)) { 2206 ret_avg = (int)averages[0]; // 1 min 2207 } else if ((__kmp_load_balance_interval >= 180 && 2208 __kmp_load_balance_interval < 600) && 2209 (res >= 2)) { 2210 ret_avg = (int)averages[1]; // 5 min 2211 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) { 2212 ret_avg = (int)averages[2]; // 15 min 2213 } else { // Error occurred 2214 return -1; 2215 } 2216 2217 return ret_avg; 2218 } 2219 2220 #else // Linux* OS 2221 2222 // The function returns number of running (not sleeping) threads, or -1 in case 2223 // of error. Error could be reported if Linux* OS kernel too old (without 2224 // "/proc" support). Counting running threads stops if max running threads 2225 // encountered. 2226 int __kmp_get_load_balance(int max) { 2227 static int permanent_error = 0; 2228 static int glb_running_threads = 0; // Saved count of the running threads for 2229 // the thread balance algorithm 2230 static double glb_call_time = 0; /* Thread balance algorithm call time */ 2231 2232 int running_threads = 0; // Number of running threads in the system. 2233 2234 DIR *proc_dir = NULL; // Handle of "/proc/" directory. 2235 struct dirent *proc_entry = NULL; 2236 2237 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path. 2238 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory. 2239 struct dirent *task_entry = NULL; 2240 int task_path_fixed_len; 2241 2242 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path. 2243 int stat_file = -1; 2244 int stat_path_fixed_len; 2245 2246 int total_processes = 0; // Total number of processes in system. 2247 int total_threads = 0; // Total number of threads in system. 2248 2249 double call_time = 0.0; 2250 2251 __kmp_str_buf_init(&task_path); 2252 __kmp_str_buf_init(&stat_path); 2253 2254 __kmp_elapsed(&call_time); 2255 2256 if (glb_call_time && 2257 (call_time - glb_call_time < __kmp_load_balance_interval)) { 2258 running_threads = glb_running_threads; 2259 goto finish; 2260 } 2261 2262 glb_call_time = call_time; 2263 2264 // Do not spend time on scanning "/proc/" if we have a permanent error. 2265 if (permanent_error) { 2266 running_threads = -1; 2267 goto finish; 2268 } 2269 2270 if (max <= 0) { 2271 max = INT_MAX; 2272 } 2273 2274 // Open "/proc/" directory. 2275 proc_dir = opendir("/proc"); 2276 if (proc_dir == NULL) { 2277 // Cannot open "/prroc/". Probably the kernel does not support it. Return an 2278 // error now and in subsequent calls. 2279 running_threads = -1; 2280 permanent_error = 1; 2281 goto finish; 2282 } 2283 2284 // Initialize fixed part of task_path. This part will not change. 2285 __kmp_str_buf_cat(&task_path, "/proc/", 6); 2286 task_path_fixed_len = task_path.used; // Remember number of used characters. 2287 2288 proc_entry = readdir(proc_dir); 2289 while (proc_entry != NULL) { 2290 // Proc entry is a directory and name starts with a digit. Assume it is a 2291 // process' directory. 2292 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) { 2293 2294 ++total_processes; 2295 // Make sure init process is the very first in "/proc", so we can replace 2296 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes == 2297 // 1. We are going to check that total_processes == 1 => d_name == "1" is 2298 // true (where "=>" is implication). Since C++ does not have => operator, 2299 // let us replace it with its equivalent: a => b == ! a || b. 2300 KMP_DEBUG_ASSERT(total_processes != 1 || 2301 strcmp(proc_entry->d_name, "1") == 0); 2302 2303 // Construct task_path. 2304 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/". 2305 __kmp_str_buf_cat(&task_path, proc_entry->d_name, 2306 KMP_STRLEN(proc_entry->d_name)); 2307 __kmp_str_buf_cat(&task_path, "/task", 5); 2308 2309 task_dir = opendir(task_path.str); 2310 if (task_dir == NULL) { 2311 // Process can finish between reading "/proc/" directory entry and 2312 // opening process' "task/" directory. So, in general case we should not 2313 // complain, but have to skip this process and read the next one. But on 2314 // systems with no "task/" support we will spend lot of time to scan 2315 // "/proc/" tree again and again without any benefit. "init" process 2316 // (its pid is 1) should exist always, so, if we cannot open 2317 // "/proc/1/task/" directory, it means "task/" is not supported by 2318 // kernel. Report an error now and in the future. 2319 if (strcmp(proc_entry->d_name, "1") == 0) { 2320 running_threads = -1; 2321 permanent_error = 1; 2322 goto finish; 2323 } 2324 } else { 2325 // Construct fixed part of stat file path. 2326 __kmp_str_buf_clear(&stat_path); 2327 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used); 2328 __kmp_str_buf_cat(&stat_path, "/", 1); 2329 stat_path_fixed_len = stat_path.used; 2330 2331 task_entry = readdir(task_dir); 2332 while (task_entry != NULL) { 2333 // It is a directory and name starts with a digit. 