1 /* 2 * z_Linux_util.cpp -- platform specific routines. 3 */ 4 5 //===----------------------------------------------------------------------===// 6 // 7 // The LLVM Compiler Infrastructure 8 // 9 // This file is dual licensed under the MIT and the University of Illinois Open 10 // Source Licenses. See LICENSE.txt for details. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "kmp.h" 15 #include "kmp_affinity.h" 16 #include "kmp_i18n.h" 17 #include "kmp_io.h" 18 #include "kmp_itt.h" 19 #include "kmp_lock.h" 20 #include "kmp_stats.h" 21 #include "kmp_str.h" 22 #include "kmp_wait_release.h" 23 #include "kmp_wrapper_getpid.h" 24 25 #if !KMP_OS_FREEBSD && !KMP_OS_NETBSD 26 #include <alloca.h> 27 #endif 28 #include <math.h> // HUGE_VAL. 29 #include <sys/resource.h> 30 #include <sys/syscall.h> 31 #include <sys/time.h> 32 #include <sys/times.h> 33 #include <unistd.h> 34 35 #if KMP_OS_LINUX && !KMP_OS_CNK 36 #include <sys/sysinfo.h> 37 #if KMP_USE_FUTEX 38 // We should really include <futex.h>, but that causes compatibility problems on 39 // different Linux* OS distributions that either require that you include (or 40 // break when you try to include) <pci/types.h>. Since all we need is the two 41 // macros below (which are part of the kernel ABI, so can't change) we just 42 // define the constants here and don't include <futex.h> 43 #ifndef FUTEX_WAIT 44 #define FUTEX_WAIT 0 45 #endif 46 #ifndef FUTEX_WAKE 47 #define FUTEX_WAKE 1 48 #endif 49 #endif 50 #elif KMP_OS_DARWIN 51 #include <mach/mach.h> 52 #include <sys/sysctl.h> 53 #elif KMP_OS_FREEBSD 54 #include <pthread_np.h> 55 #endif 56 57 #include <ctype.h> 58 #include <dirent.h> 59 #include <fcntl.h> 60 61 #include "tsan_annotations.h" 62 63 struct kmp_sys_timer { 64 struct timespec start; 65 }; 66 67 // Convert timespec to nanoseconds. 68 #define TS2NS(timespec) (((timespec).tv_sec * 1e9) + (timespec).tv_nsec) 69 70 static struct kmp_sys_timer __kmp_sys_timer_data; 71 72 #if KMP_HANDLE_SIGNALS 73 typedef void (*sig_func_t)(int); 74 STATIC_EFI2_WORKAROUND struct sigaction __kmp_sighldrs[NSIG]; 75 static sigset_t __kmp_sigset; 76 #endif 77 78 static int __kmp_init_runtime = FALSE; 79 80 static int __kmp_fork_count = 0; 81 82 static pthread_condattr_t __kmp_suspend_cond_attr; 83 static pthread_mutexattr_t __kmp_suspend_mutex_attr; 84 85 static kmp_cond_align_t __kmp_wait_cv; 86 static kmp_mutex_align_t __kmp_wait_mx; 87 88 kmp_uint64 __kmp_ticks_per_msec = 1000000; 89 90 #ifdef DEBUG_SUSPEND 91 static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) { 92 KMP_SNPRINTF(buffer, 128, "(cond (lock (%ld, %d)), (descr (%p)))", 93 cond->c_cond.__c_lock.__status, cond->c_cond.__c_lock.__spinlock, 94 cond->c_cond.__c_waiting); 95 } 96 #endif 97 98 #if (KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED) 99 100 /* Affinity support */ 101 102 void __kmp_affinity_bind_thread(int which) { 103 KMP_ASSERT2(KMP_AFFINITY_CAPABLE(), 104 "Illegal set affinity operation when not capable"); 105 106 kmp_affin_mask_t *mask; 107 KMP_CPU_ALLOC_ON_STACK(mask); 108 KMP_CPU_ZERO(mask); 109 KMP_CPU_SET(which, mask); 110 __kmp_set_system_affinity(mask, TRUE); 111 KMP_CPU_FREE_FROM_STACK(mask); 112 } 113 114 /* Determine if we can access affinity functionality on this version of 115 * Linux* OS by checking __NR_sched_{get,set}affinity system calls, and set 116 * __kmp_affin_mask_size to the appropriate value (0 means not capable). */ 117 void __kmp_affinity_determine_capable(const char *env_var) { 118 // Check and see if the OS supports thread affinity. 119 120 #define KMP_CPU_SET_SIZE_LIMIT (1024 * 1024) 121 122 int gCode; 123 int sCode; 124 unsigned char *buf; 125 buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT); 126 127 // If Linux* OS: 128 // If the syscall fails or returns a suggestion for the size, 129 // then we don't have to search for an appropriate size. 130 gCode = syscall(__NR_sched_getaffinity, 0, KMP_CPU_SET_SIZE_LIMIT, buf); 131 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 132 "initial getaffinity call returned %d errno = %d\n", 133 gCode, errno)); 134 135 // if ((gCode < 0) && (errno == ENOSYS)) 136 if (gCode < 0) { 137 // System call not supported 138 if (__kmp_affinity_verbose || 139 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) && 140 (__kmp_affinity_type != affinity_default) && 141 (__kmp_affinity_type != affinity_disabled))) { 142 int error = errno; 143 kmp_msg_t err_code = KMP_ERR(error); 144 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 145 err_code, __kmp_msg_null); 146 if (__kmp_generate_warnings == kmp_warnings_off) { 147 __kmp_str_free(&err_code.str); 148 } 149 } 150 KMP_AFFINITY_DISABLE(); 151 KMP_INTERNAL_FREE(buf); 152 return; 153 } 154 if (gCode > 0) { // Linux* OS only 155 // The optimal situation: the OS returns the size of the buffer it expects. 156 // 157 // A verification of correct behavior is that Isetaffinity on a NULL 158 // buffer with the same size fails with errno set to EFAULT. 159 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL); 160 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 161 "setaffinity for mask size %d returned %d errno = %d\n", 162 gCode, sCode, errno)); 163 if (sCode < 0) { 164 if (errno == ENOSYS) { 165 if (__kmp_affinity_verbose || 166 (__kmp_affinity_warnings && 167 (__kmp_affinity_type != affinity_none) && 168 (__kmp_affinity_type != affinity_default) && 169 (__kmp_affinity_type != affinity_disabled))) { 170 int error = errno; 171 kmp_msg_t err_code = KMP_ERR(error); 172 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var), 173 err_code, __kmp_msg_null); 174 if (__kmp_generate_warnings == kmp_warnings_off) { 175 __kmp_str_free(&err_code.str); 176 } 177 } 178 KMP_AFFINITY_DISABLE(); 179 KMP_INTERNAL_FREE(buf); 180 } 181 if (errno == EFAULT) { 182 KMP_AFFINITY_ENABLE(gCode); 183 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 184 "affinity supported (mask size %d)\n", 185 (int)__kmp_affin_mask_size)); 186 KMP_INTERNAL_FREE(buf); 187 return; 188 } 189 } 190 } 191 192 // Call the getaffinity system call repeatedly with increasing set sizes 193 // until we succeed, or reach an upper bound on the search. 194 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 195 "searching for proper set size\n")); 196 int size; 197 for (size = 1; size <= KMP_CPU_SET_SIZE_LIMIT; size *= 2) { 198 gCode = syscall(__NR_sched_getaffinity, 0, size, buf); 199 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 200 "getaffinity for mask size %d returned %d errno = %d\n", 201 size, gCode, errno)); 202 203 if (gCode < 0) { 204 if (errno == ENOSYS) { 205 // We shouldn't get here 206 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 207 "inconsistent OS call behavior: errno == ENOSYS for mask " 208 "size %d\n", 209 size)); 210 if (__kmp_affinity_verbose || 211 (__kmp_affinity_warnings && 212 (__kmp_affinity_type != affinity_none) && 213 (__kmp_affinity_type != affinity_default) && 214 (__kmp_affinity_type != affinity_disabled))) { 215 int error = errno; 216 kmp_msg_t err_code = KMP_ERR(error); 217 __kmp_msg(kmp_ms_warning, KMP_MSG(GetAffSysCallNotSupported, env_var), 218 err_code, __kmp_msg_null); 219 if (__kmp_generate_warnings == kmp_warnings_off) { 220 __kmp_str_free(&err_code.str); 221 } 222 } 223 KMP_AFFINITY_DISABLE(); 224 KMP_INTERNAL_FREE(buf); 225 return; 226 } 227 continue; 228 } 229 230 sCode = syscall(__NR_sched_setaffinity, 0, gCode, NULL); 231 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 232 "setaffinity for mask size %d returned %d errno = %d\n", 233 gCode, sCode, errno)); 234 if (sCode < 0) { 235 if (errno == ENOSYS) { // Linux* OS only 236 // We shouldn't get here 237 KA_TRACE(30, ("__kmp_affinity_determine_capable: " 238 "inconsistent OS call behavior: errno == ENOSYS for mask " 239 "size %d\n", 240 size)); 241 if (__kmp_affinity_verbose || 242 (__kmp_affinity_warnings && 243 (__kmp_affinity_type != affinity_none) && 244 (__kmp_affinity_type != affinity_default) && 245 (__kmp_affinity_type != affinity_disabled))) { 246 int error = errno; 247 kmp_msg_t err_code = KMP_ERR(error); 248 __kmp_msg(kmp_ms_warning, KMP_MSG(SetAffSysCallNotSupported, env_var), 249 err_code, __kmp_msg_null); 250 if (__kmp_generate_warnings == kmp_warnings_off) { 251 __kmp_str_free(&err_code.str); 252 } 253 } 254 KMP_AFFINITY_DISABLE(); 255 KMP_INTERNAL_FREE(buf); 256 return; 257 } 258 if (errno == EFAULT) { 259 KMP_AFFINITY_ENABLE(gCode); 260 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 261 "affinity supported (mask size %d)\n", 262 (int)__kmp_affin_mask_size)); 263 KMP_INTERNAL_FREE(buf); 264 return; 265 } 266 } 267 } 268 // save uncaught error code 269 // int error = errno; 270 KMP_INTERNAL_FREE(buf); 271 // restore uncaught error code, will be printed at the next KMP_WARNING below 272 // errno = error; 273 274 // Affinity is not supported 275 KMP_AFFINITY_DISABLE(); 276 KA_TRACE(10, ("__kmp_affinity_determine_capable: " 277 "cannot determine mask size - affinity not supported\n")); 278 if (__kmp_affinity_verbose || 279 (__kmp_affinity_warnings && (__kmp_affinity_type != affinity_none) && 280 (__kmp_affinity_type != affinity_default) && 281 (__kmp_affinity_type != affinity_disabled))) { 282 KMP_WARNING(AffCantGetMaskSize, env_var); 283 } 284 } 285 286 #endif // KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 287 288 #if KMP_USE_FUTEX 289 290 int __kmp_futex_determine_capable() { 291 int loc = 0; 292 int rc = syscall(__NR_futex, &loc, FUTEX_WAKE, 1, NULL, NULL, 0); 293 int retval = (rc == 0) || (errno != ENOSYS); 294 295 KA_TRACE(10, 296 ("__kmp_futex_determine_capable: rc = %d errno = %d\n", rc, errno)); 297 KA_TRACE(10, ("__kmp_futex_determine_capable: futex syscall%s supported\n", 298 retval ? "" : " not")); 299 300 return retval; 301 } 302 303 #endif // KMP_USE_FUTEX 304 305 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (!KMP_ASM_INTRINS) 306 /* Only 32-bit "add-exchange" instruction on IA-32 architecture causes us to 307 use compare_and_store for these routines */ 308 309 