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 #include <sys/resource.h> 29 #include <sys/syscall.h> 30 #include <sys/time.h> 31 #include <sys/times.h> 32 #include <unistd.h> 33 34 #if KMP_OS_LINUX && !KMP_OS_CNK 35 #include <sys/sysinfo.h> 36 #if KMP_USE_FUTEX 37 // We should really include <futex.h>, but that causes compatibility problems on 38 // different Linux* OS distributions that either require that you include (or 39 // break when you try to include) <pci/types.h>. Since all we need is the two 40 // macros below (which are part of the kernel ABI, so can't change) we just 41 // define the constants here and don't include <futex.h> 42 #ifndef FUTEX_WAIT 43 #define FUTEX_WAIT 0 44 #endif 45 #ifndef FUTEX_WAKE 46 #define FUTEX_WAKE 1 47 #endif 48 #endif 49 #elif KMP_OS_DARWIN 50 #include <mach/mach.h> 51 #include <sys/sysctl.h> 52 #elif KMP_OS_DRAGONFLY || KMP_OS_FREEBSD 53 #include <pthread_np.h> 54 #elif KMP_OS_NETBSD 55 #include <sys/types.h> 56 #include <sys/sysctl.h> 57 #endif 58 59 #include <ctype.h> 60 #include <dirent.h> 61 #include <fcntl.h> 62 63 #include "tsan_annotations.h" 64 65 struct kmp_sys_timer { 66 struct timespec start; 67 }; 68 69 // Convert timespec to nanoseconds. 70 #define TS2NS(timespec) (((timespec).tv_sec * 1e9) + (timespec).tv_nsec) 71 72 static struct kmp_sys_timer __kmp_sys_timer_data; 73 74 #if KMP_HANDLE_SIGNALS 75 typedef void (*sig_func_t)(int); 76 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG]; 77 static sigset_t __kmp_sigset; 78 #endif 79 80 static int __kmp_init_runtime = FALSE; 81 82 static int __kmp_fork_count = 0; 83 84 static pthread_condattr_t __kmp_suspend_cond_attr; 85 static pthread_mutexattr_t __kmp_suspend_mutex_attr; 86 87 static kmp_cond_align_t __kmp_wait_cv; 88 static kmp_mutex_align_t __kmp_wait_mx; 89 90 kmp_uint64 __kmp_ticks_per_msec = 1000000; 91 92 #ifdef DEBUG_SUSPEND 93 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) { 94 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))", 95 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock, 96 cond->c_cond.__c_waiting); 97 } 98 #endif 99 100 #if (KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED) 101 102 /* Affinity support */ 103 104 void __kmp_affinity_bind_thread(int which) { 105 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(), 106 "Illegal set affinity operation when not capable"); 107 108 kmp_affin_mask_t *mask; 109 KMP_CPU_ALLOC_ON_STACK(mask); 110 KMP_CPU_ZERO(mask); 111 KMP_CPU_SET(which, mask); 112 __kmp_set_system_affinity(mask, TRUE); 113 KMP_CPU_FREE_FROM_STACK(mask); 114 } 115 116 /* Determine if we can access affinity functionality on this version of 117 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set 118 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */ 119 void __kmp_affinity_determine_capable(const char *env_var) { 120 // Check and see if the OS supports thread affinity. 121 122 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024) 123 124 int gCode; 125 int sCode; 126 unsigned char *buf; 127 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 128 129 // If Linux* OS: 130 // If the syscall fails or returns a suggestion for the size, 131 // then we don't have to search for an appropriate size. 132 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_SIZE_LIMIT, buf); 133 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 134 "initial getaffinity call returned %d errno = %d\n", 135 gCode, errno)); 136 137 // if ((gCode < 0) && (errno == ENOSYS)) 138 if (gCode < 0) { 139 // System call not supported 140 if (__kmp_affinity_verbose || 141 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) && 142 (__kmp_affinity_type != affinity_default) && 143 (__kmp_affinity_type != affinity_disabled))) { 144 int error = errno; 145 kmp_msg_t err_code = KMP_ERR(error); 146 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 147 err_code, __kmp_msg_null); 148 if (__kmp_generate_warnings == kmp_warnings_off) { 149 __kmp_str_free(&err_code.str); 150 } 151 } 152 KMP_AFFINITY_DISABLE(); 153 KMP_INTERNAL_FREE(buf); 154 return; 155 } 156 if (gCode > 0) { // Linux* OS only 157 // The optimal situation: the OS returns the size of the buffer it expects. 158 // 159 // A verification of correct behavior is that Isetaffinity on a NULL 160 // buffer with the same size fails with errno set to EFAULT. 161 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL); 162 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 163 "setaffinity for mask size %d returned %d errno = %d\n", 164 gCode, sCode, errno)); 165 if (sCode < 0) { 166 if (errno == ENOSYS) { 167 if (__kmp_affinity_verbose || 168 (__kmp_affinity_warnings && 169 (__kmp_affinity_type != affinity_none) && 170 (__kmp_affinity_type != affinity_default) && 171 (__kmp_affinity_type != affinity_disabled))) { 172 int error = errno; 173 kmp_msg_t err_code = KMP_ERR(error); 174 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var), 175 err_code, __kmp_msg_null); 176 if (__kmp_generate_warnings == kmp_warnings_off) { 177 __kmp_str_free(&err_code.str); 178 } 179 } 180 KMP_AFFINITY_DISABLE(); 181 KMP_INTERNAL_FREE(buf); 182 } 183 if (errno == EFAULT) { 184 KMP_AFFINITY_ENABLE(gCode); 185 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 186 "affinity supported (mask size %d)\n", 187 (int)__kmp_affin_mask_size)); 188 KMP_INTERNAL_FREE(buf); 189 return; 190 } 191 } 192 } 193 194 // Call the getaffinity system call repeatedly with increasing set sizes 195 // until we succeed, or reach an upper bound on the search. 196 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 197 "searching for proper set size\n")); 198 int size; 199 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) { 200 gCode = syscall(__NR_sched_getaffinity, 0, size, buf); 201 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 202 "getaffinity for mask size %d returned %d errno = %d\n", 203 size, gCode, errno)); 204 205 if (gCode < 0) { 206 if (errno == ENOSYS) { 207 // We shouldn't get here 208 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 209 "inconsistent OS call behavior: errno == ENOSYS for mask " 210 "size %d\n", 211 size)); 212 if (__kmp_affinity_verbose || 213 (__kmp_affinity_warnings && 214 (__kmp_affinity_type != affinity_none) && 215 (__kmp_affinity_type != affinity_default) && 216 (__kmp_affinity_type != affinity_disabled))) { 217 int error = errno; 218 kmp_msg_t err_code = KMP_ERR(error); 219 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 220 err_code, __kmp_msg_null); 221 if (__kmp_generate_warnings == kmp_warnings_off) { 222 __kmp_str_free(&err_code.str); 223 } 224 } 225 KMP_AFFINITY_DISABLE(); 226 KMP_INTERNAL_FREE(buf); 227 return; 228 } 229 continue; 230 } 231 232 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL); 233 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 234 "setaffinity for mask size %d returned %d errno = %d\n", 235 gCode, sCode, errno)); 236 if (sCode < 0) { 237 if (errno == ENOSYS) { // Linux* OS only 238 // We shouldn't get here 239 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 240 "inconsistent OS call behavior: errno == ENOSYS for mask " 241 "size %d\n", 242 size)); 243 if (__kmp_affinity_verbose || 244 (__kmp_affinity_warnings && 245 (__kmp_affinity_type != affinity_none) && 246 (__kmp_affinity_type != affinity_default) && 247 (__kmp_affinity_type != affinity_disabled))) { 248 int error = errno; 249 kmp_msg_t err_code = KMP_ERR(error); 250 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var), 251 err_code, __kmp_msg_null); 252 if (__kmp_generate_warnings == kmp_warnings_off) { 253 __kmp_str_free(&err_code.str); 254 } 255 } 256 KMP_AFFINITY_DISABLE(); 257 KMP_INTERNAL_FREE(buf); 258 return; 259 } 260 if (errno == EFAULT) { 261 KMP_AFFINITY_ENABLE(gCode); 262 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 263 "affinity supported (mask size %d)\n", 264 (int)__kmp_affin_mask_size)); 265 KMP_INTERNAL_FREE(buf); 266 return; 267 } 268 } 269 } 270 // save uncaught error code 271 // int error = errno; 272 KMP_INTERNAL_FREE(buf); 273 // restore uncaught error code, will be printed at the next KMP_WARNING below 274 // errno = error; 275 276 // Affinity is not supported 277 KMP_AFFINITY_DISABLE(); 278 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 279 "cannot determine mask size - affinity not supported\n")); 280 if (__kmp_affinity_verbose || 281 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) && 282 (__kmp_affinity_type != affinity_default) && 283 (__kmp_affinity_type != affinity_disabled))) { 284 KMP_WARNING(AffCantGetMaskSize, env_var); 285 } 286 } 287 288 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 289 290 #if KMP_USE_FUTEX 291 292 int __kmp_futex_determine_capable() { 293 int loc = 0; 294 int rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0); 295 int retval = (rc == 0) || (errno != ENOSYS); 296 297 KA_TRACE(10, 298 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno)); 299 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n", 300 retval ? "" : " not")); 301 302 return retval; 303 } 304 305 #endif // KMP_USE_FUTEX 306 307 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS) 308 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to 309 use compare_and_store for these routines */ 310 311 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) { 312 kmp_int8 old_value, new_value; 313 314 old_value = TCR_1(*p); 315 new_value = old_value | d; 316 317 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 318 KMP_CPU_PAUSE(); 319 old_value = TCR_1(*p); 320 new_value = old_value | d; 321 } 322 return old_value; 323 } 324 325 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) { 326 kmp_int8 old_value, new_value; 327 328 old_value = TCR_1(*p); 329 new_value = old_value & d; 330 331 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 332 KMP_CPU_PAUSE(); 333 old_value = TCR_1(*p); 334 new_value = old_value & d; 335 } 336 return old_value; 337 } 338 339 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) { 340 kmp_uint32 