2334 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) { 2335 ++total_threads; 2336 2337 // Construct complete stat file path. Easiest way would be: 2338 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str, 2339 // task_entry->d_name ); 2340 // but seriae of __kmp_str_buf_cat works a bit faster. 2341 stat_path.used = 2342 stat_path_fixed_len; // Reset stat path to its fixed part. 2343 __kmp_str_buf_cat(&stat_path, task_entry->d_name, 2344 KMP_STRLEN(task_entry->d_name)); 2345 __kmp_str_buf_cat(&stat_path, "/stat", 5); 2346 2347 // Note: Low-level API (open/read/close) is used. High-level API 2348 // (fopen/fclose) works ~ 30 % slower. 2349 stat_file = open(stat_path.str, O_RDONLY); 2350 if (stat_file == -1) { 2351 // We cannot report an error because task (thread) can terminate 2352 // just before reading this file. 2353 } else { 2354 /* Content of "stat" file looks like: 2355 24285 (program) S ... 2356 2357 It is a single line (if program name does not include funny 2358 symbols). First number is a thread id, then name of executable 2359 file name in paretheses, then state of the thread. We need just 2360 thread state. 2361 2362 Good news: Length of program name is 15 characters max. Longer 2363 names are truncated. 2364 2365 Thus, we need rather short buffer: 15 chars for program name + 2366 2 parenthesis, + 3 spaces + ~7 digits of pid = 37. 2367 2368 Bad news: Program name may contain special symbols like space, 2369 closing parenthesis, or even new line. This makes parsing 2370 "stat" file not 100 % reliable. In case of fanny program names 2371 parsing may fail (report incorrect thread state). 2372 2373 Parsing "status" file looks more promissing (due to different 2374 file structure and escaping special symbols) but reading and 2375 parsing of "status" file works slower. 2376 -- ln 2377 */ 2378 char buffer[65]; 2379 ssize_t len; 2380 len = read(stat_file, buffer, sizeof(buffer) - 1); 2381 if (len >= 0) { 2382 buffer[len] = 0; 2383 // Using scanf: 2384 // sscanf( buffer, "%*d (%*s) %c ", & state ); 2385 // looks very nice, but searching for a closing parenthesis 2386 // works a bit faster. 2387 char *close_parent = strstr(buffer, ") "); 2388 if (close_parent != NULL) { 2389 char state = *(close_parent + 2); 2390 if (state == 'R') { 2391 ++running_threads; 2392 if (running_threads >= max) { 2393 goto finish; 2394 } 2395 } 2396 } 2397 } 2398 close(stat_file); 2399 stat_file = -1; 2400 } 2401 } 2402 task_entry = readdir(task_dir); 2403 } 2404 closedir(task_dir); 2405 task_dir = NULL; 2406 } 2407 } 2408 proc_entry = readdir(proc_dir); 2409 } 2410 2411 // There _might_ be a timing hole where the thread executing this 2412 // code get skipped in the load balance, and running_threads is 0. 2413 // Assert in the debug builds only!!! 2414 KMP_DEBUG_ASSERT(running_threads > 0); 2415 if (running_threads <= 0) { 2416 running_threads = 1; 2417 } 2418 2419 finish: // Clean up and exit. 2420 if (proc_dir != NULL) { 2421 closedir(proc_dir); 2422 } 2423 __kmp_str_buf_free(&task_path); 2424 if (task_dir != NULL) { 2425 closedir(task_dir); 2426 } 2427 __kmp_str_buf_free(&stat_path); 2428 if (stat_file != -1) { 2429 close(stat_file); 2430 } 2431 2432 glb_running_threads = running_threads; 2433 2434 return running_threads; 2435 2436 } // __kmp_get_load_balance 2437 2438 #endif // KMP_OS_DARWIN 2439 2440 #endif // USE_LOAD_BALANCE 2441 2442 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2443 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || \ 2444 KMP_ARCH_PPC64 || KMP_ARCH_RISCV64) 2445 2446 // we really only need the case with 1 argument, because CLANG always build 2447 // a struct of pointers to shared variables referenced in the outlined function 2448 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc, 2449 void *p_argv[] 2450 #if OMPT_SUPPORT 2451 , 2452 void **exit_frame_ptr 2453 #endif 2454 ) { 2455 #if OMPT_SUPPORT 2456 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0); 2457 #endif 2458 2459 switch (argc) { 2460 default: 2461 fprintf(stderr, "Too many args to microtask: %d!