kmp_int8 __kmp_test_then_or8(volatile kmp_int8 *p, kmp_int8 d) { 310 kmp_int8 old_value, new_value; 311 312 old_value = TCR_1(*p); 313 new_value = old_value | d; 314 315 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 316 KMP_CPU_PAUSE(); 317 old_value = TCR_1(*p); 318 new_value = old_value | d; 319 } 320 return old_value; 321 } 322 323 kmp_int8 __kmp_test_then_and8(volatile kmp_int8 *p, kmp_int8 d) { 324 kmp_int8 old_value, new_value; 325 326 old_value = TCR_1(*p); 327 new_value = old_value & d; 328 329 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 330 KMP_CPU_PAUSE(); 331 old_value = TCR_1(*p); 332 new_value = old_value & d; 333 } 334 return old_value; 335 } 336 337 kmp_uint32 __kmp_test_then_or32(volatile kmp_uint32 *p, kmp_uint32 d) { 338 kmp_uint32 old_value, new_value; 339 340 old_value = TCR_4(*p); 341 new_value = old_value | d; 342 343 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 344 KMP_CPU_PAUSE(); 345 old_value = TCR_4(*p); 346 new_value = old_value | d; 347 } 348 return old_value; 349 } 350 351 kmp_uint32 __kmp_test_then_and32(volatile kmp_uint32 *p, kmp_uint32 d) { 352 kmp_uint32 old_value, new_value; 353 354 old_value = TCR_4(*p); 355 new_value = old_value & d; 356 357 while (!KMP_COMPARE_AND_STORE_REL32(p, old_value, new_value)) { 358 KMP_CPU_PAUSE(); 359 old_value = TCR_4(*p); 360 new_value = old_value & d; 361 } 362 return old_value; 363 } 364 365 #if KMP_ARCH_X86 366 kmp_int8 __kmp_test_then_add8(volatile kmp_int8 *p, kmp_int8 d) { 367 kmp_int8 old_value, new_value; 368 369 old_value = TCR_1(*p); 370 new_value = old_value + d; 371 372 while (!KMP_COMPARE_AND_STORE_REL8(p, old_value, new_value)) { 373 KMP_CPU_PAUSE(); 374 old_value = TCR_1(*p); 375 new_value = old_value + d; 376 } 377 return old_value; 378 } 379 380 kmp_int64 __kmp_test_then_add64(volatile kmp_int64 *p, kmp_int64 d) { 381 kmp_int64 old_value, new_value; 382 383 old_value = TCR_8(*p); 384 new_value = old_value + d; 385 386 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 387 KMP_CPU_PAUSE(); 388 old_value = TCR_8(*p); 389 new_value = old_value + d; 390 } 391 return old_value; 392 } 393 #endif /* KMP_ARCH_X86 */ 394 395 kmp_uint64 __kmp_test_then_or64(volatile kmp_uint64 *p, kmp_uint64 d) { 396 kmp_uint64 old_value, new_value; 397 398 old_value = TCR_8(*p); 399 new_value = old_value | d; 400 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 401 KMP_CPU_PAUSE(); 402 old_value = TCR_8(*p); 403 new_value = old_value | d; 404 } 405 return old_value; 406 } 407 408 kmp_uint64 __kmp_test_then_and64(volatile kmp_uint64 *p, kmp_uint64 d) { 409 kmp_uint64 old_value, new_value; 410 411 old_value = TCR_8(*p); 412 new_value = old_value & d; 413 while (!KMP_COMPARE_AND_STORE_REL64(p, old_value, new_value)) { 414 KMP_CPU_PAUSE(); 415 old_value = TCR_8(*p); 416 new_value = old_value & d; 417 } 418 return old_value; 419 } 420 421 #endif /* (KMP_ARCH_X86 || KMP_ARCH_X86_64) && (! KMP_ASM_INTRINS) */ 422 423 void __kmp_terminate_thread(int gtid) { 424 int status; 425 kmp_info_t *th = __kmp_threads[gtid]; 426 427 if (!th) 428 return; 429 430 #ifdef KMP_CANCEL_THREADS 431 KA_TRACE(10, ("__kmp_terminate_thread: kill (%d)\n", gtid)); 432 status = pthread_cancel(th->th.th_info.ds.ds_thread); 433 if (status != 0 && status != ESRCH) { 434 __kmp_fatal(KMP_MSG(CantTerminateWorkerThread), KMP_ERR(status), 435 __kmp_msg_null); 436 } 437 #endif 438 __kmp_yield(TRUE); 439 } // 440 441 /* Set thread stack info according to values returned by pthread_getattr_np(). 442 If values are unreasonable, assume call failed and use incremental stack 443 refinement method instead. Returns TRUE if the stack parameters could be 444 determined exactly, FALSE if incremental refinement is necessary. */ 445 static kmp_int32 __kmp_set_stack_info(int gtid, kmp_info_t *th) { 446 int stack_data; 447 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD 448 /* Linux* OS only -- no pthread_getattr_np support on OS X* */ 449 pthread_attr_t attr; 450 int status; 451 size_t size = 0; 452 void *addr = 0; 453 454 /* Always do incremental stack refinement for ubermaster threads since the 455 initial thread stack range can be reduced by sibling thread creation so 456 pthread_attr_getstack may cause thread gtid aliasing */ 457 if (!KMP_UBER_GTID(gtid)) { 458 459 /* Fetch the real thread attributes */ 460 status = pthread_attr_init(&attr); 461 KMP_CHECK_SYSFAIL("pthread_attr_init", status); 462 #if KMP_OS_FREEBSD || KMP_OS_NETBSD 463 status = pthread_attr_get_np(pthread_self(), &attr); 464 KMP_CHECK_SYSFAIL("pthread_attr_get_np", status); 465 #else 466 status = pthread_getattr_np(pthread_self(), &attr); 467 KMP_CHECK_SYSFAIL("pthread_getattr_np", status); 468 #endif 469 status = pthread_attr_getstack(&attr, &addr, &size); 470 KMP_CHECK_SYSFAIL("pthread_attr_getstack", status); 471 KA_TRACE(60, 472 ("__kmp_set_stack_info: T#%d pthread_attr_getstack returned size:" 473 " %lu, low addr: %p\n", 474 gtid, size, addr)); 475 status = pthread_attr_destroy(&attr); 476 KMP_CHECK_SYSFAIL("pthread_attr_destroy", status); 477 } 478 479 if (size != 0 && addr != 0) { // was stack parameter determination successful? 480 /* Store the correct base and size */ 481 TCW_PTR(th->th.th_info.ds.ds_stackbase, (((char *)addr) + size)); 482 TCW_PTR(th->th.th_info.ds.ds_stacksize, size); 483 TCW_4(th->th.th_info.ds.ds_stackgrow, FALSE); 484 return TRUE; 485 } 486 #endif /* KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD */ 487 /* Use incremental refinement starting from initial conservative estimate */ 488 TCW_PTR(th->th.th_info.ds.ds_stacksize, 0); 489 TCW_PTR(th->th.th_info.ds.ds_stackbase, &stack_data); 490 TCW_4(th->th.th_info.ds.ds_stackgrow, TRUE); 491 return FALSE; 492 } 493 494 static void *__kmp_launch_worker(void *thr) { 495 int status, old_type, old_state; 496 #ifdef KMP_BLOCK_SIGNALS 497 sigset_t new_set, old_set; 498 #endif /* KMP_BLOCK_SIGNALS */ 499 void *exit_val; 500 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD 501 void *volatile padding = 0; 502 #endif 503 int gtid; 504 505 gtid = ((kmp_info_t *)thr)->th.th_info.ds.ds_gtid; 506 __kmp_gtid_set_specific(gtid); 507 #ifdef KMP_TDATA_GTID 508 __kmp_gtid = gtid; 509 #endif 510 #if KMP_STATS_ENABLED 511 // set thread local index to point to thread-specific stats 512 __kmp_stats_thread_ptr = ((kmp_info_t *)thr)->th.th_stats; 513 __kmp_stats_thread_ptr->startLife(); 514 KMP_SET_THREAD_STATE(IDLE); 515 KMP_INIT_PARTITIONED_TIMERS(OMP_idle); 516 #endif 517 518 #if USE_ITT_BUILD 519 __kmp_itt_thread_name(gtid); 520 #endif /* USE_ITT_BUILD */ 521 522 #if KMP_AFFINITY_SUPPORTED 523 __kmp_affinity_set_init_mask(gtid, FALSE); 524 #endif 525 526 #ifdef KMP_CANCEL_THREADS 527 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 528 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 529 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 530 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 531 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 532 #endif 533 534 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 535 // Set FP control regs to be a copy of the parallel initialization thread's. 536 __kmp_clear_x87_fpu_status_word(); 537 __kmp_load_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); 538 __kmp_load_mxcsr(&__kmp_init_mxcsr); 539 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 540 541 #ifdef KMP_BLOCK_SIGNALS 542 status = sigfillset(&new_set); 543 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 544 status = pthread_sigmask(SIG_BLOCK, &new_set, &old_set); 545 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 546 #endif /* KMP_BLOCK_SIGNALS */ 547 548 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD 549 if (__kmp_stkoffset > 0 && gtid > 0) { 550 padding = KMP_ALLOCA(gtid * __kmp_stkoffset); 551 } 552 #endif 553 554 KMP_MB(); 555 __kmp_set_stack_info(gtid, (kmp_info_t *)thr); 556 557 __kmp_check_stack_overlap((kmp_info_t *)thr); 558 559 exit_val = __kmp_launch_thread((kmp_info_t *)thr); 560 561 #ifdef KMP_BLOCK_SIGNALS 562 status = pthread_sigmask(SIG_SETMASK, &old_set, NULL); 563 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 564 #endif /* KMP_BLOCK_SIGNALS */ 565 566 return exit_val; 567 } 568 569 #if KMP_USE_MONITOR 570 /* The monitor thread controls all of the threads in the complex */ 571 572 static void *__kmp_launch_monitor(void *thr) { 573 int status, old_type, old_state; 574 #ifdef KMP_BLOCK_SIGNALS 575 sigset_t new_set; 576 #endif /* KMP_BLOCK_SIGNALS */ 577 struct timespec interval; 578 int yield_count; 579 int yield_cycles = 0; 580 581 KMP_MB(); /* Flush all pending memory write invalidates. */ 582 583 KA_TRACE(10, ("__kmp_launch_monitor: #1 launched\n")); 584 585 /* register us as the monitor thread */ 586 __kmp_gtid_set_specific(KMP_GTID_MONITOR); 587 #ifdef KMP_TDATA_GTID 588 __kmp_gtid = KMP_GTID_MONITOR; 589 #endif 590 591 KMP_MB(); 592 593 #if USE_ITT_BUILD 594 // Instruct Intel(R) Threading Tools to ignore monitor thread. 595 __kmp_itt_thread_ignore(); 596 #endif /* USE_ITT_BUILD */ 597 598 __kmp_set_stack_info(((kmp_info_t *)thr)->th.th_info.ds.ds_gtid, 599 (kmp_info_t *)thr); 600 601 __kmp_check_stack_overlap((kmp_info_t *)thr); 602 603 #ifdef KMP_CANCEL_THREADS 604 status = pthread_setcanceltype(PTHREAD_CANCEL_ASYNCHRONOUS, &old_type); 605 KMP_CHECK_SYSFAIL("pthread_setcanceltype", status); 606 // josh todo: isn't PTHREAD_CANCEL_ENABLE default for newly-created threads? 607 status = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &old_state); 608 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 609 #endif 610 611 #if KMP_REAL_TIME_FIX 612 // This is a potential fix which allows application with real-time scheduling 613 // policy work. However, decision about the fix is not made yet, so it is 614 // disabled by default. 615 { // Are program started with real-time scheduling policy? 616 int sched = sched_getscheduler(0); 617 if (sched == SCHED_FIFO || sched == SCHED_RR) { 618 // Yes, we are a part of real-time application. Try to increase the 619 // priority of the monitor. 