old_value, new_value; 341 342 old_value = TCR_4(*p); 343 new_value = old_value | d; 344 345 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 346 KMP_CPU_PAUSE(); 347 old_value = TCR_4(*p); 348 new_value = old_value | d; 349 } 350 return old_value; 351 } 352 353 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) { 354 kmp_uint32 old_value, new_value; 355 356 old_value = TCR_4(*p); 357 new_value = old_value & d; 358 359 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 360 KMP_CPU_PAUSE(); 361 old_value = TCR_4(*p); 362 new_value = old_value & d; 363 } 364 return old_value; 365 } 366 367 #if KMP_ARCH_X86 368 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) { 369 kmp_int8 old_value, new_value; 370 371 old_value = TCR_1(*p); 372 new_value = old_value + d; 373 374 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 375 KMP_CPU_PAUSE(); 376 old_value = TCR_1(*p); 377 new_value = old_value + d; 378 } 379 return old_value; 380 } 381 382 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) { 383 kmp_int64 old_value, new_value; 384 385 old_value = TCR_8(*p); 386 new_value = old_value + d; 387 388 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 389 KMP_CPU_PAUSE(); 390 old_value = TCR_8(*p); 391 new_value = old_value + d; 392 } 393 return old_value; 394 } 395 #endif /* KMP_ARCH_X86 */ 396 397 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) { 398 kmp_uint64 old_value, new_value; 399 400 old_value = TCR_8(*p); 401 new_value = old_value | d; 402 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 403 KMP_CPU_PAUSE(); 404 old_value = TCR_8(*p); 405 new_value = old_value | d; 406 } 407 return old_value; 408 } 409 410 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) { 411 kmp_uint64 old_value, new_value; 412 413 old_value = TCR_8(*p); 414 new_value = old_value & d; 415 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 416 KMP_CPU_PAUSE(); 417 old_value = TCR_8(*p); 418 new_value = old_value & d; 419 } 420 return old_value; 421 } 422 423 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */ 424 425 void __kmp_terminate_thread(int gtid) { 426 int status; 427 kmp_info_t *th = __kmp_threads[gtid]; 428 429 if (!th) 430 return; 431 432 #ifdef KMP_CANCEL_THREADS 433 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid)); 434 status = pthread_cancel(th->th.th_info.ds.ds_thread); 435 if (status != 0 && status != ESRCH) { 436 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status), 437 __kmp_msg_null); 438 } 439 #endif 440 __kmp_yield(TRUE); 441 } // 442 443 /* Set thread stack info according to values returned by pthread_getattr_np(). 444 If values are unreasonable, assume call failed and use incremental stack 445 refinement method instead. Returns TRUE if the stack parameters could be 446 determined exactly, FALSE if incremental refinement is necessary. */ 447 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) { 448 int stack_data; 449 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 450 KMP_OS_HURD 451 pthread_attr_t attr; 452 int status; 453 size_t size = 0; 454 void *addr = 0; 455 456 /* Always do incremental stack refinement for ubermaster threads since the 457 initial thread stack range can be reduced by sibling thread creation so 458 pthread_attr_getstack may cause thread gtid aliasing */ 459 if (!KMP_UBER_GTID(gtid)) { 460 461 /* Fetch the real thread attributes */ 462 status = pthread_attr_init(&attr); 463 KMP_CHECK_SYSFAIL("pthread_attr_init", status); 464 #if KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD 465 status = pthread_attr_get_np(pthread_self(), &attr); 466 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status); 467 #else 468 status = pthread_getattr_np(pthread_self(), &attr); 469 KMP_CHECK_SYSFAIL("pthread_getattr_np", status); 470 #endif 471 status = pthread_attr_getstack(&attr, &addr, &size); 472 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status); 473 KA_TRACE(60, 474 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:" 475 " %lu, low addr: %p\n", 476 gtid, size, addr)); 477 status = pthread_attr_destroy(&attr); 478 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status); 479 } 480 481 if (size != 0 && addr != 0) { // was stack parameter determination successful? 482 /* Store the correct base and size */ 483 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size)); 484 TCW_PTR(th->th.th_info.ds.ds_stacksize, size); 485 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE); 486 return TRUE; 487 } 488 #endif /* KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || 489 KMP_OS_HURD */ 490 /* Use incremental refinement starting from initial conservative estimate */ 491 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0); 492 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data); 493 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE); 494 return FALSE; 495 } 496 497 static void *__kmp_launch_worker(void *thr) { 498 int status, old_type, old_state; 499 #ifdef KMP_BLOCK_SIGNALS 500 sigset_t new_set, old_set; 501 #endif /* KMP_BLOCK_SIGNALS */ 502 void *exit_val; 503 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 504 KMP_OS_OPENBSD || KMP_OS_HURD 505 void *volatile padding = 0; 506 #endif 507 int gtid; 508 509 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid; 510 __kmp_gtid_set_specific(gtid); 511 #ifdef KMP_TDATA_GTID 512 __kmp_gtid = gtid; 513 #endif 514 #if KMP_STATS_ENABLED 515 // set thread local index to point to thread-specific stats 516 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats; 517 __kmp_stats_thread_ptr->startLife(); 518 KMP_SET_THREAD_STATE(IDLE); 519 KMP_INIT_PARTITIONED_TIMERS(OMP_idle); 520 #endif 521 522 #if USE_ITT_BUILD 523 __kmp_itt_thread_name(gtid); 524 #endif /* USE_ITT_BUILD */ 525 526 #if KMP_AFFINITY_SUPPORTED 527 __kmp_affinity_set_init_mask(gtid, FALSE); 528 #endif 529 530 #ifdef KMP_CANCEL_THREADS 531 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 532 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 533 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 534 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 535 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 536 #endif 537 538 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 539 // Set FP control regs to be a copy of the parallel initialization thread's. 540 __kmp_clear_x87_fpu_status_word(); 541 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); 542 __kmp_load_mxcsr(&__kmp_init_mxcsr); 543 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 544 545 #ifdef KMP_BLOCK_SIGNALS 546 status = sigfillset(&new_set); 547 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 548 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set); 549 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 550 #endif /* KMP_BLOCK_SIGNALS */ 551 552 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 553 KMP_OS_OPENBSD 554 if (__kmp_stkoffset > 0 && gtid > 0) { 555 padding = KMP_ALLOCA(gtid * __kmp_stkoffset); 556 } 557 #endif 558 559 KMP_MB(); 560 __kmp_set_stack_info(gtid, (kmp_info_t *)thr); 561 562 __kmp_check_stack_overlap((kmp_info_t *)thr); 563 564 exit_val = __kmp_launch_thread((kmp_info_t *)thr); 565 566 #ifdef KMP_BLOCK_SIGNALS 567 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL); 568 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 569 #endif /* KMP_BLOCK_SIGNALS */ 570 571 return exit_val; 572 } 573 574 #if KMP_USE_MONITOR 575 /* The monitor thread controls all of the threads in the complex */ 576 577 static void *__kmp_launch_monitor(void *thr) { 578 int status, old_type, old_state; 579 #ifdef KMP_BLOCK_SIGNALS 580 sigset_t new_set; 581 #endif /* KMP_BLOCK_SIGNALS */ 582 struct timespec interval; 583 int yield_count; 584 int yield_cycles = 0; 585 586 KMP_MB(); /* Flush all pending memory write invalidates. */ 587 588 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n")); 589 590 /* register us as the monitor thread */ 591 __kmp_gtid_set_specific(KMP_GTID_MONITOR); 592 #ifdef KMP_TDATA_GTID 593 __kmp_gtid = KMP_GTID_MONITOR; 594 #endif 595 596 KMP_MB(); 597 598 #if USE_ITT_BUILD 599 // Instruct Intel(R) Threading Tools to ignore monitor thread. 600 __kmp_itt_thread_ignore(); 601 #endif /* USE_ITT_BUILD */ 602 603 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid, 604 (kmp_info_t *)thr); 605 606 __kmp_check_stack_overlap((kmp_info_t *)thr); 607 608 #ifdef KMP_CANCEL_THREADS 609 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 610 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 611 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 612 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 613 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 614 #endif 615 616 #if KMP_REAL_TIME_FIX 617 // This is a potential fix which allows application with real-time scheduling 618 // policy work. However, decision about the fix is not made yet, so it is 619 // disabled by default. 620 { // Are program started with real-time scheduling policy? 621 int sched = sched_getscheduler(0); 622 if (sched == SCHED_FIFO || sched == SCHED_RR) { 623 // Yes, we are a part of real-time application. Try to increase the 624 // priority of the monitor. 