\n", argc); 2462 fflush(stderr); 2463 exit(-1); 2464 case 0: 2465 (*pkfn)(>id, &tid); 2466 break; 2467 case 1: 2468 (*pkfn)(>id, &tid, p_argv[0]); 2469 break; 2470 case 2: 2471 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]); 2472 break; 2473 case 3: 2474 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]); 2475 break; 2476 case 4: 2477 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]); 2478 break; 2479 case 5: 2480 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]); 2481 break; 2482 case 6: 2483 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2484 p_argv[5]); 2485 break; 2486 case 7: 2487 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2488 p_argv[5], p_argv[6]); 2489 break; 2490 case 8: 2491 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2492 p_argv[5], p_argv[6], p_argv[7]); 2493 break; 2494 case 9: 2495 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2496 p_argv[5], p_argv[6], p_argv[7], p_argv[8]); 2497 break; 2498 case 10: 2499 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2500 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]); 2501 break; 2502 case 11: 2503 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2504 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]); 2505 break; 2506 case 12: 2507 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2508 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2509 p_argv[11]); 2510 break; 2511 case 13: 2512 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2513 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2514 p_argv[11], p_argv[12]); 2515 break; 2516 case 14: 2517 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2518 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2519 p_argv[11], p_argv[12], p_argv[13]); 2520 break; 2521 case 15: 2522 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2523 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2524 p_argv[11], p_argv[12], p_argv[13], p_argv[14]); 2525 break; 2526 } 2527 2528 return 1; 2529 } 2530 2531 #endif 2532 2533 #if KMP_OS_LINUX 2534 // Functions for hidden helper task 2535 namespace { 2536 // Condition variable for initializing hidden helper team 2537 pthread_cond_t hidden_helper_threads_initz_cond_var; 2538 pthread_mutex_t hidden_helper_threads_initz_lock; 2539 volatile int hidden_helper_initz_signaled = FALSE; 2540 2541 // Condition variable for deinitializing hidden helper team 2542 pthread_cond_t hidden_helper_threads_deinitz_cond_var; 2543 pthread_mutex_t hidden_helper_threads_deinitz_lock; 2544 volatile int hidden_helper_deinitz_signaled = FALSE; 2545 2546 // Condition variable for the wrapper function of main thread 2547 pthread_cond_t hidden_helper_main_thread_cond_var; 2548 pthread_mutex_t hidden_helper_main_thread_lock; 2549 volatile int hidden_helper_main_thread_signaled = FALSE; 2550 2551 // Semaphore for worker threads. We don't use condition variable here in case 2552 // that when multiple signals are sent at the same time, only one thread might 2553 // be waken. 2554 sem_t hidden_helper_task_sem; 2555 } // namespace 2556 2557 void __kmp_hidden_helper_worker_thread_wait() { 2558 int status = sem_wait(&hidden_helper_task_sem); 2559 KMP_CHECK_SYSFAIL("sem_wait", status); 2560 } 2561 2562 void __kmp_do_initialize_hidden_helper_threads() { 2563 // Initialize condition variable 2564 int status = 2565 pthread_cond_init(&hidden_helper_threads_initz_cond_var, nullptr); 2566 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2567 2568 status = pthread_cond_init(&hidden_helper_threads_deinitz_cond_var, nullptr); 2569 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2570 2571 status = pthread_cond_init(&hidden_helper_main_thread_cond_var, nullptr); 2572 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 2573 2574 status = pthread_mutex_init(&hidden_helper_threads_initz_lock, nullptr); 2575 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2576 2577 status = pthread_mutex_init(&hidden_helper_threads_deinitz_lock, nullptr); 2578 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2579 2580 status = pthread_mutex_init(&hidden_helper_main_thread_lock, nullptr); 2581 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 2582 2583 // Initialize the semaphore 2584 status = sem_init(&hidden_helper_task_sem, 0, 0); 2585 KMP_CHECK_SYSFAIL("sem_init", status); 2586 2587 // Create a new thread to finish initialization 2588 pthread_t handle; 2589 status = pthread_create( 2590 &handle, nullptr, 2591 [](void *) -> void * { 2592 __kmp_hidden_helper_threads_initz_routine(); 2593 return nullptr; 2594 }, 2595 nullptr); 2596 KMP_CHECK_SYSFAIL("pthread_create", status); 2597 } 2598 2599 void __kmp_hidden_helper_threads_initz_wait() { 2600 // Initial thread waits here for the completion of the initialization. The 2601 // condition variable will be notified by main thread of hidden helper teams. 2602 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2603 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2604 2605 if (!TCR_4(hidden_helper_initz_signaled)) { 2606 status = pthread_cond_wait(&hidden_helper_threads_initz_cond_var, 2607 &hidden_helper_threads_initz_lock); 2608 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2609 } 2610 2611 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2612 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2613 } 2614 2615 void __kmp_hidden_helper_initz_release() { 2616 // After all initialization, reset __kmp_init_hidden_helper_threads to false. 