620 struct sched_param param; 621 int max_priority = sched_get_priority_max(sched); 622 int rc; 623 KMP_WARNING(RealTimeSchedNotSupported); 624 sched_getparam(0, ¶m); 625 if (param.sched_priority < max_priority) { 626 param.sched_priority += 1; 627 rc = sched_setscheduler(0, sched, ¶m); 628 if (rc != 0) { 629 int error = errno; 630 kmp_msg_t err_code = KMP_ERR(error); 631 __kmp_msg(kmp_ms_warning, KMP_MSG(CantChangeMonitorPriority), 632 err_code, KMP_MSG(MonitorWillStarve), __kmp_msg_null); 633 if (__kmp_generate_warnings == kmp_warnings_off) { 634 __kmp_str_free(&err_code.str); 635 } 636 } 637 } else { 638 // We cannot abort here, because number of CPUs may be enough for all 639 // the threads, including the monitor thread, so application could 640 // potentially work... 641 __kmp_msg(kmp_ms_warning, KMP_MSG(RunningAtMaxPriority), 642 KMP_MSG(MonitorWillStarve), KMP_HNT(RunningAtMaxPriority), 643 __kmp_msg_null); 644 } 645 } 646 // AC: free thread that waits for monitor started 647 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 648 } 649 #endif // KMP_REAL_TIME_FIX 650 651 KMP_MB(); /* Flush all pending memory write invalidates. */ 652 653 if (__kmp_monitor_wakeups == 1) { 654 interval.tv_sec = 1; 655 interval.tv_nsec = 0; 656 } else { 657 interval.tv_sec = 0; 658 interval.tv_nsec = (KMP_NSEC_PER_SEC / __kmp_monitor_wakeups); 659 } 660 661 KA_TRACE(10, ("__kmp_launch_monitor: #2 monitor\n")); 662 663 if (__kmp_yield_cycle) { 664 __kmp_yielding_on = 0; /* Start out with yielding shut off */ 665 yield_count = __kmp_yield_off_count; 666 } else { 667 __kmp_yielding_on = 1; /* Yielding is on permanently */ 668 } 669 670 while (!TCR_4(__kmp_global.g.g_done)) { 671 struct timespec now; 672 struct timeval tval; 673 674 /* This thread monitors the state of the system */ 675 676 KA_TRACE(15, ("__kmp_launch_monitor: update\n")); 677 678 status = gettimeofday(&tval, NULL); 679 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 680 TIMEVAL_TO_TIMESPEC(&tval, &now); 681 682 now.tv_sec += interval.tv_sec; 683 now.tv_nsec += interval.tv_nsec; 684 685 if (now.tv_nsec >= KMP_NSEC_PER_SEC) { 686 now.tv_sec += 1; 687 now.tv_nsec -= KMP_NSEC_PER_SEC; 688 } 689 690 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 691 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 692 // AC: the monitor should not fall asleep if g_done has been set 693 if (!TCR_4(__kmp_global.g.g_done)) { // check once more under mutex 694 status = pthread_cond_timedwait(&__kmp_wait_cv.c_cond, 695 &__kmp_wait_mx.m_mutex, &now); 696 if (status != 0) { 697 if (status != ETIMEDOUT && status != EINTR) { 698 KMP_SYSFAIL("pthread_cond_timedwait", status); 699 } 700 } 701 } 702 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 703 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 704 705 if (__kmp_yield_cycle) { 706 yield_cycles++; 707 if ((yield_cycles % yield_count) == 0) { 708 if (__kmp_yielding_on) { 709 __kmp_yielding_on = 0; /* Turn it off now */ 710 yield_count = __kmp_yield_off_count; 711 } else { 712 __kmp_yielding_on = 1; /* Turn it on now */ 713 yield_count = __kmp_yield_on_count; 714 } 715 yield_cycles = 0; 716 } 717 } else { 718 __kmp_yielding_on = 1; 719 } 720 721 TCW_4(__kmp_global.g.g_time.dt.t_value, 722 TCR_4(__kmp_global.g.g_time.dt.t_value) + 1); 723 724 KMP_MB(); /* Flush all pending memory write invalidates. */ 725 } 726 727 KA_TRACE(10, ("__kmp_launch_monitor: #3 cleanup\n")); 728 729 #ifdef KMP_BLOCK_SIGNALS 730 status = sigfillset(&new_set); 731 KMP_CHECK_SYSFAIL_ERRNO("sigfillset", status); 732 status = pthread_sigmask(SIG_UNBLOCK, &new_set, NULL); 733 KMP_CHECK_SYSFAIL("pthread_sigmask", status); 734 #endif /* KMP_BLOCK_SIGNALS */ 735 736 KA_TRACE(10, ("__kmp_launch_monitor: #4 finished\n")); 737 738 if (__kmp_global.g.g_abort != 0) { 739 /* now we need to terminate the worker threads */ 740 /* the value of t_abort is the signal we caught */ 741 742 int gtid; 743 744 KA_TRACE(10, ("__kmp_launch_monitor: #5 terminate sig=%d\n", 745 __kmp_global.g.g_abort)); 746 747 /* terminate the OpenMP worker threads */ 748 /* TODO this is not valid for sibling threads!! 749 * the uber master might not be 0 anymore.. */ 750 for (gtid = 1; gtid < __kmp_threads_capacity; ++gtid) 751 __kmp_terminate_thread(gtid); 752 753 __kmp_cleanup(); 754 755 KA_TRACE(10, ("__kmp_launch_monitor: #6 raise sig=%d\n", 756 __kmp_global.g.g_abort)); 757 758 if (__kmp_global.g.g_abort > 0) 759 raise(__kmp_global.g.g_abort); 760 } 761 762 KA_TRACE(10, ("__kmp_launch_monitor: #7 exit\n")); 763 764 return thr; 765 } 766 #endif // KMP_USE_MONITOR 767 768 void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size) { 769 pthread_t handle; 770 pthread_attr_t thread_attr; 771 int status; 772 773 th->th.th_info.ds.ds_gtid = gtid; 774 775 #if KMP_STATS_ENABLED 776 // sets up worker thread stats 777 __kmp_acquire_tas_lock(&__kmp_stats_lock, gtid); 778 779 // th->th.th_stats is used to transfer thread-specific stats-pointer to 780 // __kmp_launch_worker. So when thread is created (goes into 781 // __kmp_launch_worker) it will set its thread local pointer to 782 // th->th.th_stats 783 if (!KMP_UBER_GTID(gtid)) { 784 th->th.th_stats = __kmp_stats_list->push_back(gtid); 785 } else { 786 // For root threads, __kmp_stats_thread_ptr is set in __kmp_register_root(), 787 // so set the th->th.th_stats field to it. 788 th->th.th_stats = __kmp_stats_thread_ptr; 789 } 790 __kmp_release_tas_lock(&__kmp_stats_lock, gtid); 791 792 #endif // KMP_STATS_ENABLED 793 794 if (KMP_UBER_GTID(gtid)) { 795 KA_TRACE(10, ("__kmp_create_worker: uber thread (%d)\n", gtid)); 796 th->th.th_info.ds.ds_thread = pthread_self(); 797 __kmp_set_stack_info(gtid, th); 798 __kmp_check_stack_overlap(th); 799 return; 800 } 801 802 KA_TRACE(10, ("__kmp_create_worker: try to create thread (%d)\n", gtid)); 803 804 KMP_MB(); /* Flush all pending memory write invalidates. */ 805 806 #ifdef KMP_THREAD_ATTR 807 status = pthread_attr_init(&thread_attr); 808 if (status != 0) { 809 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 810 } 811 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 812 if (status != 0) { 813 __kmp_fatal(KMP_MSG(CantSetWorkerState), KMP_ERR(status), __kmp_msg_null); 814 } 815 816 /* Set stack size for this thread now. 817 The multiple of 2 is there because on some machines, requesting an unusual 818 stacksize causes the thread to have an offset before the dummy alloca() 819 takes place to create the offset. Since we want the user to have a 820 sufficient stacksize AND support a stack offset, we alloca() twice the 821 offset so that the upcoming alloca() does not eliminate any premade offset, 822 and also gives the user the stack space they requested for all threads */ 823 stack_size += gtid * __kmp_stkoffset * 2; 824 825 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 826 "__kmp_stksize = %lu bytes, final stacksize = %lu bytes\n", 827 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 828 829 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 830 status = pthread_attr_setstacksize(&thread_attr, stack_size); 831 #ifdef KMP_BACKUP_STKSIZE 832 if (status != 0) { 833 if (!__kmp_env_stksize) { 834 stack_size = KMP_BACKUP_STKSIZE + gtid * __kmp_stkoffset; 835 __kmp_stksize = KMP_BACKUP_STKSIZE; 836 KA_TRACE(10, ("__kmp_create_worker: T#%d, default stacksize = %lu bytes, " 837 "__kmp_stksize = %lu bytes, (backup) final stacksize = %lu " 838 "bytes\n", 839 gtid, KMP_DEFAULT_STKSIZE, __kmp_stksize, stack_size)); 840 status = pthread_attr_setstacksize(&thread_attr, stack_size); 841 } 842 } 843 #endif /* KMP_BACKUP_STKSIZE */ 844 if (status != 0) { 845 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 846 KMP_HNT(ChangeWorkerStackSize), __kmp_msg_null); 847 } 848 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 849 850 #endif /* KMP_THREAD_ATTR */ 851 852 status = 853 pthread_create(&handle, &thread_attr, __kmp_launch_worker, (void *)th); 854 if (status != 0 || !handle) { // ??? Why do we check handle?? 855 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 856 if (status == EINVAL) { 857 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 858 KMP_HNT(IncreaseWorkerStackSize), __kmp_msg_null); 859 } 860 if (status == ENOMEM) { 861 __kmp_fatal(KMP_MSG(CantSetWorkerStackSize, stack_size), KMP_ERR(status), 862 KMP_HNT(DecreaseWorkerStackSize), __kmp_msg_null); 863 } 864 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 865 if (status == EAGAIN) { 866 __kmp_fatal(KMP_MSG(NoResourcesForWorkerThread), KMP_ERR(status), 867 KMP_HNT(Decrease_NUM_THREADS), __kmp_msg_null); 868 } 869 KMP_SYSFAIL("pthread_create", status); 870 } 871 872 th->th.th_info.ds.ds_thread = handle; 873 874 #ifdef KMP_THREAD_ATTR 875 status = pthread_attr_destroy(&thread_attr); 876 if (status) { 877 kmp_msg_t err_code = KMP_ERR(status); 878 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 879 __kmp_msg_null); 880 if (__kmp_generate_warnings == kmp_warnings_off) { 881 __kmp_str_free(&err_code.str); 882 } 883 } 884 #endif /* KMP_THREAD_ATTR */ 885 886 KMP_MB(); /* Flush all pending memory write invalidates. */ 887 888 KA_TRACE(10, ("__kmp_create_worker: done creating thread (%d)\n", gtid)); 889 890 } // __kmp_create_worker 891 892 #if KMP_USE_MONITOR 893 void __kmp_create_monitor(kmp_info_t *th) { 894 pthread_t handle; 895 pthread_attr_t thread_attr; 896 size_t size; 897 int status; 898 int auto_adj_size = FALSE; 899 900 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { 901 // We don't need monitor thread in case of MAX_BLOCKTIME 902 KA_TRACE(10, ("__kmp_create_monitor: skipping monitor thread because of " 903 "MAX blocktime\n")); 904 th->th.th_info.ds.ds_tid = 0; // this makes reap_monitor no-op 905 th->th.th_info.ds.ds_gtid = 0; 906 return; 907 } 908 KA_TRACE(10, ("__kmp_create_monitor: try to create monitor\n")); 909 910 KMP_MB(); /* Flush all pending memory write invalidates. */ 911 912 th->th.th_info.ds.ds_tid = KMP_GTID_MONITOR; 913 th->th.th_info.ds.ds_gtid = KMP_GTID_MONITOR; 914 #if KMP_REAL_TIME_FIX 915 TCW_4(__kmp_global.g.g_time.dt.t_value, 916 -1); // Will use it for synchronization a bit later. 