625 struct sched_param param; 626 int max_priority = sched_get_priority_max(sched); 627 int rc; 628 KMP_WARNING(RealTimeSchedNotSupported); 629 sched_getparam(0, ¶m); 630 if (param.sched_priority < max_priority) { 631 param.sched_priority += 1; 632 rc = sched_setscheduler(0, sched, ¶m); 633 if (rc != 0) { 634 int error = errno; 635 kmp_msg_t err_code = KMP_ERR(error); 636 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority), 637 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null); 638 if (__kmp_generate_warnings == kmp_warnings_off) { 639 __kmp_str_free(&err_code.str); 640 } 641 } 642 } else { 643 // We cannot abort here, because number of CPUs may be enough for all 644 // the threads, including the monitor thread, so application could 645 // potentially work... 646 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority), 647 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority), 648 __kmp_msg_null); 649 } 650 } 651 // AC: free thread that waits for monitor started 652 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 653 } 654 #endif // KMP_REAL_TIME_FIX 655 656 KMP_MB(); /* Flush all pending memory write invalidates. */ 657 658 if (__kmp_monitor_wakeups == 1) { 659 interval.tv_sec = 1; 660 interval.tv_nsec = 0; 661 } else { 662 interval.tv_sec = 0; 663 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups); 664 } 665 666 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n")); 667 668 if (__kmp_yield_cycle) { 669 __kmp_yielding_on = 0; /* Start out with yielding shut off */ 670 yield_count = __kmp_yield_off_count; 671 } else { 672 __kmp_yielding_on = 1; /* Yielding is on permanently */ 673 } 674 675 while (!TCR_4(__kmp_global.g.g_done)) { 676 struct timespec now; 677 struct timeval tval; 678 679 /* This thread monitors the state of the system */ 680 681 KA_TRACE(15, ("__kmp_launch_monitor: update\n")); 682 683 status = gettimeofday(&tval, NULL); 684 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 685 TIMEVAL_TO_TIMESPEC(&tval, &now); 686 687 now.tv_sec += interval.tv_sec; 688 now.tv_nsec += interval.tv_nsec; 689 690 if (now.tv_nsec >= KMP_NSEC_PER_SEC) { 691 now.tv_sec += 1; 692 now.tv_nsec -= KMP_NSEC_PER_SEC; 693 } 694 695 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 696 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 697 // AC: the monitor should not fall asleep if g_done has been set 698 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex 699 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond, 700 &__kmp_wait_mx.m_mutex, &now); 701 if (status != 0) { 702 if (status != ETIMEDOUT && status != EINTR) { 703 KMP_SYSFAIL("pthread_cond_timedwait", status); 704 } 705 } 706 } 707 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 708 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 709 710 if (__kmp_yield_cycle) { 711 yield_cycles++; 712 if ((yield_cycles % yield_count) == 0) { 713 if (__kmp_yielding_on) { 714 __kmp_yielding_on = 0; /* Turn it off now */ 715 yield_count = __kmp_yield_off_count; 716 } else { 717 __kmp_yielding_on = 1; /* Turn it on now */ 718 yield_count = __kmp_yield_on_count; 719 } 720 yield_cycles = 0; 721 } 722 } else { 723 __kmp_yielding_on = 1; 724 } 725 726 TCW_4(__kmp_global.g.g_time.dt.t_value, 727 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1); 728 729 KMP_MB(); /* Flush all pending memory write invalidates. */ 730 } 731 732 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n")); 733 734 #ifdef KMP_BLOCK_SIGNALS 735 status = sigfillset(&new_set); 736 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 737 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL); 738 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 739 #endif /* KMP_BLOCK_SIGNALS */ 740 741 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n")); 742 743 if (__kmp_global.g.g_abort != 0) { 744 /* now we need to terminate the worker threads */ 745 /* the value of t_abort is the signal we caught */ 746 747 int gtid; 748 749 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n", 750 __kmp_global.g.g_abort)); 751 752 /* terminate the OpenMP worker threads */ 753 /* TODO this is not valid for sibling threads!! 754 * the uber master might not be 0 anymore.. */ 755 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid) 756 __kmp_terminate_thread(gtid); 757 758 __kmp_cleanup(); 759 760 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n", 761 __kmp_global.g.g_abort)); 762 763 if (__kmp_global.g.g_abort > 0) 764 raise(__kmp_global.g.g_abort); 765 } 766 767 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n")); 768 769 return thr; 770 } 771 #endif // KMP_USE_MONITOR 772 773 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) { 774 pthread_t handle; 775 pthread_attr_t thread_attr; 776 int status; 777 778 th->th.th_info.ds.ds_gtid = gtid; 779 780 #if KMP_STATS_ENABLED 781 // sets up worker thread stats 782 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid); 783 784 // th->th.th_stats is used to transfer thread-specific stats-pointer to 785 // __kmp_launch_worker. So when thread is created (goes into 786 // __kmp_launch_worker) it will set its thread local pointer to 787 // th->th.th_stats 788 if (!KMP_UBER_GTID(gtid)) { 789 th->th.th_stats = __kmp_stats_list->push_back(gtid); 790 } else { 791 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(), 792 // so set the th->th.th_stats field to it. 793 th->th.th_stats = __kmp_stats_thread_ptr; 794 } 795 __kmp_release_tas_lock(&__kmp_stats_lock, gtid); 796 797 #endif // KMP_STATS_ENABLED 798 799 if (KMP_UBER_GTID(gtid)) { 800 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid)); 801 th->th.th_info.ds.ds_thread = pthread_self(); 802 __kmp_set_stack_info(gtid, th); 803 __kmp_check_stack_overlap(th); 804 return; 805 } 806 807 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid)); 808 809 KMP_MB(); /* Flush all pending memory write invalidates. */ 810 811 #ifdef KMP_THREAD_ATTR 812 status = pthread_attr_init(&thread_attr); 813 if (status != 0) { 814 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 815 } 816 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 817 if (status != 0) { 818 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null); 819 } 820 821 /* Set stack size for this thread now. 822 The multiple of 2 is there because on some machines, requesting an unusual 823 stacksize causes the thread to have an offset before the dummy alloca() 824 takes place to create the offset. Since we want the user to have a 825 sufficient stacksize AND support a stack offset, we alloca() twice the 826 offset so that the upcoming alloca() does not eliminate any premade offset, 827 and also gives the user the stack space they requested for all threads */ 828 stack_size += gtid * __kmp_stkoffset * 2; 829 830 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 831 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n", 832 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 833 834 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 835 status = pthread_attr_setstacksize(&thread_attr, stack_size); 836 #ifdef KMP_BACKUP_STKSIZE 837 if (status != 0) { 838 if (!__kmp_env_stksize) { 839 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset; 840 __kmp_stksize = KMP_BACKUP_STKSIZE; 841 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 842 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu " 843 "bytes\n", 844 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 845 status = pthread_attr_setstacksize(&thread_attr, stack_size); 846 } 847 } 848 #endif /* KMP_BACKUP_STKSIZE */ 849 if (status != 0) { 850 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 851 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null); 852 } 853 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 854 855 #endif /* KMP_THREAD_ATTR */ 856 857 status = 858 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th); 859 if (status != 0 || !handle) { // ??? Why do we check handle?? 860 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 861 if (status == EINVAL) { 862 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 863 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null); 864 } 865 if (status == ENOMEM) { 866 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 867 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null); 868 } 869 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 870 if (status == EAGAIN) { 871 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status), 872 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null); 873 } 874 KMP_SYSFAIL("pthread_create", status); 875 } 876 877 th->th.th_info.ds.ds_thread = handle; 878 879 #ifdef KMP_THREAD_ATTR 880 status = pthread_attr_destroy(&thread_attr); 881 if (status) { 882 kmp_msg_t err_code = KMP_ERR(status); 883 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 884 __kmp_msg_null); 885 if (__kmp_generate_warnings == kmp_warnings_off) { 886 __kmp_str_free(&err_code.str); 887 } 888 } 889 #endif /* KMP_THREAD_ATTR */ 890 891 KMP_MB(); /* Flush all pending memory write invalidates. */ 892 893 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid)); 894 895 } // __kmp_create_worker 896 897 #if KMP_USE_MONITOR 898 void __kmp_create_monitor(kmp_info_t *th) { 899 pthread_t handle; 900 pthread_attr_t thread_attr; 901 size_t size; 902 int status; 903 int auto_adj_size = FALSE; 904 905 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { 906 // We don't need monitor thread in case of MAX_BLOCKTIME 907 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of " 908 "MAX blocktime\n")); 909 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op 910 th->th.th_info.ds.ds_gtid = 0; 911 return; 912 } 913 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n")); 914 915 KMP_MB(); /* Flush all pending memory write invalidates. */ 916 917 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR; 918 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR; 919 #if KMP_REAL_TIME_FIX 920 TCW_4(__kmp_global.g.g_time.dt.t_value, 921 -1); // Will use it for synchronization a bit later. 