2617 int status = pthread_mutex_lock(&hidden_helper_threads_initz_lock); 2618 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2619 2620 status = pthread_cond_signal(&hidden_helper_threads_initz_cond_var); 2621 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2622 2623 TCW_SYNC_4(hidden_helper_initz_signaled, TRUE); 2624 2625 status = pthread_mutex_unlock(&hidden_helper_threads_initz_lock); 2626 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2627 } 2628 2629 void __kmp_hidden_helper_main_thread_wait() { 2630 // The main thread of hidden helper team will be blocked here. The 2631 // condition variable can only be signal in the destructor of RTL. 2632 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2633 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2634 2635 if (!TCR_4(hidden_helper_main_thread_signaled)) { 2636 status = pthread_cond_wait(&hidden_helper_main_thread_cond_var, 2637 &hidden_helper_main_thread_lock); 2638 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2639 } 2640 2641 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2642 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2643 } 2644 2645 void __kmp_hidden_helper_main_thread_release() { 2646 // The initial thread of OpenMP RTL should call this function to wake up the 2647 // main thread of hidden helper team. 2648 int status = pthread_mutex_lock(&hidden_helper_main_thread_lock); 2649 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2650 2651 status = pthread_cond_signal(&hidden_helper_main_thread_cond_var); 2652 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 2653 2654 // The hidden helper team is done here 2655 TCW_SYNC_4(hidden_helper_main_thread_signaled, TRUE); 2656 2657 status = pthread_mutex_unlock(&hidden_helper_main_thread_lock); 2658 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2659 } 2660 2661 void __kmp_hidden_helper_worker_thread_signal() { 2662 int status = sem_post(&hidden_helper_task_sem); 2663 KMP_CHECK_SYSFAIL("sem_post", status); 2664 } 2665 2666 void __kmp_hidden_helper_threads_deinitz_wait() { 2667 // Initial thread waits here for the completion of the deinitialization. The 2668 // condition variable will be notified by main thread of hidden helper teams. 2669 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2670 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2671 2672 if (!TCR_4(hidden_helper_deinitz_signaled)) { 2673 status = pthread_cond_wait(&hidden_helper_threads_deinitz_cond_var, 2674 &hidden_helper_threads_deinitz_lock); 2675 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2676 } 2677 2678 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2679 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2680 } 2681 2682 void __kmp_hidden_helper_threads_deinitz_release() { 2683 int status = pthread_mutex_lock(&hidden_helper_threads_deinitz_lock); 2684 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 2685 2686 status = pthread_cond_signal(&hidden_helper_threads_deinitz_cond_var); 2687 KMP_CHECK_SYSFAIL("pthread_cond_wait", status); 2688 2689 TCW_SYNC_4(hidden_helper_deinitz_signaled, TRUE); 2690 2691 status = pthread_mutex_unlock(&hidden_helper_threads_deinitz_lock); 2692 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 2693 } 2694 #else // KMP_OS_LINUX 2695 void __kmp_hidden_helper_worker_thread_wait() { 2696 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2697 } 2698 2699 void __kmp_do_initialize_hidden_helper_threads() { 2700 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2701 } 2702 2703 void __kmp_hidden_helper_threads_initz_wait() { 2704 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2705 } 2706 2707 void __kmp_hidden_helper_initz_release() { 2708 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2709 } 2710 2711 void __kmp_hidden_helper_main_thread_wait() { 2712 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2713 } 2714 2715 void __kmp_hidden_helper_main_thread_release() { 2716 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2717 } 2718 2719 void __kmp_hidden_helper_worker_thread_signal() { 2720 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2721 } 2722 2723 void __kmp_hidden_helper_threads_deinitz_wait() { 2724 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2725 } 2726 2727 void __kmp_hidden_helper_threads_deinitz_release() { 2728 KMP_ASSERT(0 && "Hidden helper task is not supported on this OS"); 2729 } 2730 #endif // KMP_OS_LINUX 2731 2732 // end of file // 2733