917 #else 918 TCW_4(__kmp_global.g.g_time.dt.t_value, 0); 919 #endif // KMP_REAL_TIME_FIX 920 921 #ifdef KMP_THREAD_ATTR 922 if (__kmp_monitor_stksize == 0) { 923 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 924 auto_adj_size = TRUE; 925 } 926 status = pthread_attr_init(&thread_attr); 927 if (status != 0) { 928 __kmp_fatal(KMP_MSG(CantInitThreadAttrs), KMP_ERR(status), __kmp_msg_null); 929 } 930 status = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_JOINABLE); 931 if (status != 0) { 932 __kmp_fatal(KMP_MSG(CantSetMonitorState), KMP_ERR(status), __kmp_msg_null); 933 } 934 935 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 936 status = pthread_attr_getstacksize(&thread_attr, &size); 937 KMP_CHECK_SYSFAIL("pthread_attr_getstacksize", status); 938 #else 939 size = __kmp_sys_min_stksize; 940 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 941 #endif /* KMP_THREAD_ATTR */ 942 943 if (__kmp_monitor_stksize == 0) { 944 __kmp_monitor_stksize = KMP_DEFAULT_MONITOR_STKSIZE; 945 } 946 if (__kmp_monitor_stksize < __kmp_sys_min_stksize) { 947 __kmp_monitor_stksize = __kmp_sys_min_stksize; 948 } 949 950 KA_TRACE(10, ("__kmp_create_monitor: default stacksize = %lu bytes," 951 "requested stacksize = %lu bytes\n", 952 size, __kmp_monitor_stksize)); 953 954 retry: 955 956 /* Set stack size for this thread now. */ 957 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 958 KA_TRACE(10, ("__kmp_create_monitor: setting stacksize = %lu bytes,", 959 __kmp_monitor_stksize)); 960 status = pthread_attr_setstacksize(&thread_attr, __kmp_monitor_stksize); 961 if (status != 0) { 962 if (auto_adj_size) { 963 __kmp_monitor_stksize *= 2; 964 goto retry; 965 } 966 kmp_msg_t err_code = KMP_ERR(status); 967 __kmp_msg(kmp_ms_warning, // should this be fatal? BB 968 KMP_MSG(CantSetMonitorStackSize, (long int)__kmp_monitor_stksize), 969 err_code, KMP_HNT(ChangeMonitorStackSize), __kmp_msg_null); 970 if (__kmp_generate_warnings == kmp_warnings_off) { 971 __kmp_str_free(&err_code.str); 972 } 973 } 974 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 975 976 status = 977 pthread_create(&handle, &thread_attr, __kmp_launch_monitor, (void *)th); 978 979 if (status != 0) { 980 #ifdef _POSIX_THREAD_ATTR_STACKSIZE 981 if (status == EINVAL) { 982 if (auto_adj_size && (__kmp_monitor_stksize < (size_t)0x40000000)) { 983 __kmp_monitor_stksize *= 2; 984 goto retry; 985 } 986 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 987 KMP_ERR(status), KMP_HNT(IncreaseMonitorStackSize), 988 __kmp_msg_null); 989 } 990 if (status == ENOMEM) { 991 __kmp_fatal(KMP_MSG(CantSetMonitorStackSize, __kmp_monitor_stksize), 992 KMP_ERR(status), KMP_HNT(DecreaseMonitorStackSize), 993 __kmp_msg_null); 994 } 995 #endif /* _POSIX_THREAD_ATTR_STACKSIZE */ 996 if (status == EAGAIN) { 997 __kmp_fatal(KMP_MSG(NoResourcesForMonitorThread), KMP_ERR(status), 998 KMP_HNT(DecreaseNumberOfThreadsInUse), __kmp_msg_null); 999 } 1000 KMP_SYSFAIL("pthread_create", status); 1001 } 1002 1003 th->th.th_info.ds.ds_thread = handle; 1004 1005 #if KMP_REAL_TIME_FIX 1006 // Wait for the monitor thread is really started and set its *priority*. 1007 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 1008 sizeof(__kmp_global.g.g_time.dt.t_value)); 1009 __kmp_wait_yield_4((kmp_uint32 volatile *)&__kmp_global.g.g_time.dt.t_value, 1010 -1, &__kmp_neq_4, NULL); 1011 #endif // KMP_REAL_TIME_FIX 1012 1013 #ifdef KMP_THREAD_ATTR 1014 status = pthread_attr_destroy(&thread_attr); 1015 if (status != 0) { 1016 kmp_msg_t err_code = KMP_ERR(status); 1017 __kmp_msg(kmp_ms_warning, KMP_MSG(CantDestroyThreadAttrs), err_code, 1018 __kmp_msg_null); 1019 if (__kmp_generate_warnings == kmp_warnings_off) { 1020 __kmp_str_free(&err_code.str); 1021 } 1022 } 1023 #endif 1024 1025 KMP_MB(); /* Flush all pending memory write invalidates. */ 1026 1027 KA_TRACE(10, ("__kmp_create_monitor: monitor created %#.8lx\n", 1028 th->th.th_info.ds.ds_thread)); 1029 1030 } // __kmp_create_monitor 1031 #endif // KMP_USE_MONITOR 1032 1033 void __kmp_exit_thread(int exit_status) { 1034 pthread_exit((void *)(intptr_t)exit_status); 1035 } // __kmp_exit_thread 1036 1037 #if KMP_USE_MONITOR 1038 void __kmp_resume_monitor(); 1039 1040 void __kmp_reap_monitor(kmp_info_t *th) { 1041 int status; 1042 void *exit_val; 1043 1044 KA_TRACE(10, ("__kmp_reap_monitor: try to reap monitor thread with handle" 1045 " %#.8lx\n", 1046 th->th.th_info.ds.ds_thread)); 1047 1048 // If monitor has been created, its tid and gtid should be KMP_GTID_MONITOR. 1049 // If both tid and gtid are 0, it means the monitor did not ever start. 1050 // If both tid and gtid are KMP_GTID_DNE, the monitor has been shut down. 1051 KMP_DEBUG_ASSERT(th->th.th_info.ds.ds_tid == th->th.th_info.ds.ds_gtid); 1052 if (th->th.th_info.ds.ds_gtid != KMP_GTID_MONITOR) { 1053 KA_TRACE(10, ("__kmp_reap_monitor: monitor did not start, returning\n")); 1054 return; 1055 } 1056 1057 KMP_MB(); /* Flush all pending memory write invalidates. */ 1058 1059 /* First, check to see whether the monitor thread exists to wake it up. This 1060 is to avoid performance problem when the monitor sleeps during 1061 blocktime-size interval */ 1062 1063 status = pthread_kill(th->th.th_info.ds.ds_thread, 0); 1064 if (status != ESRCH) { 1065 __kmp_resume_monitor(); // Wake up the monitor thread 1066 } 1067 KA_TRACE(10, ("__kmp_reap_monitor: try to join with monitor\n")); 1068 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1069 if (exit_val != th) { 1070 __kmp_fatal(KMP_MSG(ReapMonitorError), KMP_ERR(status), __kmp_msg_null); 1071 } 1072 1073 th->th.th_info.ds.ds_tid = KMP_GTID_DNE; 1074 th->th.th_info.ds.ds_gtid = KMP_GTID_DNE; 1075 1076 KA_TRACE(10, ("__kmp_reap_monitor: done reaping monitor thread with handle" 1077 " %#.8lx\n", 1078 th->th.th_info.ds.ds_thread)); 1079 1080 KMP_MB(); /* Flush all pending memory write invalidates. */ 1081 } 1082 #endif // KMP_USE_MONITOR 1083 1084 void __kmp_reap_worker(kmp_info_t *th) { 1085 int status; 1086 void *exit_val; 1087 1088 KMP_MB(); /* Flush all pending memory write invalidates. */ 1089 1090 KA_TRACE( 1091 10, ("__kmp_reap_worker: try to reap T#%d\n", th->th.th_info.ds.ds_gtid)); 1092 1093 status = pthread_join(th->th.th_info.ds.ds_thread, &exit_val); 1094 #ifdef KMP_DEBUG 1095 /* Don't expose these to the user until we understand when they trigger */ 1096 if (status != 0) { 1097 __kmp_fatal(KMP_MSG(ReapWorkerError), KMP_ERR(status), __kmp_msg_null); 1098 } 1099 if (exit_val != th) { 1100 KA_TRACE(10, ("__kmp_reap_worker: worker T#%d did not reap properly, " 1101 "exit_val = %p\n", 1102 th->th.th_info.ds.ds_gtid, exit_val)); 1103 } 1104 #endif /* KMP_DEBUG */ 1105 1106 KA_TRACE(10, ("__kmp_reap_worker: done reaping T#%d\n", 1107 th->th.th_info.ds.ds_gtid)); 1108 1109 KMP_MB(); /* Flush all pending memory write invalidates. */ 1110 } 1111 1112 #if KMP_HANDLE_SIGNALS 1113 1114 static void __kmp_null_handler(int signo) { 1115 // Do nothing, for doing SIG_IGN-type actions. 1116 } // __kmp_null_handler 1117 1118 static void __kmp_team_handler(int signo) { 1119 if (__kmp_global.g.g_abort == 0) { 1120 /* Stage 1 signal handler, let's shut down all of the threads */ 1121 #ifdef KMP_DEBUG 1122 __kmp_debug_printf("__kmp_team_handler: caught signal = %d\n", signo); 1123 #endif 1124 switch (signo) { 1125 case SIGHUP: 1126 case SIGINT: 1127 case SIGQUIT: 1128 case SIGILL: 1129 case SIGABRT: 1130 case SIGFPE: 1131 case SIGBUS: 1132 case SIGSEGV: 1133 #ifdef SIGSYS 1134 case SIGSYS: 1135 #endif 1136 case SIGTERM: 1137 if (__kmp_debug_buf) { 1138 __kmp_dump_debug_buffer(); 1139 } 1140 KMP_MB(); // Flush all pending memory write invalidates. 1141 TCW_4(__kmp_global.g.g_abort, signo); 1142 KMP_MB(); // Flush all pending memory write invalidates. 1143 TCW_4(__kmp_global.g.g_done, TRUE); 1144 KMP_MB(); // Flush all pending memory write invalidates. 1145 break; 1146 default: 1147 #ifdef KMP_DEBUG 1148 __kmp_debug_printf("__kmp_team_handler: unknown signal type"); 1149 #endif 1150 break; 1151 } 1152 } 1153 } // __kmp_team_handler 1154 1155 static void __kmp_sigaction(int signum, const struct sigaction *act, 1156 struct sigaction *oldact) { 1157 int rc = sigaction(signum, act, oldact); 1158 KMP_CHECK_SYSFAIL_ERRNO("sigaction", rc); 1159 } 1160 1161 static void __kmp_install_one_handler(int sig, sig_func_t handler_func, 1162 int parallel_init) { 1163 KMP_MB(); // Flush all pending memory write invalidates. 1164 KB_TRACE(60, 1165 ("__kmp_install_one_handler( %d, ..., %d )\n", sig, parallel_init)); 1166 if (parallel_init) { 1167 struct sigaction new_action; 1168 struct sigaction old_action; 1169 new_action.sa_handler = handler_func; 1170 new_action.sa_flags = 0; 1171 sigfillset(&new_action.sa_mask); 1172 __kmp_sigaction(sig, &new_action, &old_action); 1173 if (old_action.sa_handler == __kmp_sighldrs[sig].sa_handler) { 1174 sigaddset(&__kmp_sigset, sig); 1175 } else { 1176 // Restore/keep user's handler if one previously installed. 1177 __kmp_sigaction(sig, &old_action, NULL); 1178 } 1179 } else { 1180 // Save initial/system signal handlers to see if user handlers installed. 1181 __kmp_sigaction(sig, NULL, &__kmp_sighldrs[sig]); 1182 } 1183 KMP_MB(); // Flush all pending memory write invalidates. 1184 } // __kmp_install_one_handler 1185 1186 static void __kmp_remove_one_handler(int sig) { 1187 KB_TRACE(60, ("__kmp_remove_one_handler( %d )\n", sig)); 1188 if (sigismember(&__kmp_sigset, sig)) { 1189 struct sigaction old; 1190 KMP_MB(); // Flush all pending memory write invalidates. 1191 __kmp_sigaction(sig, &__kmp_sighldrs[sig], &old); 1192 if ((old.sa_handler != __kmp_team_handler) && 1193 (old.sa_handler != __kmp_null_handler)) { 1194 // Restore the users signal handler. 1195 KB_TRACE(10, ("__kmp_remove_one_handler: oops, not our handler, " 1196 "restoring: sig=%d\n", 1197 sig)); 1198 __kmp_sigaction(sig, &old, NULL); 1199 } 1200 sigdelset(&__kmp_sigset, sig); 1201 KMP_MB(); // Flush all pending memory write invalidates. 