922 #else 923 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 924 #endif // KMP_REAL_TIME_FIX 925 926 #ifdef KMP_THREAD_ATTR 927 if (__kmp_monitor_stksize == 0) { 928 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 929 auto_adj_size = TRUE; 930 } 931 status = pthread_attr_init(&thread_attr); 932 if (status != 0) { 933 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 934 } 935 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 936 if (status != 0) { 937 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null); 938 } 939 940 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 941 status = pthread_attr_getstacksize(&thread_attr, &size); 942 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status); 943 #else 944 size = __kmp_sys_min_stksize; 945 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 946 #endif /* KMP_THREAD_ATTR */ 947 948 if (__kmp_monitor_stksize == 0) { 949 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 950 } 951 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) { 952 __kmp_monitor_stksize = __kmp_sys_min_stksize; 953 } 954 955 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes," 956 "requested stacksize = %lu bytes\n", 957 size, __kmp_monitor_stksize)); 958 959 retry: 960 961 /* Set stack size for this thread now. */ 962 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 963 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,", 964 __kmp_monitor_stksize)); 965 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize); 966 if (status != 0) { 967 if (auto_adj_size) { 968 __kmp_monitor_stksize *= 2; 969 goto retry; 970 } 971 kmp_msg_t err_code = KMP_ERR(status); 972 __kmp_msg(kmp_ms_warning, // should this be fatal? BB 973 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize), 974 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null); 975 if (__kmp_generate_warnings == kmp_warnings_off) { 976 __kmp_str_free(&err_code.str); 977 } 978 } 979 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 980 981 status = 982 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th); 983 984 if (status != 0) { 985 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 986 if (status == EINVAL) { 987 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) { 988 __kmp_monitor_stksize *= 2; 989 goto retry; 990 } 991 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 992 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize), 993 __kmp_msg_null); 994 } 995 if (status == ENOMEM) { 996 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 997 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize), 998 __kmp_msg_null); 999 } 1000 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 1001 if (status == EAGAIN) { 1002 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status), 1003 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null); 1004 } 1005 KMP_SYSFAIL("pthread_create", status); 1006 } 1007 1008 th->th.th_info.ds.ds_thread = handle; 1009 1010 #if KMP_REAL_TIME_FIX 1011 // Wait for the monitor thread is really started and set its *priority*. 1012 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 1013 sizeof(__kmp_global.g.g_time.dt.t_value)); 1014 __kmp_wait_yield_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, 1015 -1, &__kmp_neq_4, NULL); 1016 #endif // KMP_REAL_TIME_FIX 1017 1018 #ifdef KMP_THREAD_ATTR 1019 status = pthread_attr_destroy(&thread_attr); 1020 if (status != 0) { 1021 kmp_msg_t err_code = KMP_ERR(status); 1022 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 1023 __kmp_msg_null); 1024 if (__kmp_generate_warnings == kmp_warnings_off) { 1025 __kmp_str_free(&err_code.str); 1026 } 1027 } 1028 #endif 1029 1030 KMP_MB(); /* Flush all pending memory write invalidates. */ 1031 1032 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n", 1033 th->th.th_info.ds.ds_thread)); 1034 1035 } // __kmp_create_monitor 1036 #endif // KMP_USE_MONITOR 1037 1038 void __kmp_exit_thread(int exit_status) { 1039 pthread_exit((void *)(intptr_t)exit_status); 1040 } // __kmp_exit_thread 1041 1042 #if KMP_USE_MONITOR 1043 void __kmp_resume_monitor(); 1044 1045 void __kmp_reap_monitor(kmp_info_t *th) { 1046 int status; 1047 void *exit_val; 1048 1049 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle" 1050 " %#.8lx\n", 1051 th->th.th_info.ds.ds_thread)); 1052 1053 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR. 1054 // If both tid and gtid are 0, it means the monitor did not ever start. 1055 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down. 1056 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid); 1057 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) { 1058 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n")); 1059 return; 1060 } 1061 1062 KMP_MB(); /* Flush all pending memory write invalidates. */ 1063 1064 /* First, check to see whether the monitor thread exists to wake it up. This 1065 is to avoid performance problem when the monitor sleeps during 1066 blocktime-size interval */ 1067 1068 status = pthread_kill(th->th.th_info.ds.ds_thread, 0); 1069 if (status != ESRCH) { 1070 __kmp_resume_monitor(); // Wake up the monitor thread 1071 } 1072 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n")); 1073 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1074 if (exit_val != th) { 1075 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null); 1076 } 1077 1078 th->th.th_info.ds.ds_tid = KMP_GTID_DNE; 1079 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE; 1080 1081 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle" 1082 " %#.8lx\n", 1083 th->th.th_info.ds.ds_thread)); 1084 1085 KMP_MB(); /* Flush all pending memory write invalidates. */ 1086 } 1087 #endif // KMP_USE_MONITOR 1088 1089 void __kmp_reap_worker(kmp_info_t *th) { 1090 int status; 1091 void *exit_val; 1092 1093 KMP_MB(); /* Flush all pending memory write invalidates. */ 1094 1095 KA_TRACE( 1096 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid)); 1097 1098 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1099 #ifdef KMP_DEBUG 1100 /* Don't expose these to the user until we understand when they trigger */ 1101 if (status != 0) { 1102 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null); 1103 } 1104 if (exit_val != th) { 1105 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, " 1106 "exit_val = %p\n", 1107 th->th.th_info.ds.ds_gtid, exit_val)); 1108 } 1109 #endif /* KMP_DEBUG */ 1110 1111 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n", 1112 th->th.th_info.ds.ds_gtid)); 1113 1114 KMP_MB(); /* Flush all pending memory write invalidates. */ 1115 } 1116 1117 #if KMP_HANDLE_SIGNALS 1118 1119 static void __kmp_null_handler(int signo) { 1120 // Do nothing, for doing SIG_IGN-type actions. 1121 } // __kmp_null_handler 1122 1123 static void __kmp_team_handler(int signo) { 1124 if (__kmp_global.g.g_abort == 0) { 1125 /* Stage 1 signal handler, let's shut down all of the threads */ 1126 #ifdef KMP_DEBUG 1127 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo); 1128 #endif 1129 switch (signo) { 1130 case SIGHUP: 1131 case SIGINT: 1132 case SIGQUIT: 1133 case SIGILL: 1134 case SIGABRT: 1135 case SIGFPE: 1136 case SIGBUS: 1137 case SIGSEGV: 1138 #ifdef SIGSYS 1139 case SIGSYS: 1140 #endif 1141 case SIGTERM: 1142 if (__kmp_debug_buf) { 1143 __kmp_dump_debug_buffer(); 1144 } 1145 KMP_MB(); // Flush all pending memory write invalidates. 1146 TCW_4(__kmp_global.g.g_abort, signo); 1147 KMP_MB(); // Flush all pending memory write invalidates. 1148 TCW_4(__kmp_global.g.g_done, TRUE); 1149 KMP_MB(); // Flush all pending memory write invalidates. 1150 break; 1151 default: 1152 #ifdef KMP_DEBUG 1153 __kmp_debug_printf("__kmp_team_handler: unknown signal type"); 1154 #endif 1155 break; 1156 } 1157 } 1158 } // __kmp_team_handler 1159 1160 static void __kmp_sigaction(int signum, const struct sigaction *act, 1161 struct sigaction *oldact) { 1162 int rc = sigaction(signum, act, oldact); 1163 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc); 1164 } 1165 1166 static void __kmp_install_one_handler(int sig, sig_func_t handler_func, 1167 int parallel_init) { 1168 KMP_MB(); // Flush all pending memory write invalidates. 1169 KB_TRACE(60, 1170 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init)); 1171 if (parallel_init) { 1172 struct sigaction new_action; 1173 struct sigaction old_action; 1174 new_action.sa_handler = handler_func; 1175 new_action.sa_flags = 0; 1176 sigfillset(&new_action.sa_mask); 1177 __kmp_sigaction(sig, &new_action, &old_action); 1178 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) { 1179 sigaddset(&__kmp_sigset, sig); 1180 } else { 1181 // Restore/keep user's handler if one previously installed. 1182 __kmp_sigaction(sig, &old_action, NULL); 1183 } 1184 } else { 1185 // Save initial/system signal handlers to see if user handlers installed. 1186 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]); 1187 } 1188 KMP_MB(); // Flush all pending memory write invalidates. 1189 } // __kmp_install_one_handler 1190 1191 static void __kmp_remove_one_handler(int sig) { 1192 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig)); 1193 if (sigismember(&__kmp_sigset, sig)) { 1194 struct sigaction old; 1195 KMP_MB(); // Flush all pending memory write invalidates. 1196 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old); 1197 if ((old.sa_handler != __kmp_team_handler) && 1198 (old.sa_handler != __kmp_null_handler)) { 1199 // Restore the users signal handler. 1200 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, " 1201 "restoring: sig=%d\n", 1202 sig)); 1203 __kmp_sigaction(sig, &old, NULL); 1204 } 1205 sigdelset(&__kmp_sigset, sig); 1206 KMP_MB(); // Flush all pending memory write invalidates. 1207 } 1208 } // __kmp_remove_one_handler 1209 1210 void __kmp_install_signals(int parallel_init) { 1211 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init)); 1212 if (__kmp_handle_signals || !parallel_init) { 1213 // If ! parallel_init, we do not install handlers, just save original 1214 // handlers. Let us do it even __handle_signals is 0. 1215 sigemptyset(&__kmp_sigset); 1216 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init); 1217 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init); 1218 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init); 1219 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init); 1220 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init); 1221 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init); 1222 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init); 1223 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init); 1224 #ifdef SIGSYS 1225 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init); 1226 #endif // SIGSYS 1227 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init); 1228 #ifdef SIGPIPE 1229 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init); 1230 #endif // SIGPIPE 1231 } 1232 } // __kmp_install_signals 1233 1234 void __kmp_remove_signals(void) { 1235 int sig; 1236 KB_TRACE(10, ("__kmp_remove_signals()\n")); 1237 for (sig = 1; sig < NSIG; ++sig) { 1238 __kmp_remove_one_handler(sig); 1239 } 1240 } // __kmp_remove_signals 1241 1242 #endif // KMP_HANDLE_SIGNALS 1243 1244 void __kmp_enable(int new_state) { 1245 #ifdef KMP_CANCEL_THREADS 1246 int status, old_state; 1247 status = pthread_setcancelstate(new_state, &old_state); 1248 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1249 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE); 1250 #endif 1251 } 1252 1253 void __kmp_disable(int *old_state) { 1254 #ifdef KMP_CANCEL_THREADS 1255 int status; 1256 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state); 1257 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1258 #endif 1259 } 1260 1261 static void __kmp_atfork_prepare(void) { 1262 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 1263 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1264 } 1265 1266 static void __kmp_atfork_parent(void) { 1267 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1268 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1269 } 1270 1271 /* Reset the library so execution in the child starts "all over again" with 1272 clean data structures in initial states. Don't worry about freeing memory 1273 allocated by parent, just abandon it to be safe. */ 1274 static void __kmp_atfork_child(void) { 1275 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1276 /* TODO make sure this is done right for nested/sibling */ 1277 // ATT: Memory leaks are here? TODO: Check it and fix. 1278 /* KMP_ASSERT( 0 ); */ 1279 1280 ++__kmp_fork_count; 1281 1282 #if KMP_AFFINITY_SUPPORTED 1283 #if KMP_OS_LINUX 1284 // reset the affinity in the child to the initial thread 1285 // affinity in the parent 1286 kmp_set_thread_affinity_mask_initial(); 1287 #endif 1288 // Set default not to bind threads tightly in the child (we’re expecting 1289 // over-subscription after the fork and this can improve things for 1290 // scripting languages that use OpenMP inside process-parallel code). 1291 __kmp_affinity_type = affinity_none; 1292 #if OMP_40_ENABLED 1293 if (__kmp_nested_proc_bind.bind_types != NULL) { 1294 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false; 1295 } 1296 #endif // OMP_40_ENABLED 1297 #endif // KMP_AFFINITY_SUPPORTED 1298 1299 __kmp_init_runtime = FALSE; 1300 #if KMP_USE_MONITOR 1301 __kmp_init_monitor = 0; 1302 #endif 1303 __kmp_init_parallel = FALSE; 1304 __kmp_init_middle = FALSE; 1305 __kmp_init_serial = FALSE; 1306 TCW_4(__kmp_init_gtid, FALSE); 1307 __kmp_init_common = FALSE; 1308 1309 TCW_4(__kmp_init_user_locks, FALSE); 1310 #if !KMP_USE_DYNAMIC_LOCK 1311 __kmp_user_lock_table.used = 1; 1312 __kmp_user_lock_table.allocated = 0; 1313 __kmp_user_lock_table.table = NULL; 1314 __kmp_lock_blocks = NULL; 1315 #endif 1316 1317 __kmp_all_nth = 0; 1318 TCW_4(__kmp_nth, 0); 1319 1320 __kmp_thread_pool = NULL; 1321 __kmp_thread_pool_insert_pt = NULL; 1322 __kmp_team_pool = NULL; 1323 1324 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate 1325 here so threadprivate doesn't use stale data */ 1326 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n", 1327 __kmp_threadpriv_cache_list)); 1328 1329 while (__kmp_threadpriv_cache_list != NULL) { 1330 1331 if (*__kmp_threadpriv_cache_list->addr != NULL) { 1332 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n", 1333 &(*__kmp_threadpriv_cache_list->addr))); 1334 1335 *__kmp_threadpriv_cache_list->addr = NULL; 1336 } 1337 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next; 1338 } 1339 1340 __kmp_init_runtime = FALSE; 1341 1342 /* reset statically initialized locks */ 1343 __kmp_init_bootstrap_lock(&__kmp_initz_lock); 1344 __kmp_init_bootstrap_lock(&__kmp_stdio_lock); 1345 __kmp_init_bootstrap_lock(&__kmp_console_lock); 1346 __kmp_init_bootstrap_lock(&__kmp_task_team_lock); 1347 1348 #if USE_ITT_BUILD 1349 __kmp_itt_reset(); // reset ITT's global state 1350 #endif /* USE_ITT_BUILD */ 1351 1352 /* This is necessary to make sure no stale data is left around */ 1353 /* AC: customers complain that we use unsafe routines in the atfork 1354 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen 1355 in dynamic_link when check the presence of shared tbbmalloc library. 1356 Suggestion is to make the library initialization lazier, similar 1357 to what done for __kmpc_begin(). */ 1358 // TODO: synchronize all static initializations with regular library 1359 // startup; look at kmp_global.cpp and etc. 1360 //__kmp_internal_begin (); 1361 } 1362 1363 void __kmp_register_atfork(void) { 1364 if (__kmp_need_register_atfork) { 1365 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent, 1366 __kmp_atfork_child); 1367 KMP_CHECK_SYSFAIL("pthread_atfork", status); 1368 __kmp_need_register_atfork = FALSE; 1369 } 1370 } 1371 1372 void __kmp_suspend_initialize(void) { 1373 int status; 1374 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr); 1375 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1376 status = pthread_condattr_init(&__kmp_suspend_cond_attr); 1377 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1378 } 1379 1380 static void __kmp_suspend_initialize_thread(kmp_info_t *th) { 1381 ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count); 1382 if (th->th.th_suspend_init_count <= __kmp_fork_count) { 1383 /* this means we haven't initialized the suspension pthread objects for this 1384 thread in this instance of the process */ 1385 int status; 1386 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond, 1387 &__kmp_suspend_cond_attr); 1388 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1389 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex, 1390 &__kmp_suspend_mutex_attr); 1391 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1392 *(volatile int *)&th->th.th_suspend_init_count = __kmp_fork_count + 1; 1393 ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count); 1394 } 1395 } 1396 1397 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { 1398 if (th->th.th_suspend_init_count > __kmp_fork_count) { 1399 /* this means we have initialize the suspension pthread objects for this 1400 thread in this instance of the process */ 1401 int status; 1402 1403 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond); 1404 if (status != 0 && status != EBUSY) { 1405 KMP_SYSFAIL("pthread_cond_destroy", status); 1406 } 1407 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex); 1408 if (status != 0 && status != EBUSY) { 1409 KMP_SYSFAIL("pthread_mutex_destroy", status); 1410 } 1411 --th->th.th_suspend_init_count; 1412 KMP_DEBUG_ASSERT(th->th.th_suspend_init_count == __kmp_fork_count); 1413 } 1414 } 1415 1416 // return true if lock obtained, false otherwise 1417 int __kmp_try_suspend_mx(kmp_info_t *th) { 1418 return (pthread_mutex_trylock(&th->th.th_suspend_mx.m_mutex) == 0); 1419 } 1420 1421 void __kmp_lock_suspend_mx(kmp_info_t *th) { 1422 int status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1423 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1424 } 1425 1426 void __kmp_unlock_suspend_mx(kmp_info_t *th) { 1427 int status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1428 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1429 } 1430 1431 /* This routine puts the calling thread to sleep after setting the 1432 sleep bit for the indicated flag variable to true. */ 1433 template <class C> 1434 static inline void __kmp_suspend_template(int th_gtid, C *flag) { 1435 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend); 1436 kmp_info_t *th = __kmp_threads[th_gtid]; 1437 int status; 1438 typename C::flag_t old_spin; 1439 1440 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid, 1441 flag->get())); 1442 1443 __kmp_suspend_initialize_thread(th); 1444 1445 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1446 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1447 1448 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n", 1449 th_gtid, flag->get())); 1450 1451 /* TODO: shouldn't this use release semantics to ensure that 1452 __kmp_suspend_initialize_thread gets called first? */ 1453 old_spin = flag->set_sleeping(); 1454 #if OMP_50_ENABLED 1455 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME && 1456 __kmp_pause_status != kmp_soft_paused) { 1457 flag->unset_sleeping(); 1458 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1459 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1460 return; 1461 } 1462 #endif 1463 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1464 " was %x\n", 1465 th_gtid, flag->get(), flag->load(), old_spin)); 1466 1467 if (flag->done_check_val(old_spin)) { 1468 old_spin = flag->unset_sleeping(); 1469 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1470 "for spin(%p)\n", 1471 th_gtid, flag->get())); 1472 } else { 1473 /* Encapsulate in a loop as the documentation states that this may 1474 "with low probability" return when the condition variable has 1475 not been signaled or broadcast */ 1476 int deactivated = FALSE; 1477 TCW_PTR(th->th.th_sleep_loc, (void *)flag); 1478 1479 while (flag->is_sleeping()) { 1480 #ifdef DEBUG_SUSPEND 1481 char buffer[128]; 1482 __kmp_suspend_count++; 1483 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1484 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid, 1485 buffer); 1486 #endif 1487 // Mark the thread as no longer active (only in the first iteration of the 1488 // loop). 