1202 } 1203 } // __kmp_remove_one_handler 1204 1205 void __kmp_install_signals(int parallel_init) { 1206 KB_TRACE(10, ("__kmp_install_signals( %d )\n", parallel_init)); 1207 if (__kmp_handle_signals || !parallel_init) { 1208 // If ! parallel_init, we do not install handlers, just save original 1209 // handlers. Let us do it even __handle_signals is 0. 1210 sigemptyset(&__kmp_sigset); 1211 __kmp_install_one_handler(SIGHUP, __kmp_team_handler, parallel_init); 1212 __kmp_install_one_handler(SIGINT, __kmp_team_handler, parallel_init); 1213 __kmp_install_one_handler(SIGQUIT, __kmp_team_handler, parallel_init); 1214 __kmp_install_one_handler(SIGILL, __kmp_team_handler, parallel_init); 1215 __kmp_install_one_handler(SIGABRT, __kmp_team_handler, parallel_init); 1216 __kmp_install_one_handler(SIGFPE, __kmp_team_handler, parallel_init); 1217 __kmp_install_one_handler(SIGBUS, __kmp_team_handler, parallel_init); 1218 __kmp_install_one_handler(SIGSEGV, __kmp_team_handler, parallel_init); 1219 #ifdef SIGSYS 1220 __kmp_install_one_handler(SIGSYS, __kmp_team_handler, parallel_init); 1221 #endif // SIGSYS 1222 __kmp_install_one_handler(SIGTERM, __kmp_team_handler, parallel_init); 1223 #ifdef SIGPIPE 1224 __kmp_install_one_handler(SIGPIPE, __kmp_team_handler, parallel_init); 1225 #endif // SIGPIPE 1226 } 1227 } // __kmp_install_signals 1228 1229 void __kmp_remove_signals(void) { 1230 int sig; 1231 KB_TRACE(10, ("__kmp_remove_signals()\n")); 1232 for (sig = 1; sig < NSIG; ++sig) { 1233 __kmp_remove_one_handler(sig); 1234 } 1235 } // __kmp_remove_signals 1236 1237 #endif // KMP_HANDLE_SIGNALS 1238 1239 void __kmp_enable(int new_state) { 1240 #ifdef KMP_CANCEL_THREADS 1241 int status, old_state; 1242 status = pthread_setcancelstate(new_state, &old_state); 1243 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1244 KMP_DEBUG_ASSERT(old_state == PTHREAD_CANCEL_DISABLE); 1245 #endif 1246 } 1247 1248 void __kmp_disable(int *old_state) { 1249 #ifdef KMP_CANCEL_THREADS 1250 int status; 1251 status = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, old_state); 1252 KMP_CHECK_SYSFAIL("pthread_setcancelstate", status); 1253 #endif 1254 } 1255 1256 static void __kmp_atfork_prepare(void) { 1257 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 1258 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1259 } 1260 1261 static void __kmp_atfork_parent(void) { 1262 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 1263 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1264 } 1265 1266 /* Reset the library so execution in the child starts "all over again" with 1267 clean data structures in initial states. Don't worry about freeing memory 1268 allocated by parent, just abandon it to be safe. */ 1269 static void __kmp_atfork_child(void) { 1270 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1271 /* TODO make sure this is done right for nested/sibling */ 1272 // ATT: Memory leaks are here? TODO: Check it and fix. 1273 /* KMP_ASSERT( 0 ); */ 1274 1275 ++__kmp_fork_count; 1276 1277 #if KMP_AFFINITY_SUPPORTED 1278 #if KMP_OS_LINUX 1279 // reset the affinity in the child to the initial thread 1280 // affinity in the parent 1281 kmp_set_thread_affinity_mask_initial(); 1282 #endif 1283 // Set default not to bind threads tightly in the child (we’re expecting 1284 // over-subscription after the fork and this can improve things for 1285 // scripting languages that use OpenMP inside process-parallel code). 1286 __kmp_affinity_type = affinity_none; 1287 #if OMP_40_ENABLED 1288 if (__kmp_nested_proc_bind.bind_types != NULL) { 1289 __kmp_nested_proc_bind.bind_types[0] = proc_bind_false; 1290 } 1291 #endif // OMP_40_ENABLED 1292 #endif // KMP_AFFINITY_SUPPORTED 1293 1294 __kmp_init_runtime = FALSE; 1295 #if KMP_USE_MONITOR 1296 __kmp_init_monitor = 0; 1297 #endif 1298 __kmp_init_parallel = FALSE; 1299 __kmp_init_middle = FALSE; 1300 __kmp_init_serial = FALSE; 1301 TCW_4(__kmp_init_gtid, FALSE); 1302 __kmp_init_common = FALSE; 1303 1304 TCW_4(__kmp_init_user_locks, FALSE); 1305 #if !KMP_USE_DYNAMIC_LOCK 1306 __kmp_user_lock_table.used = 1; 1307 __kmp_user_lock_table.allocated = 0; 1308 __kmp_user_lock_table.table = NULL; 1309 __kmp_lock_blocks = NULL; 1310 #endif 1311 1312 __kmp_all_nth = 0; 1313 TCW_4(__kmp_nth, 0); 1314 1315 __kmp_thread_pool = NULL; 1316 __kmp_thread_pool_insert_pt = NULL; 1317 __kmp_team_pool = NULL; 1318 1319 /* Must actually zero all the *cache arguments passed to __kmpc_threadprivate 1320 here so threadprivate doesn't use stale data */ 1321 KA_TRACE(10, ("__kmp_atfork_child: checking cache address list %p\n", 1322 __kmp_threadpriv_cache_list)); 1323 1324 while (__kmp_threadpriv_cache_list != NULL) { 1325 1326 if (*__kmp_threadpriv_cache_list->addr != NULL) { 1327 KC_TRACE(50, ("__kmp_atfork_child: zeroing cache at address %p\n", 1328 &(*__kmp_threadpriv_cache_list->addr))); 1329 1330 *__kmp_threadpriv_cache_list->addr = NULL; 1331 } 1332 __kmp_threadpriv_cache_list = __kmp_threadpriv_cache_list->next; 1333 } 1334 1335 __kmp_init_runtime = FALSE; 1336 1337 /* reset statically initialized locks */ 1338 __kmp_init_bootstrap_lock(&__kmp_initz_lock); 1339 __kmp_init_bootstrap_lock(&__kmp_stdio_lock); 1340 __kmp_init_bootstrap_lock(&__kmp_console_lock); 1341 __kmp_init_bootstrap_lock(&__kmp_task_team_lock); 1342 1343 #if USE_ITT_BUILD 1344 __kmp_itt_reset(); // reset ITT's global state 1345 #endif /* USE_ITT_BUILD */ 1346 1347 /* This is necessary to make sure no stale data is left around */ 1348 /* AC: customers complain that we use unsafe routines in the atfork 1349 handler. Mathworks: dlsym() is unsafe. We call dlsym and dlopen 1350 in dynamic_link when check the presence of shared tbbmalloc library. 1351 Suggestion is to make the library initialization lazier, similar 1352 to what done for __kmpc_begin(). */ 1353 // TODO: synchronize all static initializations with regular library 1354 // startup; look at kmp_global.cpp and etc. 1355 //__kmp_internal_begin (); 1356 } 1357 1358 void __kmp_register_atfork(void) { 1359 if (__kmp_need_register_atfork) { 1360 int status = pthread_atfork(__kmp_atfork_prepare, __kmp_atfork_parent, 1361 __kmp_atfork_child); 1362 KMP_CHECK_SYSFAIL("pthread_atfork", status); 1363 __kmp_need_register_atfork = FALSE; 1364 } 1365 } 1366 1367 void __kmp_suspend_initialize(void) { 1368 int status; 1369 status = pthread_mutexattr_init(&__kmp_suspend_mutex_attr); 1370 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1371 status = pthread_condattr_init(&__kmp_suspend_cond_attr); 1372 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1373 } 1374 1375 static void __kmp_suspend_initialize_thread(kmp_info_t *th) { 1376 ANNOTATE_HAPPENS_AFTER(&th->th.th_suspend_init_count); 1377 if (th->th.th_suspend_init_count <= __kmp_fork_count) { 1378 /* this means we haven't initialized the suspension pthread objects for this 1379 thread in this instance of the process */ 1380 int status; 1381 status = pthread_cond_init(&th->th.th_suspend_cv.c_cond, 1382 &__kmp_suspend_cond_attr); 1383 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1384 status = pthread_mutex_init(&th->th.th_suspend_mx.m_mutex, 1385 &__kmp_suspend_mutex_attr); 1386 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1387 *(volatile int *)&th->th.th_suspend_init_count = __kmp_fork_count + 1; 1388 ANNOTATE_HAPPENS_BEFORE(&th->th.th_suspend_init_count); 1389 } 1390 } 1391 1392 void __kmp_suspend_uninitialize_thread(kmp_info_t *th) { 1393 if (th->th.th_suspend_init_count > __kmp_fork_count) { 1394 /* this means we have initialize the suspension pthread objects for this 1395 thread in this instance of the process */ 1396 int status; 1397 1398 status = pthread_cond_destroy(&th->th.th_suspend_cv.c_cond); 1399 if (status != 0 && status != EBUSY) { 1400 KMP_SYSFAIL("pthread_cond_destroy", status); 1401 } 1402 status = pthread_mutex_destroy(&th->th.th_suspend_mx.m_mutex); 1403 if (status != 0 && status != EBUSY) { 1404 KMP_SYSFAIL("pthread_mutex_destroy", status); 1405 } 1406 --th->th.th_suspend_init_count; 1407 KMP_DEBUG_ASSERT(th->th.th_suspend_init_count == __kmp_fork_count); 1408 } 1409 } 1410 1411 /* This routine puts the calling thread to sleep after setting the 1412 sleep bit for the indicated flag variable to true. */ 1413 template <class C> 1414 static inline void __kmp_suspend_template(int th_gtid, C *flag) { 1415 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_suspend); 1416 kmp_info_t *th = __kmp_threads[th_gtid]; 1417 int status; 1418 typename C::flag_t old_spin; 1419 1420 KF_TRACE(30, ("__kmp_suspend_template: T#%d enter for flag = %p\n", th_gtid, 1421 flag->get())); 1422 1423 __kmp_suspend_initialize_thread(th); 1424 1425 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1426 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1427 1428 KF_TRACE(10, ("__kmp_suspend_template: T#%d setting sleep bit for spin(%p)\n", 1429 th_gtid, flag->get())); 1430 1431 /* TODO: shouldn't this use release semantics to ensure that 1432 __kmp_suspend_initialize_thread gets called first? */ 1433 old_spin = flag->set_sleeping(); 1434 1435 KF_TRACE(5, ("__kmp_suspend_template: T#%d set sleep bit for spin(%p)==%x," 1436 " was %x\n", 1437 th_gtid, flag->get(), flag->load(), old_spin)); 1438 1439 if (flag->done_check_val(old_spin)) { 1440 old_spin = flag->unset_sleeping(); 1441 KF_TRACE(5, ("__kmp_suspend_template: T#%d false alarm, reset sleep bit " 1442 "for spin(%p)\n", 1443 th_gtid, flag->get())); 1444 } else { 1445 /* Encapsulate in a loop as the documentation states that this may 1446 "with low probability" return when the condition variable has 1447 not been signaled or broadcast */ 1448 int deactivated = FALSE; 1449 TCW_PTR(th->th.th_sleep_loc, (void *)flag); 1450 1451 while (flag->is_sleeping()) { 1452 #ifdef DEBUG_SUSPEND 1453 char buffer[128]; 1454 __kmp_suspend_count++; 1455 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1456 __kmp_printf("__kmp_suspend_template: suspending T#%d: %s\n", th_gtid, 1457 buffer); 1458 #endif 1459 // Mark the thread as no longer active (only in the first iteration of the 1460 // loop). 1461 if (!deactivated) { 1462 th->th.th_active = FALSE; 1463 if (th->th.th_active_in_pool) { 1464 th->th.th_active_in_pool = FALSE; 1465 KMP_ATOMIC_DEC(&__kmp_thread_pool_active_nth); 1466 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 1467 } 1468 deactivated = TRUE; 1469 } 1470 1471 #if USE_SUSPEND_TIMEOUT 1472 struct timespec now; 1473 struct timeval tval; 1474 int msecs; 1475 1476 status = gettimeofday(&tval, NULL); 1477 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1478 TIMEVAL_TO_TIMESPEC(&tval, &now); 1479 1480 msecs = (4 * __kmp_dflt_blocktime) + 200; 1481 now.tv_sec += msecs / 1000; 1482 now.tv_nsec += (msecs % 1000) * 1000; 1483 1484 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform " 1485 "pthread_cond_timedwait\n", 1486 th_gtid)); 1487 status = pthread_cond_timedwait(&th->th.th_suspend_cv.c_cond, 1488 &th->th.th_suspend_mx.m_mutex, &now); 1489 #else 1490 KF_TRACE(15, ("__kmp_suspend_template: T#%d about to perform" 1491 " pthread_cond_wait\n", 1492 th_gtid)); 1493 status = pthread_cond_wait(&th->th.th_suspend_cv.c_cond, 1494 &th->th.th_suspend_mx.m_mutex); 1495 #endif 1496 1497 if ((status != 0) && (status != EINTR) && (status != ETIMEDOUT)) { 1498 KMP_SYSFAIL("pthread_cond_wait", status); 1499 } 1500 #ifdef KMP_DEBUG 1501 if (status == ETIMEDOUT) { 1502 if (flag->is_sleeping()) { 1503 KF_TRACE(100, 1504 ("__kmp_suspend_template: T#%d timeout wakeup\n", th_gtid)); 1505 } else { 1506 KF_TRACE(2, ("__kmp_suspend_template: T#%d timeout wakeup, sleep bit " 1507 "not set!\n", 1508 th_gtid)); 1509 } 1510 } else if (flag->is_sleeping()) { 1511 KF_TRACE(100, 1512 ("__kmp_suspend_template: T#%d spurious wakeup\n", th_gtid)); 1513 } 1514 #endif 1515 } // while 1516 1517 // Mark the thread as active again (if it was previous marked as inactive) 1518 if (deactivated) { 1519 th->th.th_active = TRUE; 1520 if (TCR_4(th->th.th_in_pool)) { 1521 KMP_ATOMIC_INC(&__kmp_thread_pool_active_nth); 1522 th->th.th_active_in_pool = TRUE; 1523 } 1524 } 1525 } 1526 #ifdef DEBUG_SUSPEND 1527 { 1528 char buffer[128]; 1529 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1530 __kmp_printf("__kmp_suspend_template: T#%d has awakened: %s\n", th_gtid, 1531 buffer); 1532 } 1533 #endif 1534 1535 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1536 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1537 KF_TRACE(30, ("__kmp_suspend_template: T#%d exit\n", th_gtid)); 1538 } 1539 1540 void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag) { 1541 __kmp_suspend_template(th_gtid, flag); 1542 } 1543 void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag) { 1544 __kmp_suspend_template(th_gtid, flag); 1545 } 1546 void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag) { 1547 __kmp_suspend_template(th_gtid, flag); 1548 } 1549 1550 /* This routine signals the thread specified by target_gtid to wake up 1551 after setting the sleep bit indicated by the flag argument to FALSE. 1552 The target thread must already have called __kmp_suspend_template() */ 1553 template <class C> 1554 static inline void __kmp_resume_template(int target_gtid, C *flag) { 1555 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1556 kmp_info_t *th = __kmp_threads[target_gtid]; 1557 int status; 1558 1559 #ifdef KMP_DEBUG 1560 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1561 #endif 1562 1563 KF_TRACE(30, ("__kmp_resume_template: T#%d wants to wakeup T#%d enter\n", 1564 gtid, target_gtid)); 1565 KMP_DEBUG_ASSERT(gtid != target_gtid); 1566 1567 __kmp_suspend_initialize_thread(th); 1568 1569 status = pthread_mutex_lock(&th->th.th_suspend_mx.m_mutex); 1570 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1571 1572 if (!flag) { // coming from __kmp_null_resume_wrapper 1573 flag = (C *)CCAST(void *, th->th.th_sleep_loc); 1574 } 1575 1576 // First, check if the flag is null or its type has changed. If so, someone 1577 // else woke it up. 1578 if (!flag || flag->get_type() != flag->get_ptr_type()) { // get_ptr_type 1579 // simply shows what 1580 // flag was cast to 1581 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1582 "awake: flag(%p)\n", 1583 gtid, target_gtid, NULL)); 1584 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1585 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1586 return; 1587 } else { // if multiple threads are sleeping, flag should be internally 1588 // referring to a specific thread here 1589 typename C::flag_t old_spin = flag->unset_sleeping(); 1590 if (!flag->is_sleeping_val(old_spin)) { 1591 KF_TRACE(5, ("__kmp_resume_template: T#%d exiting, thread T#%d already " 1592 "awake: flag(%p): " 1593 "%u => %u\n", 1594 gtid, target_gtid, flag->get(), old_spin, flag->load())); 1595 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1596 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1597 return; 1598 } 1599 KF_TRACE(5, ("__kmp_resume_template: T#%d about to wakeup T#%d, reset " 1600 "sleep bit for flag's loc(%p): " 1601 "%u => %u\n", 1602 gtid, target_gtid, flag->get(), old_spin, flag->load())); 1603 } 1604 TCW_PTR(th->th.th_sleep_loc, NULL); 1605 1606 #ifdef DEBUG_SUSPEND 1607 { 1608 char buffer[128]; 1609 __kmp_print_cond(buffer, &th->th.th_suspend_cv); 1610 __kmp_printf("__kmp_resume_template: T#%d resuming T#%d: %s\n", gtid, 1611 target_gtid, buffer); 1612 } 1613 #endif 1614 status = pthread_cond_signal(&th->th.th_suspend_cv.c_cond); 1615 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1616 status = pthread_mutex_unlock(&th->th.th_suspend_mx.m_mutex); 1617 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1618 KF_TRACE(30, ("__kmp_resume_template: T#%d exiting after signaling wake up" 1619 " for T#%d\n", 1620 gtid, target_gtid)); 1621 } 1622 1623 void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag) { 1624 __kmp_resume_template(target_gtid, flag); 1625 } 1626 void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag) { 1627 __kmp_resume_template(target_gtid, flag); 1628 } 1629 void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag) { 1630 __kmp_resume_template(target_gtid, flag); 1631 } 1632 1633 #if KMP_USE_MONITOR 1634 void __kmp_resume_monitor() { 1635 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(USER_resume); 1636 int status; 1637 #ifdef KMP_DEBUG 1638 int gtid = TCR_4(__kmp_init_gtid) ? __kmp_get_gtid() : -1; 1639 KF_TRACE(30, ("__kmp_resume_monitor: T#%d wants to wakeup T#%d enter\n", gtid, 1640 KMP_GTID_MONITOR)); 1641 KMP_DEBUG_ASSERT(gtid != KMP_GTID_MONITOR); 1642 #endif 1643 status = pthread_mutex_lock(&__kmp_wait_mx.m_mutex); 1644 KMP_CHECK_SYSFAIL("pthread_mutex_lock", status); 1645 #ifdef DEBUG_SUSPEND 1646 { 1647 char buffer[128]; 1648 __kmp_print_cond(buffer, &__kmp_wait_cv.c_cond); 1649 __kmp_printf("__kmp_resume_monitor: T#%d resuming T#%d: %s\n", gtid, 1650 KMP_GTID_MONITOR, buffer); 1651 } 1652 #endif 1653 status = pthread_cond_signal(&__kmp_wait_cv.c_cond); 1654 KMP_CHECK_SYSFAIL("pthread_cond_signal", status); 1655 status = pthread_mutex_unlock(&__kmp_wait_mx.m_mutex); 1656 KMP_CHECK_SYSFAIL("pthread_mutex_unlock", status); 1657 KF_TRACE(30, ("__kmp_resume_monitor: T#%d exiting after signaling wake up" 1658 " for T#%d\n", 1659 gtid, KMP_GTID_MONITOR)); 1660 } 1661 #endif // KMP_USE_MONITOR 1662 1663 void __kmp_yield(int cond) { 1664 if (!cond) 1665 return; 1666 #if KMP_USE_MONITOR 1667 if (!__kmp_yielding_on) 1668 return; 1669 #else 1670 if (__kmp_yield_cycle && !KMP_YIELD_NOW()) 1671 return; 1672 #endif 1673 sched_yield(); 1674 } 1675 1676 void __kmp_gtid_set_specific(int gtid) { 1677 if (__kmp_init_gtid) { 1678 int status; 1679 status = pthread_setspecific(__kmp_gtid_threadprivate_key, 1680 (void *)(intptr_t)(gtid + 1)); 1681 KMP_CHECK_SYSFAIL("pthread_setspecific", status); 1682 } else { 1683 KA_TRACE(50, ("__kmp_gtid_set_specific: runtime shutdown, returning\n")); 1684 } 1685 } 1686 1687 int __kmp_gtid_get_specific() { 1688 int gtid; 1689 if (!__kmp_init_gtid) { 1690 KA_TRACE(50, ("__kmp_gtid_get_specific: runtime shutdown, returning " 1691 "KMP_GTID_SHUTDOWN\n")); 1692 return KMP_GTID_SHUTDOWN; 1693 } 1694 gtid = (int)(size_t)pthread_getspecific(__kmp_gtid_threadprivate_key); 1695 if (gtid == 0) { 1696 gtid = KMP_GTID_DNE; 1697 } else { 1698 gtid--; 1699 } 1700 KA_TRACE(50, ("__kmp_gtid_get_specific: key:%d gtid:%d\n", 1701 __kmp_gtid_threadprivate_key, gtid)); 1702 return gtid; 1703 } 1704 1705 double __kmp_read_cpu_time(void) { 1706 /*clock_t t;*/ 1707 struct tms buffer; 1708 1709 /*t =*/times(&buffer); 1710 1711 return (buffer.tms_utime + buffer.tms_cutime) / (double)CLOCKS_PER_SEC; 1712 } 1713 1714 int __kmp_read_system_info(struct kmp_sys_info *info) { 1715 int status; 1716 struct rusage r_usage; 1717 1718 memset(info, 0, sizeof(*info)); 1719 1720 status = getrusage(RUSAGE_SELF, &r_usage); 1721 KMP_CHECK_SYSFAIL_ERRNO("getrusage", status); 1722 1723 // The maximum resident set size utilized (in kilobytes) 1724 info->maxrss = r_usage.ru_maxrss; 1725 // The number of page faults serviced without any I/O 1726 info->minflt = r_usage.ru_minflt; 1727 // The number of page faults serviced that required I/O 1728 info->majflt = r_usage.ru_majflt; 1729 // The number of times a process was "swapped" out of memory 1730 info->nswap = r_usage.ru_nswap; 1731 // The number of times the file system had to perform input 1732 info->inblock = r_usage.ru_inblock; 1733 // The number of times the file system had to perform output 1734 info->oublock = r_usage.ru_oublock; 1735 // The number of times a context switch was voluntarily 1736 info->nvcsw = r_usage.ru_nvcsw; 1737 // The number of times a context switch was forced 1738 info->nivcsw = r_usage.ru_nivcsw; 1739 1740 return (status != 0); 1741 } 1742 1743 void __kmp_read_system_time(double *delta) { 1744 double t_ns; 1745 struct timeval tval; 1746 struct timespec stop; 1747 int status; 1748 1749 status = gettimeofday(&tval, NULL); 1750 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1751 TIMEVAL_TO_TIMESPEC(&tval, &stop); 1752 t_ns = TS2NS(stop) - TS2NS(__kmp_sys_timer_data.start); 1753 *delta = (t_ns * 1e-9); 1754 } 1755 1756 void __kmp_clear_system_time(void) { 1757 struct timeval tval; 1758 int status; 1759 status = gettimeofday(&tval, NULL); 1760 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1761 TIMEVAL_TO_TIMESPEC(&tval, &__kmp_sys_timer_data.start); 1762 } 1763 1764 static int __kmp_get_xproc(void) { 1765 1766 int r = 0; 1767 1768 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD 1769 1770 r = sysconf(_SC_NPROCESSORS_ONLN); 1771 1772 #elif KMP_OS_DARWIN 1773 1774 // Bug C77011 High "OpenMP Threads and number of active cores". 