1489 if (!deactivated) { 1490 th->th.th_active = FALSE; 1491 if (th->th.th_active_in_pool) { 1492 th->th.th_active_in_pool = FALSE; 1493 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); 1494 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 1495 } 1496 deactivated = TRUE; 1497 } 1498 1499 #if USE_SUSPEND_TIMEOUT 1500 struct timespec now; 1501 struct timeval tval; 1502 int msecs; 1503 1504 status = gettimeofday(&tval, NULL); 1505 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1506 TIMEVAL_TO_TIMESPEC(&tval, &now); 1507 1508 msecs = (4 * __kmp_dflt_blocktime) + 200; 1509 now.tv_sec += msecs / 1000; 1510 now.tv_nsec += (msecs % 1000) * 1000; 1511 1512 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " 1513 "pthread_cond_timedwait\n", 1514 th_gtid)); 1515 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond, 1516 &th->th.th_suspend_mx.m_mutex, &now); 1517 #else 1518 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform" 1519 " pthread_cond_wait\n", 1520 th_gtid)); 1521 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond, 1522 &th->th.th_suspend_mx.m_mutex); 1523 #endif 1524 1525 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) { 1526 KMP_SYSFAIL("pthread_cond_wait", status); 1527 } 1528 #ifdef KMP_DEBUG 1529 if (status == ETIMEDOUT) { 1530 if (flag->is_sleeping()) { 1531 KF_TRACE(100, 1532 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid)); 1533 } else { 1534 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1535 "not set!\n", 1536 th_gtid)); 1537 } 1538 } else if (flag->is_sleeping()) { 1539 KF_TRACE(100, 1540 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid)); 1541 } 1542 #endif 1543 } // while 1544 1545 // Mark the thread as active again (if it was previous marked as inactive) 1546 if (deactivated) { 1547 th->th.th_active = TRUE; 1548 if (TCR_4(th->th.th_in_pool)) { 1549 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); 1550 th->th.th_active_in_pool = TRUE; 1551 } 1552 } 1553 } 1554 #ifdef DEBUG_SUSPEND 1555 { 1556 char buffer[128]; 1557 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1558 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid, 1559 buffer); 1560 } 1561 #endif 1562 1563 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1564 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1565 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid)); 1566 } 1567 1568 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) { 1569 __kmp_suspend_template(th_gtid, flag); 1570 } 1571 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) { 1572 __kmp_suspend_template(th_gtid, flag); 1573 } 1574 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { 1575 __kmp_suspend_template(th_gtid, flag); 1576 } 1577 1578 /* This routine signals the thread specified by target_gtid to wake up 1579 after setting the sleep bit indicated by the flag argument to FALSE. 1580 The target thread must already have called __kmp_suspend_template() */ 1581 template <class C> 1582 static inline void __kmp_resume_template(int target_gtid, C *flag) { 1583 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1584 kmp_info_t *th = __kmp_threads[target_gtid]; 1585 int status; 1586 1587 #ifdef KMP_DEBUG 1588 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1589 #endif 1590 1591 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n", 1592 gtid, target_gtid)); 1593 KMP_DEBUG_ASSERT(gtid != target_gtid); 1594 1595 __kmp_suspend_initialize_thread(th); 1596 1597 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1598 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1599 1600 if (!flag) { // coming from __kmp_null_resume_wrapper 1601 flag = (C *)CCAST(void *, th->th.th_sleep_loc); 1602 } 1603 1604 // First, check if the flag is null or its type has changed. If so, someone 1605 // else woke it up. 1606 if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type 1607 // simply shows what 1608 // flag was cast to 1609 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1610 "awake: flag(%p)\n", 1611 gtid, target_gtid, NULL)); 1612 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1613 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1614 return; 1615 } else { // if multiple threads are sleeping, flag should be internally 1616 // referring to a specific thread here 1617 typename C::flag_t old_spin = flag->unset_sleeping(); 1618 if (!flag->is_sleeping_val(old_spin)) { 1619 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1620 "awake: flag(%p): " 1621 "%u => %u\n", 1622 gtid, target_gtid, flag->get(), old_spin, flag->load())); 1623 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1624 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1625 return; 1626 } 1627 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1628 "sleep bit for flag's loc(%p): " 1629 "%u => %u\n", 1630 gtid, target_gtid, flag->get(), old_spin, flag->load())); 1631 } 1632 TCW_PTR(th->th.th_sleep_loc, NULL); 1633 1634 #ifdef DEBUG_SUSPEND 1635 { 1636 char buffer[128]; 1637 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1638 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid, 1639 target_gtid, buffer); 1640 } 1641 #endif 1642 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond); 1643 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1644 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1645 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1646 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" 1647 " for T#%d\n", 1648 gtid, target_gtid)); 1649 } 1650 1651 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) { 1652 __kmp_resume_template(target_gtid, flag); 1653 } 1654 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) { 1655 __kmp_resume_template(target_gtid, flag); 1656 } 1657 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { 1658 __kmp_resume_template(target_gtid, flag); 1659 } 1660 1661 #if KMP_USE_MONITOR 1662 void __kmp_resume_monitor() { 1663 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1664 int status; 1665 #ifdef KMP_DEBUG 1666 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1667 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid, 1668 KMP_GTID_MONITOR)); 1669 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR); 1670 #endif 1671 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 1672 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1673 #ifdef DEBUG_SUSPEND 1674 { 1675 char buffer[128]; 1676 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond); 1677 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid, 1678 KMP_GTID_MONITOR, buffer); 1679 } 1680 #endif 1681 status = pthread_cond_signal(&__kmp_wait_cv.c_cond); 1682 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1683 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 1684 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1685 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up" 1686 " for T#%d\n", 1687 gtid, KMP_GTID_MONITOR)); 1688 } 1689 #endif // KMP_USE_MONITOR 1690 1691 void __kmp_yield(int cond) { 1692 if (!cond) 1693 return; 1694 #if KMP_USE_MONITOR 1695 if (!__kmp_yielding_on) 1696 return; 1697 #else 1698 if (__kmp_yield_cycle && !KMP_YIELD_NOW()) 1699 return; 1700 #endif 1701 sched_yield(); 1702 } 1703 1704 void __kmp_gtid_set_specific(int gtid) { 1705 if (__kmp_init_gtid) { 1706 int status; 1707 status = pthread_setspecific(__kmp_gtid_threadprivate_key, 1708 (void *)(intptr_t)(gtid + 1)); 1709 KMP_CHECK_SYSFAIL("pthread_setspecific", status); 1710 } else { 1711 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n")); 1712 } 1713 } 1714 1715 int __kmp_gtid_get_specific() { 1716 int gtid; 1717 if (!__kmp_init_gtid) { 1718 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " 1719 "KMP_GTID_SHUTDOWN\n")); 1720 return KMP_GTID_SHUTDOWN; 1721 } 1722 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key); 1723 if (gtid == 0) { 1724 gtid = KMP_GTID_DNE; 1725 } else { 1726 gtid--; 1727 } 1728 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n", 1729 __kmp_gtid_threadprivate_key, gtid)); 1730 return gtid; 1731 } 1732 1733 double __kmp_read_cpu_time(void) { 1734 /*clock_t t;*/ 1735 struct tms buffer; 1736 1737 /*t =*/times(&buffer); 1738 1739 return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC; 1740 } 1741 1742 int __kmp_read_system_info(struct kmp_sys_info *info) { 1743 int status; 1744 struct rusage r_usage; 1745 1746 memset(info, 0, sizeof(*info)); 1747 1748 status = getrusage(RUSAGE_SELF, &r_usage); 1749 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status); 1750 1751 // The maximum resident set size utilized (in kilobytes) 1752 info->maxrss = r_usage.ru_maxrss; 1753 // The number of page faults serviced without any I/O 1754 info->minflt = r_usage.ru_minflt; 1755 // The number of page faults serviced that required I/O 1756 info->majflt = r_usage.ru_majflt; 1757 // The number of times a process was "swapped" out of memory 1758 info->nswap = r_usage.ru_nswap; 1759 // The number of times the file system had to perform input 1760 info->inblock = r_usage.ru_inblock; 1761 // The number of times the file system had to perform output 1762 info->oublock = r_usage.ru_oublock; 1763 // The number of times a context switch was voluntarily 1764 info->nvcsw = r_usage.ru_nvcsw; 1765 // The number of times a context switch was forced 1766 info->nivcsw = r_usage.ru_nivcsw; 1767 1768 return (status != 0); 1769 } 1770 1771 void __kmp_read_system_time(double *delta) { 1772 double t_ns; 1773 struct timeval tval; 1774 struct timespec stop; 1775 int status; 1776 1777 status = gettimeofday(&tval, NULL); 1778 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1779 TIMEVAL_TO_TIMESPEC(&tval, &stop); 1780 t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start); 1781 *delta = (t_ns * 1e-9); 1782 } 1783 1784 void __kmp_clear_system_time(void) { 1785 struct timeval tval; 1786 int status; 1787 status = gettimeofday(&tval, NULL); 1788 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1789 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start); 1790 } 1791 1792 static int __kmp_get_xproc(void) { 1793 1794 int r = 0; 1795 1796 #if KMP_OS_LINUX || KMP_OS_DRAGONFLY || KMP_OS_FREEBSD || KMP_OS_NETBSD || \ 1797 KMP_OS_OPENBSD || KMP_OS_HURD 1798 1799 r = sysconf(_SC_NPROCESSORS_ONLN); 1800 1801 #elif KMP_OS_DARWIN 1802 1803 // Bug C77011 High "OpenMP Threads and number of active cores". 