1775 1776 // Find the number of available CPUs. 1777 kern_return_t rc; 1778 host_basic_info_data_t info; 1779 mach_msg_type_number_t num = HOST_BASIC_INFO_COUNT; 1780 rc = host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&info, &num); 1781 if (rc == 0 && num == HOST_BASIC_INFO_COUNT) { 1782 // Cannot use KA_TRACE() here because this code works before trace support 1783 // is initialized. 1784 r = info.avail_cpus; 1785 } else { 1786 KMP_WARNING(CantGetNumAvailCPU); 1787 KMP_INFORM(AssumedNumCPU); 1788 } 1789 1790 #else 1791 1792 #error "Unknown or unsupported OS." 1793 1794 #endif 1795 1796 return r > 0 ? r : 2; /* guess value of 2 if OS told us 0 */ 1797 1798 } // __kmp_get_xproc 1799 1800 int __kmp_read_from_file(char const *path, char const *format, ...) { 1801 int result; 1802 va_list args; 1803 1804 va_start(args, format); 1805 FILE *f = fopen(path, "rb"); 1806 if (f == NULL) 1807 return 0; 1808 result = vfscanf(f, format, args); 1809 fclose(f); 1810 1811 return result; 1812 } 1813 1814 void __kmp_runtime_initialize(void) { 1815 int status; 1816 pthread_mutexattr_t mutex_attr; 1817 pthread_condattr_t cond_attr; 1818 1819 if (__kmp_init_runtime) { 1820 return; 1821 } 1822 1823 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1824 if (!__kmp_cpuinfo.initialized) { 1825 __kmp_query_cpuid(&__kmp_cpuinfo); 1826 } 1827 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1828 1829 __kmp_xproc = __kmp_get_xproc(); 1830 1831 if (sysconf(_SC_THREADS)) { 1832 1833 /* Query the maximum number of threads */ 1834 __kmp_sys_max_nth = sysconf(_SC_THREAD_THREADS_MAX); 1835 if (__kmp_sys_max_nth == -1) { 1836 /* Unlimited threads for NPTL */ 1837 __kmp_sys_max_nth = INT_MAX; 1838 } else if (__kmp_sys_max_nth <= 1) { 1839 /* Can't tell, just use PTHREAD_THREADS_MAX */ 1840 __kmp_sys_max_nth = KMP_MAX_NTH; 1841 } 1842 1843 /* Query the minimum stack size */ 1844 __kmp_sys_min_stksize = sysconf(_SC_THREAD_STACK_MIN); 1845 if (__kmp_sys_min_stksize <= 1) { 1846 __kmp_sys_min_stksize = KMP_MIN_STKSIZE; 1847 } 1848 } 1849 1850 /* Set up minimum number of threads to switch to TLS gtid */ 1851 __kmp_tls_gtid_min = KMP_TLS_GTID_MIN; 1852 1853 status = pthread_key_create(&__kmp_gtid_threadprivate_key, 1854 __kmp_internal_end_dest); 1855 KMP_CHECK_SYSFAIL("pthread_key_create", status); 1856 status = pthread_mutexattr_init(&mutex_attr); 1857 KMP_CHECK_SYSFAIL("pthread_mutexattr_init", status); 1858 status = pthread_mutex_init(&__kmp_wait_mx.m_mutex, &mutex_attr); 1859 KMP_CHECK_SYSFAIL("pthread_mutex_init", status); 1860 status = pthread_condattr_init(&cond_attr); 1861 KMP_CHECK_SYSFAIL("pthread_condattr_init", status); 1862 status = pthread_cond_init(&__kmp_wait_cv.c_cond, &cond_attr); 1863 KMP_CHECK_SYSFAIL("pthread_cond_init", status); 1864 #if USE_ITT_BUILD 1865 __kmp_itt_initialize(); 1866 #endif /* USE_ITT_BUILD */ 1867 1868 __kmp_init_runtime = TRUE; 1869 } 1870 1871 void __kmp_runtime_destroy(void) { 1872 int status; 1873 1874 if (!__kmp_init_runtime) { 1875 return; // Nothing to do. 1876 } 1877 1878 #if USE_ITT_BUILD 1879 __kmp_itt_destroy(); 1880 #endif /* USE_ITT_BUILD */ 1881 1882 status = pthread_key_delete(__kmp_gtid_threadprivate_key); 1883 KMP_CHECK_SYSFAIL("pthread_key_delete", status); 1884 1885 status = pthread_mutex_destroy(&__kmp_wait_mx.m_mutex); 1886 if (status != 0 && status != EBUSY) { 1887 KMP_SYSFAIL("pthread_mutex_destroy", status); 1888 } 1889 status = pthread_cond_destroy(&__kmp_wait_cv.c_cond); 1890 if (status != 0 && status != EBUSY) { 1891 KMP_SYSFAIL("pthread_cond_destroy", status); 1892 } 1893 #if KMP_AFFINITY_SUPPORTED 1894 __kmp_affinity_uninitialize(); 1895 #endif 1896 1897 __kmp_init_runtime = FALSE; 1898 } 1899 1900 /* Put the thread to sleep for a time period */ 1901 /* NOTE: not currently used anywhere */ 1902 void __kmp_thread_sleep(int millis) { sleep((millis + 500) / 1000); } 1903 1904 /* Calculate the elapsed wall clock time for the user */ 1905 void __kmp_elapsed(double *t) { 1906 int status; 1907 #ifdef FIX_SGI_CLOCK 1908 struct timespec ts; 1909 1910 status = clock_gettime(CLOCK_PROCESS_CPUTIME_ID, &ts); 1911 KMP_CHECK_SYSFAIL_ERRNO("clock_gettime", status); 1912 *t = 1913 (double)ts.tv_nsec * (1.0 / (double)KMP_NSEC_PER_SEC) + (double)ts.tv_sec; 1914 #else 1915 struct timeval tv; 1916 1917 status = gettimeofday(&tv, NULL); 1918 KMP_CHECK_SYSFAIL_ERRNO("gettimeofday", status); 1919 *t = 1920 (double)tv.tv_usec * (1.0 / (double)KMP_USEC_PER_SEC) + (double)tv.tv_sec; 1921 #endif 1922 } 1923 1924 /* Calculate the elapsed wall clock tick for the user */ 1925 void __kmp_elapsed_tick(double *t) { *t = 1 / (double)CLOCKS_PER_SEC; } 1926 1927 /* Return the current time stamp in nsec */ 1928 kmp_uint64 __kmp_now_nsec() { 1929 struct timeval t; 1930 gettimeofday(&t, NULL); 1931 return KMP_NSEC_PER_SEC * t.tv_sec + 1000 * t.tv_usec; 1932 } 1933 1934 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1935 /* Measure clock ticks per millisecond */ 1936 void __kmp_initialize_system_tick() { 1937 kmp_uint64 delay = 100000; // 50~100 usec on most machines. 1938 kmp_uint64 nsec = __kmp_now_nsec(); 1939 kmp_uint64 goal = __kmp_hardware_timestamp() + delay; 1940 kmp_uint64 now; 1941 while ((now = __kmp_hardware_timestamp()) < goal) 1942 ; 1943 __kmp_ticks_per_msec = 1944 (kmp_uint64)(1e6 * (delay + (now - goal)) / (__kmp_now_nsec() - nsec)); 1945 } 1946 #endif 1947 1948 /* Determine whether the given address is mapped into the current address 1949 space. */ 1950 1951 int __kmp_is_address_mapped(void *addr) { 1952 1953 int found = 0; 1954 int rc; 1955 1956 #if KMP_OS_LINUX || KMP_OS_FREEBSD 1957 1958 /* On Linux* OS, read the /proc/<pid>/maps pseudo-file to get all the address 1959 ranges mapped into the address space. */ 1960 1961 char *name = __kmp_str_format("/proc/%d/maps", getpid()); 1962 FILE *file = NULL; 1963 1964 file = fopen(name, "r"); 1965 KMP_ASSERT(file != NULL); 1966 1967 for (;;) { 1968 1969 void *beginning = NULL; 1970 void *ending = NULL; 1971 char perms[5]; 1972 1973 rc = fscanf(file, "%p-%p %4s %*[^\n]\n", &beginning, &ending, perms); 1974 if (rc == EOF) { 1975 break; 1976 } 1977 KMP_ASSERT(rc == 3 && 1978 KMP_STRLEN(perms) == 4); // Make sure all fields are read. 1979 1980 // Ending address is not included in the region, but beginning is. 1981 if ((addr >= beginning) && (addr < ending)) { 1982 perms[2] = 0; // 3th and 4th character does not matter. 1983 if (strcmp(perms, "rw") == 0) { 1984 // Memory we are looking for should be readable and writable. 1985 found = 1; 1986 } 1987 break; 1988 } 1989 } 1990 1991 // Free resources. 1992 fclose(file); 1993 KMP_INTERNAL_FREE(name); 1994 1995 #elif KMP_OS_DARWIN 1996 1997 /* On OS X*, /proc pseudo filesystem is not available. Try to read memory 1998 using vm interface. */ 1999 2000 int buffer; 2001 vm_size_t count; 2002 rc = vm_read_overwrite( 2003 mach_task_self(), // Task to read memory of. 2004 (vm_address_t)(addr), // Address to read from. 2005 1, // Number of bytes to be read. 2006 (vm_address_t)(&buffer), // Address of buffer to save read bytes in. 2007 &count // Address of var to save number of read bytes in. 2008 ); 2009 if (rc == 0) { 2010 // Memory successfully read. 2011 found = 1; 2012 } 2013 2014 #elif KMP_OS_FREEBSD || KMP_OS_NETBSD 2015 2016 // FIXME(FreeBSD, NetBSD): Implement this 2017 found = 1; 2018 2019 #else 2020 2021 #error "Unknown or unsupported OS" 2022 2023 #endif 2024 2025 return found; 2026 2027 } // __kmp_is_address_mapped 2028 2029 #ifdef USE_LOAD_BALANCE 2030 2031 #if KMP_OS_DARWIN 2032 2033 // The function returns the rounded value of the system load average 2034 // during given time interval which depends on the value of 2035 // __kmp_load_balance_interval variable (default is 60 sec, other values 2036 // may be 300 sec or 900 sec). 2037 // It returns -1 in case of error. 2038 int __kmp_get_load_balance(int max) { 2039 double averages[3]; 2040 int ret_avg = 0; 2041 2042 int res = getloadavg(averages, 3); 2043 2044 // Check __kmp_load_balance_interval to determine which of averages to use. 2045 // getloadavg() may return the number of samples less than requested that is 2046 // less than 3. 2047 if (__kmp_load_balance_interval < 180 && (res >= 1)) { 2048 ret_avg = averages[0]; // 1 min 2049 } else if ((__kmp_load_balance_interval >= 180 && 2050 __kmp_load_balance_interval < 600) && 2051 (res >= 2)) { 2052 ret_avg = averages[1]; // 5 min 2053 } else if ((__kmp_load_balance_interval >= 600) && (res == 3)) { 2054 ret_avg = averages[2]; // 15 min 2055 } else { // Error occurred 2056 return -1; 2057 } 2058 2059 return ret_avg; 2060 } 2061 2062 #else // Linux* OS 2063 2064 // The fuction returns number of running (not sleeping) threads, or -1 in case 2065 // of error. Error could be reported if Linux* OS kernel too old (without 2066 // "/proc" support). Counting running threads stops if max running threads 2067 // encountered. 2068 int __kmp_get_load_balance(int max) { 2069 static int permanent_error = 0; 2070 static int glb_running_threads = 0; // Saved count of the running threads for 2071 // the thread balance algortihm 2072 static double glb_call_time = 0; /* Thread balance algorithm call time */ 2073 2074 int running_threads = 0; // Number of running threads in the system. 2075 2076 DIR *proc_dir = NULL; // Handle of "/proc/" directory. 2077 struct dirent *proc_entry = NULL; 2078 2079 kmp_str_buf_t task_path; // "/proc/<pid>/task/<tid>/" path. 2080 DIR *task_dir = NULL; // Handle of "/proc/<pid>/task/<tid>/" directory. 2081 struct dirent *task_entry = NULL; 2082 int task_path_fixed_len; 2083 2084 kmp_str_buf_t stat_path; // "/proc/<pid>/task/<tid>/stat" path. 2085 int stat_file = -1; 2086 int stat_path_fixed_len; 2087 2088 int total_processes = 0; // Total number of processes in system. 2089 int total_threads = 0; // Total number of threads in system. 2090 2091 double call_time = 0.0; 2092 2093 __kmp_str_buf_init(&task_path); 2094 __kmp_str_buf_init(&stat_path); 2095 2096 __kmp_elapsed(&call_time); 2097 2098 if (glb_call_time && 2099 (call_time - glb_call_time < __kmp_load_balance_interval)) { 2100 running_threads = glb_running_threads; 2101 goto finish; 2102 } 2103 2104 glb_call_time = call_time; 2105 2106 // Do not spend time on scanning "/proc/" if we have a permanent error. 2107 if (permanent_error) { 2108 running_threads = -1; 2109 goto finish; 2110 } 2111 2112 if (max <= 0) { 2113 max = INT_MAX; 2114 } 2115 2116 // Open "/proc/" directory. 2117 proc_dir = opendir("/proc"); 2118 if (proc_dir == NULL) { 2119 // Cannot open "/prroc/". Probably the kernel does not support it. Return an 2120 // error now and in subsequent calls. 2121 running_threads = -1; 2122 permanent_error = 1; 2123 goto finish; 2124 } 2125 2126 // Initialize fixed part of task_path. This part will not change. 2127 __kmp_str_buf_cat(&task_path, "/proc/", 6); 2128 task_path_fixed_len = task_path.used; // Remember number of used characters. 2129 2130 proc_entry = readdir(proc_dir); 2131 while (proc_entry != NULL) { 2132 // Proc entry is a directory and name starts with a digit. Assume it is a 2133 // process' directory. 2134 if (proc_entry->d_type == DT_DIR && isdigit(proc_entry->d_name[0])) { 2135 2136 ++total_processes; 2137 // Make sure init process is the very first in "/proc", so we can replace 2138 // strcmp( proc_entry->d_name, "1" ) == 0 with simpler total_processes == 2139 // 1. We are going to check that total_processes == 1 => d_name == "1" is 2140 // true (where "=>" is implication). Since C++ does not have => operator, 2141 // let us replace it with its equivalent: a => b == ! a || b. 2142 KMP_DEBUG_ASSERT(total_processes != 1 || 2143 strcmp(proc_entry->d_name, "1") == 0); 2144 2145 // Construct task_path. 2146 task_path.used = task_path_fixed_len; // Reset task_path to "/proc/". 2147 __kmp_str_buf_cat(&task_path, proc_entry->d_name, 2148 KMP_STRLEN(proc_entry->d_name)); 2149 __kmp_str_buf_cat(&task_path, "/task", 5); 2150 2151 task_dir = opendir(task_path.str); 2152 if (task_dir == NULL) { 2153 // Process can finish between reading "/proc/" directory entry and 2154 // opening process' "task/" directory. So, in general case we should not 2155 // complain, but have to skip this process and read the next one. But on 2156 // systems with no "task/" support we will spend lot of time to scan 2157 // "/proc/" tree again and again without any benefit. "init" process 2158 // (its pid is 1) should exist always, so, if we cannot open 2159 // "/proc/1/task/" directory, it means "task/" is not supported by 2160 // kernel. Report an error now and in the future. 2161 if (strcmp(proc_entry->d_name, "1") == 0) { 2162 running_threads = -1; 2163 permanent_error = 1; 2164 goto finish; 2165 } 2166 } else { 2167 // Construct fixed part of stat file path. 2168 __kmp_str_buf_clear(&stat_path); 2169 __kmp_str_buf_cat(&stat_path, task_path.str, task_path.used); 2170 __kmp_str_buf_cat(&stat_path, "/", 1); 2171 stat_path_fixed_len = stat_path.used; 2172 2173 task_entry = readdir(task_dir); 2174 while (task_entry != NULL) { 2175 // It is a directory and name starts with a digit. 2176 if (proc_entry->d_type == DT_DIR && isdigit(task_entry->d_name[0])) { 2177 ++total_threads; 2178 2179 // Consruct complete stat file path. Easiest way would be: 2180 // __kmp_str_buf_print( & stat_path, "%s/%s/stat", task_path.str, 2181 // task_entry->d_name ); 2182 // but seriae of __kmp_str_buf_cat works a bit faster. 2183 stat_path.used = 2184 stat_path_fixed_len; // Reset stat path to its fixed part. 2185 __kmp_str_buf_cat(&stat_path, task_entry->d_name, 2186 KMP_STRLEN(task_entry->d_name)); 2187 __kmp_str_buf_cat(&stat_path, "/stat", 5); 2188 2189 // Note: Low-level API (open/read/close) is used. High-level API 2190 // (fopen/fclose) works ~ 30 % slower. 2191 stat_file = open(stat_path.str, O_RDONLY); 2192 if (stat_file == -1) { 2193 // We cannot report an error because task (thread) can terminate 2194 // just before reading this file. 2195 } else { 2196 /* Content of "stat" file looks like: 2197 24285 (program) S ... 2198 2199 It is a single line (if program name does not include funny 2200 symbols). First number is a thread id, then name of executable 2201 file name in paretheses, then state of the thread. We need just 2202 thread state. 2203 2204 Good news: Length of program name is 15 characters max. Longer 2205 names are truncated. 2206 2207 Thus, we need rather short buffer: 15 chars for program name + 2208 2 parenthesis, + 3 spaces + ~7 digits of pid = 37. 2209 2210 Bad news: Program name may contain special symbols like space, 2211 closing parenthesis, or even new line. This makes parsing 2212 "stat" file not 100 % reliable. In case of fanny program names 2213 parsing may fail (report incorrect thread state). 2214 2215 Parsing "status" file looks more promissing (due to different 2216 file structure and escaping special symbols) but reading and 2217 parsing of "status" file works slower. 2218 -- ln 2219 */ 2220 char buffer[65]; 2221 int len; 2222 len = read(stat_file, buffer, sizeof(buffer) - 1); 2223 if (len >= 0) { 2224 buffer[len] = 0; 2225 // Using scanf: 2226 // sscanf( buffer, "%*d (%*s) %c ", & state ); 2227 // looks very nice, but searching for a closing parenthesis 2228 // works a bit faster. 2229 char *close_parent = strstr(buffer, ") "); 2230 if (close_parent != NULL) { 2231 char state = *(close_parent + 2); 2232 if (state == 'R') { 2233 ++running_threads; 2234 if (running_threads >= max) { 2235 goto finish; 2236 } 2237 } 2238 } 2239 } 2240 close(stat_file); 2241 stat_file = -1; 2242 } 2243 } 2244 task_entry = readdir(task_dir); 2245 } 2246 closedir(task_dir); 2247 task_dir = NULL; 2248 } 2249 } 2250 proc_entry = readdir(proc_dir); 2251 } 2252 2253 // There _might_ be a timing hole where the thread executing this 2254 // code get skipped in the load balance, and running_threads is 0. 2255 // Assert in the debug builds only!!! 2256 KMP_DEBUG_ASSERT(running_threads > 0); 2257 if (running_threads <= 0) { 2258 running_threads = 1; 2259 } 2260 2261 finish: // Clean up and exit. 2262 if (proc_dir != NULL) { 2263 closedir(proc_dir); 2264 } 2265 __kmp_str_buf_free(&task_path); 2266 if (task_dir != NULL) { 2267 closedir(task_dir); 2268 } 2269 __kmp_str_buf_free(&stat_path); 2270 if (stat_file != -1) { 2271 close(stat_file); 2272 } 2273 2274 glb_running_threads = running_threads; 2275 2276 return running_threads; 2277 2278 } // __kmp_get_load_balance 2279 2280 #endif // KMP_OS_DARWIN 2281 2282 #endif // USE_LOAD_BALANCE 2283 2284 #if !(KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_MIC || \ 2285 ((KMP_OS_LINUX || KMP_OS_DARWIN) && KMP_ARCH_AARCH64) || KMP_ARCH_PPC64) 2286 2287 // we really only need the case with 1 argument, because CLANG always build 2288 // a struct of pointers to shared variables referenced in the outlined function 2289 int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int tid, int argc, 2290 void *p_argv[] 2291 #if OMPT_SUPPORT 2292 , 2293 void **exit_frame_ptr 2294 #endif 2295 ) { 2296 #if OMPT_SUPPORT 2297 *exit_frame_ptr = OMPT_GET_FRAME_ADDRESS(0); 2298 #endif 2299 2300 switch (argc) { 2301 default: 2302 fprintf(stderr, "Too many args to microtask: %d!\n", argc); 2303 fflush(stderr); 2304 exit(-1); 2305 case 0: 2306 (*pkfn)(>id, &tid); 2307 break; 2308 case 1: 2309 (*pkfn)(>id, &tid, p_argv[0]); 2310 break; 2311 case 2: 2312 (*pkfn)(>id, &tid, p_argv[0], p_argv[1]); 2313 break; 2314 case 3: 2315 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2]); 2316 break; 2317 case 4: 2318 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3]); 2319 break; 2320 case 5: 2321 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4]); 2322 break; 2323 case 6: 2324 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2325 p_argv[5]); 2326 break; 2327 case 7: 2328 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2329 p_argv[5], p_argv[6]); 2330 break; 2331 case 8: 2332 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2333 p_argv[5], p_argv[6], p_argv[7]); 2334 break; 2335 case 9: 2336 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2337 p_argv[5], p_argv[6], p_argv[7], p_argv[8]); 2338 break; 2339 case 10: 2340 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2341 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9]); 2342 break; 2343 case 11: 2344 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2345 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10]); 2346 break; 2347 case 12: 2348 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2349 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2350 p_argv[11]); 2351 break; 2352 case 13: 2353 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2354 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2355 p_argv[11], p_argv[12]); 2356 break; 2357 case 14: 2358 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2359 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2360 p_argv[11], p_argv[12], p_argv[13]); 2361 break; 2362 case 15: 2363 (*pkfn)(>id, &tid, p_argv[0], p_argv[1], p_argv[2], p_argv[3], p_argv[4], 2364 p_argv[5], p_argv[6], p_argv[7], p_argv[8], p_argv[9], p_argv[10], 2365 p_argv[11], p_argv[12], p_argv[13], p_argv[14]); 2366 break; 2367 } 2368 2369 #if OMPT_SUPPORT 2370 *exit_frame_ptr = 0; 2371 #endif 2372 2373 return 1; 2374 } 2375 2376 #endif 2377 2378 // end of file // 2379