1804 1805 // Find the number of available CPUs. 1806 kern_return_t rc; 1807 host_basic_info_data_t info; 1808 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT; 1809 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num); 1810 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) { 1811 // Cannot use KA_TRACE() here because this code works before trace support 1812 // is initialized. 1813 r = info.avail_cpus; 1814 } else { 1815 KMP_WARNING(CantGetNumAvailCPU); 1816 KMP_INFORM(AssumedNumCPU); 1817 } 1818 1819 #else 1820 1821 #error "Unknown or unsupported OS." 1822 1823 #endif 1824 1825 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */ 1826 1827 } // __kmp_get_xproc 1828 1829 int __kmp_read_from_file(char const *path, char const *format, ...) { 1830 int result; 1831 va_list args; 1832 1833 va_start(args, format); 1834 FILE *f = fopen(path, "rb"); 1835 if (f == NULL) 1836 return 0; 1837 result = vfscanf(f, format, args); 1838 fclose(f); 1839 1840 return result; 1841 } 1842 1843 void __kmp_runtime_initialize(void) { 1844 int status; 1845 pthread_mutexattr_t mutex_attr; 1846 pthread_condattr_t cond_attr; 1847 1848 if (__kmp_init_runtime) { 1849 return; 1850 } 1851 1852 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1853 if (!__kmp_cpuinfo.initialized) { 1854 __kmp_query_cpuid(&__kmp_cpuinfo); 1855 } 1856 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1857 1858 __kmp_xproc = __kmp_get_xproc(); 1859 1860 if (sysconf(_SC_THREADS)) { 1861 1862 /* Query the maximum number of threads */ 1863 __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX); 1864 if (__kmp_sys_max_nth == -1) { 1865 /* Unlimited threads for NPTL */ 1866 __kmp_sys_max_nth = INT_MAX; 1867 } else if (__kmp_sys_max_nth <= 1) { 1868 /* Can't tell, just use PTHREAD_THREADS_MAX */ 1869 __kmp_sys_max_nth = KMP_MAX_NTH; 1870 } 1871 1872 /* Query the minimum stack size */ 1873 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN); 1874 if (__kmp_sys_min_stksize <= 1) { 1875 __kmp_sys_min_stksize = KMP_MIN_STKSIZE; 1876 } 1877 } 1878 1879 /* Set up minimum number of threads to switch to TLS gtid */ 1880 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; 1881 1882 status = pthread_key_create(&__kmp_gtid_threadprivate_key, 1883 __kmp_internal_end_dest); 1884 KMP_CHECK_SYSFAIL("pthread_key_create", status); 1885 status = pthread_mutexattr_init(&mutex_attr); 1886 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1887 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr); 1888 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1889 status = pthread_condattr_init(&cond_attr); 1890 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1891 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr); 1892 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1893 #if USE_ITT_BUILD 1894 __kmp_itt_initialize(); 1895 #endif /* USE_ITT_BUILD */ 1896 1897 __kmp_init_runtime = TRUE; 1898 } 1899 1900 void __kmp_runtime_destroy(void) { 1901 int status; 1902 1903 if (!__kmp_init_runtime) { 1904 return; // Nothing to do. 1905 } 1906 1907 #if USE_ITT_BUILD 1908 __kmp_itt_destroy(); 1909 #endif /* USE_ITT_BUILD */ 1910 1911 status = pthread_key_delete(__kmp_gtid_threadprivate_key); 1912 KMP_CHECK_SYSFAIL("pthread_key_delete", status); 1913 1914 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex); 1915 if (status != 0 && status != EBUSY) { 1916 KMP_SYSFAIL("pthread_mutex_destroy", status); 1917 } 1918 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond); 1919 if (status != 0 && status != EBUSY) { 1920 KMP_SYSFAIL("pthread_cond_destroy", status); 1921 } 1922 #if KMP_AFFINITY_SUPPORTED 1923 __kmp_affinity_uninitialize(); 1924 #endif 1925 1926 __kmp_init_runtime = FALSE; 1927 } 1928 1929 /* Put the thread to sleep for a time period */ 1930 /* NOTE: not currently used anywhere */ 1931 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); } 1932 1933 /* Calculate the elapsed wall clock time for the user */ 1934 void __kmp_elapsed(double *t) { 1935 int status; 1936 #ifdef FIX_SGI_CLOCK 1937 struct timespec ts; 1938 1939 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); 1940 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status); 1941 *t = 1942 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec; 1943 #else 1944 struct timeval tv; 1945 1946 status = gettimeofday(&tv, NULL); 1947 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1948 *t = 1949 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec; 1950 #endif 1951 } 1952 1953 /* Calculate the elapsed wall clock tick for the user */ 1954 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; } 1955 1956 /* Return the current time stamp in nsec */ 1957 kmp_uint64 __kmp_now_nsec() { 1958 struct timeval t; 1959 gettimeofday(&t, NULL); 1960 kmp_uint64 nsec = (kmp_uint64)KMP_NSEC_PER_SEC * (kmp_uint64)t.tv_sec + 1961 (kmp_uint64)1000 * (kmp_uint64)t.tv_usec; 1962 return nsec; 1963 } 1964 1965 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1966 /* Measure clock ticks per millisecond */ 1967 void __kmp_initialize_system_tick() { 1968 kmp_uint64 now, nsec2, diff; 1969 kmp_uint64 delay = 100000; // 50~100 usec on most machines. 1970 kmp_uint64 nsec = __kmp_now_nsec(); 1971 kmp_uint64 goal = __kmp_hardware_timestamp() + delay; 1972 while ((now = __kmp_hardware_timestamp()) < goal) 1973 ; 1974 nsec2 = __kmp_now_nsec(); 1975 diff = nsec2 - nsec; 1976 if (diff > 0) { 1977 kmp_uint64 tpms = (kmp_uint64)(1e6 * (delay + (now - goal)) / diff); 1978 if (tpms > 0) 1979 __kmp_ticks_per_msec = tpms; 1980 } 1981 } 1982 #endif 1983 1984 /* Determine whether the given address is mapped into the current address 1985 space. */ 1986 1987 int __kmp_is_address_mapped(void *addr) { 1988 1989 int found = 0; 1990 int rc; 1991 1992 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_HURD 1993 1994 /* On GNUish OSes, read the /proc/<pid>/maps pseudo-file to get all the address 1995 ranges mapped into the address space. */ 1996 1997 char *name = __kmp_str_format("/proc/%d/maps", getpid()); 1998 FILE *file = NULL; 1999 2000 file = fopen(name, "r"); 2001 KMP_ASSERT(file != NULL); 2002 2003 for (;;) { 2004 2005 void *beginning = NULL; 2006 void *ending = NULL; 2007 char perms[5]; 2008 2009 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms); 2010 if (rc == EOF) { 2011 break; 2012 } 2013 KMP_ASSERT(rc == 3 && 2014 KMP_STRLEN(perms) == 4); // Make sure all fields are read. 2015 2016 // Ending address is not included in the region, but beginning is. 2017 if ((addr >= beginning) && (addr < ending)) { 2018 perms[2] = 0; // 3th and 4th character does not matter. 2019 if (strcmp(perms, "rw") == 0) { 2020 // Memory we are looking for should be readable and writable. 2021 found = 1; 2022 } 2023 break; 2024 } 2025 } 2026 2027 // Free resources. 2028 fclose(file); 2029 KMP_INTERNAL_FREE(name); 2030 2031 #elif KMP_OS_DARWIN 2032 2033 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory 2034 using vm interface. */ 2035 2036 int buffer; 2037 vm_size_t count; 2038 rc = vm_read_overwrite( 2039 mach_task_self(), // Task to read memory of. 2040 (vm_address_t)(addr), // Address to read from. 2041 1, // Number of bytes to be read. 2042 (vm_address_t)(&buffer), // Address of buffer to save read bytes in. 2043 &count // Address of var to save number of read bytes in. 2044 ); 2045 if (rc == 0) { 2046 // Memory successfully read. 2047 found = 1; 2048 } 2049 2050 #elif KMP_OS_NETBSD 2051 2052 int mib[5]; 2053 mib[0] = CTL_VM; 2054 mib[1] = VM_PROC; 2055 mib[2] = VM_PROC_MAP; 2056 mib[3] = getpid(); 2057 mib[4] = sizeof(struct kinfo_vmentry); 2058 2059 size_t size; 2060 rc = sysctl(mib, __arraycount(mib), NULL, &size, NULL, 0); 2061 KMP_ASSERT(!rc); 2062 KMP_ASSERT(size); 2063 2064 size = size * 4 / 3; 2065 struct kinfo_vmentry *kiv = (struct kinfo_vmentry *)KMP_INTERNAL_MALLOC(size); 2066 KMP_ASSERT(kiv); 2067 2068 rc = sysctl(mib, __arraycount(mib), kiv, &size, NULL, 0); 2069 KMP_ASSERT(!rc); 2070 KMP_ASSERT(size); 2071 2072 for (size_t i = 0; i < size; i++) { 2073 if (kiv[i].kve_start >= (uint64_t)addr && 2074 kiv[i].kve_end <= (uint64_t)addr) { 2075 found = 1; 2076 break; 2077 } 2078 } 2079 KMP_INTERNAL_FREE(kiv); 2080 #elif KMP_OS_DRAGONFLY || KMP_OS_OPENBSD 2081 2082 // FIXME(DragonFly, OpenBSD): Implement this 2083 found = 1; 2084 2085 #else 2086 2087 #error "Unknown or unsupported OS" 2088 2089 #endif 2090 2091 return found; 2092 2093 } // __kmp_is_address_mapped 2094 2095 #ifdef USE_LOAD_BALANCE 2096 2097 #if KMP_OS_DARWIN || KMP_OS_NETBSD 2098 2099 // The function returns the rounded value of the system load average 2100 // during given time interval which depends on the value of 2101 // __kmp_load_balance_interval variable (default is 60 sec, other values 2102 // may be 300 sec or 900 sec). 2103 // It returns -1 in case of error. 2104 int __kmp_get_load_balance(int max) { 2105 double averages[3]; 2106 int ret_avg = 0; 2107 2108 int res = getloadavg(averages, 3); 2109 2110 // Check __kmp_load_balance_interval to determine which of averages to use. 2111 // getloadavg() may return the number of samples less than requested that is 2112 // less than 3. 2113 if (__kmp_load_balance_interval < 180 && (res >= 1)) { 2114 ret_avg = averages[0]; // 1 min 2115 } else if ((__kmp_load_balance_interval >= 180 && 2116 __kmp_load_balance_interval < 600) && 2117 (res >= 2)) { 2118 ret_avg = averages[1]; // 5 min 2119 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) { 2120 ret_avg = averages[2]; // 15 min 2121 } else { // Error occurred 2122 return -1; 2123 } 2124 2125 return ret_avg; 2126 } 2127 2128 #else // Linux* OS 2129 2130 // The fuction returns number of running (not sleeping) threads, or -1 in case 2131 // of error. Error could be reported if Linux* OS kernel too old (without 2132 // "/proc" support). Counting running threads stops if max running threads 2133 // encountered. 2134 int __kmp_get_load_balance(int max) { 2135 static int permanent_error = 0; 2136 static int glb_running_threads = 0; // Saved count of the running threads for 2137 // the thread balance algortihm 2138 static double glb_call_time = 0; /* Thread balance algorithm call time */ 2139 2140 int running_threads = 0; // Number of running threads in the system. 2141 2142 DIR *proc_dir = NULL; // Handle of "/proc/" directory. 2143 struct dirent *proc_entry = NULL; 2144 2145 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path. 2146 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory. 2147 struct dirent *task_entry = NULL; 2148 int task_path_fixed_len; 2149 2150 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path. 2151 int stat_file = -1; 2152 int stat_path_fixed_len; 2153 2154 int total_processes = 0; // Total number of processes in system. 2155 int total_threads = 0; // Total number of threads in system. 2156 2157 double call_time = 0.0; 2158 2159 __kmp_str_buf_init(&task_path); 2160 __kmp_str_buf_init(&stat_path); 2161 2162 __kmp_elapsed(&call_time); 2163 2164 if (glb_call_time && 2165 (call_time - glb_call_time < __kmp_load_balance_interval)) { 2166 running_threads = glb_running_threads; 2167 goto finish; 2168 } 2169 2170 glb_call_time = call_time; 2171 2172 // Do not spend time on scanning "/proc/" if we have a permanent error. 2173 if (permanent_error) { 2174 running_threads = -1; 2175 goto finish; 2176 } 2177 2178 if (max <= 0) { 2179 max = INT_MAX; 2180 } 2181 2182 // Open "/proc/" directory. 2183 proc_dir = opendir("/proc"); 2184 if (proc_dir == NULL) { 2185 // Cannot open "/prroc/". Probably the kernel does not support it. Return an 2186 // error now and in subsequent calls. 2187 running_threads = -1; 2188 permanent_error = 1; 2189 goto finish; 2190 } 2191 2192 // Initialize fixed part of task_path. This part will not change. 2193 __kmp_str_buf_cat(&task_path, "/proc/", 6); 2194 task_path_fixed_len = task_path.used; // Remember number of used characters. 2195 2196 proc_entry = readdir(proc_dir); 2197 while (proc_entry != NULL) { 2198 // Proc entry is a directory and name starts with a digit. Assume it is a 2199 // process' directory. 2200 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) { 2201 2202 ++total_processes; 2203 // Make sure init process is the very first in "/proc", so we can replace 2204 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes == 2205 // 1. We are going to check that total_processes == 1 => d_name == "1" is 2206 // true (where "=>" is implication). Since C++ does not have => operator, 2207 // let us replace it with its equivalent: a => b == ! a || b. 2208 KMP_DEBUG_ASSERT(total_processes != 1 || 2209 strcmp(proc_entry->d_name, "1") == 0); 2210 2211 // Construct task_path. 2212 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/". 2213 __kmp_str_buf_cat(&task_path, proc_entry->d_name, 2214 KMP_STRLEN(proc_entry->d_name)); 2215 __kmp_str_buf_cat(&task_path, "/task", 5); 2216 2217 task_dir = opendir(task_path.str); 2218 if (task_dir == NULL) { 2219 // Process can finish between reading "/proc/" directory entry and 2220 // opening process' "task/" directory. So, in general case we should not 2221 // complain, but have to skip this process and read the next one. But on 2222 // systems with no "task/" support we will spend lot of time to scan 2223 // "/proc/" tree again and again without any benefit. "init" process 2224 // (its pid is 1) should exist always, so, if we cannot open 2225 // "/proc/1/task/" directory, it means "task/" is not supported by 2226 // kernel. Report an error now and in the future. 2227 if (strcmp(proc_entry->d_name, "1") == 0) { 2228 running_threads = -1; 2229 permanent_error = 1; 2230 goto finish; 2231 } 2232 } else { 2233 // Construct fixed part of stat file path. 2234 __kmp_str_buf_clear(&stat_path); 2235 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used); 2236 __kmp_str_buf_cat(&stat_path, "/", 1); 2237 stat_path_fixed_len = stat_path.used; 2238 2239 task_entry = readdir(task_dir); 2240 while (task_entry != NULL) { 2241 // It is a directory and name starts with a digit. 2242 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) { 2243 ++total_threads; 2244 2245 // Consruct complete stat file path. Easiest way would be: 2246 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str, 2247 // task_entry->d_name ); 2248 // but seriae of __kmp_str_buf_cat works a bit faster. 2249 stat_path.used = 2250 stat_path_fixed_len; // Reset stat path to its fixed part. 2251 __kmp_str_buf_cat(&stat_path, task_entry->d_name, 2252 KMP_STRLEN(task_entry->d_name)); 2253 __kmp_str_buf_cat(&stat_path, "/stat", 5); 2254 2255 // Note: Low-level API (open/read/close) is used. High-level API 2256 // (fopen/fclose) works ~ 30 % slower. 2257 stat_file = open(stat_path.str, O_RDONLY); 2258 if (stat_file == -1) { 2259 // We cannot report an error because task (thread) can terminate 2260 // just before reading this file. 2261 } else { 2262 /* Content of "stat" file looks like: 2263 24285 (program) S ... 2264 2265 It is a single line (if program name does not include funny 2266 symbols). First number is a thread id, then name of executable 2267 file name in paretheses, then state of the thread. We need just 2268 thread state. 2269 2270 Good news: Length of program name is 15 characters max. Longer 2271 names are truncated. 2272 2273 Thus, we need rather short buffer: 15 chars for program name + 2274 2 parenthesis, + 3 spaces + ~7 digits of pid = 37. 2275 2276 Bad news: Program name may contain special symbols like space, 2277 closing parenthesis, or even new line. This makes parsing 2278 "stat" file not 100 % reliable. In case of fanny program names 2279 parsing may fail (report incorrect thread state). 2280 2281 Parsing "status" file looks more promissing (due to different 2282 file structure and escaping special symbols) but reading and 2283 parsing of "status" file works slower. 2284 -- ln 2285 */ 2286 char buffer[65]; 2287 int len; 2288 len = read(stat_file, buffer, sizeof(buffer) - 1); 2289 if (len >= 0) { 2290 buffer[len] = 0; 2291 // Using scanf: 2292 // sscanf( buffer, "%*d (%*s) %c ", & state ); 2293 // looks very nice, but searching for a closing parenthesis 2294 // works a bit faster. 2295 char *close_parent = strstr(buffer, ") "); 2296 if (close_parent != NULL) { 2297 char state = *(close_parent + 2); 2298 if (state == 'R') { 2299 ++running_threads; 2300 if (running_threads >= max) { 2301 goto finish; 2302 } 2303 } 2304 } 2305 } 2306 close(stat_file); 2307 stat_file = -1; 2308 } 2309 } 2310 task_entry = readdir(task_dir); 2311 } 2312 closedir(task_dir); 2313 task_dir = NULL; 2314 } 2315 } 2316 proc_entry = readdir(proc_dir); 2317 } 2318 2319 // There _might_ be a timing hole where the thread executing this 2320 // code get skipped in the load balance, and running_threads is 0. 2321 // Assert in the debug builds only!!! 2322 KMP_DEBUG_ASSERT(running_threads > 0); 2323 if (running_threads <= 0) { 2324 running_threads = 1; 2325 } 2326 2327 finish: // Clean up and exit. 2328 if (proc_dir != NULL) { 2329 closedir(proc_dir); 2330 } 2331 __kmp_str_buf_free(&task_path); 2332 if (task_dir != NULL) { 2333 closedir(task_dir); 2334 } 2335 __kmp_str_buf_free(&stat_path); 2336 if (stat_file != -1) { 2337 close(stat_file); 2338 } 2339 2340 glb_running_threads = running_threads; 2341 2342 return running_threads; 2343 2344 } // __kmp_get_load_balance 2345 2346 #endif // KMP_OS_DARWIN 2347 2348 #endif // USE_LOAD_BALANCE 2349 2350 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2351 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || KMP_ARCH_PPC64) 2352 2353 // we really only need the case with 1 argument, because CLANG always build 2354 // a struct of pointers to shared variables referenced in the outlined function 2355 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc, 2356 void *p_argv[] 2357 #if OMPT_SUPPORT 2358 , 2359 void **exit_frame_ptr 2360 #endif 2361 ) { 2362 #if OMPT_SUPPORT 2363 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0); 2364 #endif 2365 2366 switch (argc) { 2367 default: 2368 fprintf(stderr, "Too many args to microtask: %d!\n", argc); 2369 fflush(stderr); 2370 exit(-1); 2371 case 0: 2372 (*pkfn)(>id, &tid); 2373 break; 2374 case 1: 2375 (*pkfn)(>id, &tid, p_argv[0]); 2376 break; 2377 case 2: 2378 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]); 2379 break; 2380 case 3: 2381 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]); 2382 break; 2383 case 4: 2384 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]); 2385 break; 2386 case 5: 2387 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]); 2388 break; 2389 case 6: 2390 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2391 p_argv[5]); 2392 break; 2393 case 7: 2394 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2395 p_argv[5], p_argv[6]); 2396 break; 2397 case 8: 2398 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2399 p_argv[5], p_argv[6], p_argv[7]); 2400 break; 2401 case 9: 2402 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2403 p_argv[5], p_argv[6], p_argv[7], p_argv[8]); 2404 break; 2405 case 10: 2406 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2407 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]); 2408 break; 2409 case 11: 2410 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2411 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]); 2412 break; 2413 case 12: 2414 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2415 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2416 p_argv[11]); 2417 break; 2418 case 13: 2419 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2420 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2421 p_argv[11], p_argv[12]); 2422 break; 2423 case 14: 2424 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2425 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2426 p_argv[11], p_argv[12], p_argv[13]); 2427 break; 2428 case 15: 2429 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2430 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2431 p_argv[11], p_argv[12], p_argv[13], p_argv[14]); 2432 break; 2433 } 2434 2435 #if OMPT_SUPPORT 2436 *exit_frame_ptr = 0; 2437 #endif 2438 2439 return 1; 2440 } 2441 2442 #endif 2443 2444 // end of file // 2445