1 /* 2 * kmp_lock.cpp -- lock-related functions 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 <stddef.h> 14 #include <atomic> 15 16 #include "kmp.h" 17 #include "kmp_i18n.h" 18 #include "kmp_io.h" 19 #include "kmp_itt.h" 20 #include "kmp_lock.h" 21 #include "kmp_wait_release.h" 22 #include "kmp_wrapper_getpid.h" 23 24 #if KMP_USE_FUTEX 25 #include <sys/syscall.h> 26 #include <unistd.h> 27 // We should really include <futex.h>, but that causes compatibility problems on 28 // different Linux* OS distributions that either require that you include (or 29 // break when you try to include) <pci/types.h>. Since all we need is the two 30 // macros below (which are part of the kernel ABI, so can't change) we just 31 // define the constants here and don't include <futex.h> 32 #ifndef FUTEX_WAIT 33 #define FUTEX_WAIT 0 34 #endif 35 #ifndef FUTEX_WAKE 36 #define FUTEX_WAKE 1 37 #endif 38 #endif 39 40 /* Implement spin locks for internal library use. */ 41 /* The algorithm implemented is Lamport's bakery lock [1974]. */ 42 43 void __kmp_validate_locks(void) { 44 int i; 45 kmp_uint32 x, y; 46 47 /* Check to make sure unsigned arithmetic does wraps properly */ 48 x = ~((kmp_uint32)0) - 2; 49 y = x - 2; 50 51 for (i = 0; i < 8; ++i, ++x, ++y) { 52 kmp_uint32 z = (x - y); 53 KMP_ASSERT(z == 2); 54 } 55 56 KMP_ASSERT(offsetof(kmp_base_queuing_lock, tail_id) % 8 == 0); 57 } 58 59 /* ------------------------------------------------------------------------ */ 60 /* test and set locks */ 61 62 // For the non-nested locks, we can only assume that the first 4 bytes were 63 // allocated, since gcc only allocates 4 bytes for omp_lock_t, and the Intel 64 // compiler only allocates a 4 byte pointer on IA-32 architecture. On 65 // Windows* OS on Intel(R) 64, we can assume that all 8 bytes were allocated. 66 // 67 // gcc reserves >= 8 bytes for nested locks, so we can assume that the 68 // entire 8 bytes were allocated for nested locks on all 64-bit platforms. 69 70 static kmp_int32 __kmp_get_tas_lock_owner(kmp_tas_lock_t *lck) { 71 return KMP_LOCK_STRIP(KMP_ATOMIC_LD_RLX(&lck->lk.poll)) - 1; 72 } 73 74 static inline bool __kmp_is_tas_lock_nestable(kmp_tas_lock_t *lck) { 75 return lck->lk.depth_locked != -1; 76 } 77 78 __forceinline static int 79 __kmp_acquire_tas_lock_timed_template(kmp_tas_lock_t *lck, kmp_int32 gtid) { 80 KMP_MB(); 81 82 #ifdef USE_LOCK_PROFILE 83 kmp_uint32 curr = KMP_LOCK_STRIP(lck->lk.poll); 84 if ((curr != 0) && (curr != gtid + 1)) 85 __kmp_printf("LOCK CONTENTION: %p\n", lck); 86 /* else __kmp_printf( "." );*/ 87 #endif /* USE_LOCK_PROFILE */ 88 89 kmp_int32 tas_free = KMP_LOCK_FREE(tas); 90 kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas); 91 92 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free && 93 __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) { 94 KMP_FSYNC_ACQUIRED(lck); 95 return KMP_LOCK_ACQUIRED_FIRST; 96 } 97 98 kmp_uint32 spins; 99 KMP_FSYNC_PREPARE(lck); 100 KMP_INIT_YIELD(spins); 101 kmp_backoff_t backoff = __kmp_spin_backoff_params; 102 do { 103 __kmp_spin_backoff(&backoff); 104 KMP_YIELD_OVERSUB_ELSE_SPIN(spins); 105 } while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free || 106 !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)); 107 KMP_FSYNC_ACQUIRED(lck); 108 return KMP_LOCK_ACQUIRED_FIRST; 109 } 110 111 int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) { 112 int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid); 113 return retval; 114 } 115 116 static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck, 117 kmp_int32 gtid) { 118 char const *const func = "omp_set_lock"; 119 if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) && 120 __kmp_is_tas_lock_nestable(lck)) { 121 KMP_FATAL(LockNestableUsedAsSimple, func); 122 } 123 if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) { 124 KMP_FATAL(LockIsAlreadyOwned, func); 125 } 126 return __kmp_acquire_tas_lock(lck, gtid); 127 } 128 129 int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) { 130 kmp_int32 tas_free = KMP_LOCK_FREE(tas); 131 kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas); 132 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free && 133 __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) { 134 KMP_FSYNC_ACQUIRED(lck); 135 return TRUE; 136 } 137 return FALSE; 138 } 139 140 static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck, 141 kmp_int32 gtid) { 142 char const *const func = "omp_test_lock"; 143 if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) && 144 __kmp_is_tas_lock_nestable(lck)) { 145 KMP_FATAL(LockNestableUsedAsSimple, func); 146 } 147 return __kmp_test_tas_lock(lck, gtid); 148 } 149 150 int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) { 151 KMP_MB(); /* Flush all pending memory write invalidates. */ 152 153 KMP_FSYNC_RELEASING(lck); 154 KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas)); 155 KMP_MB(); /* Flush all pending memory write invalidates. */ 156 157 KMP_YIELD_OVERSUB(); 158 return KMP_LOCK_RELEASED; 159 } 160 161 static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck, 162 kmp_int32 gtid) { 163 char const *const func = "omp_unset_lock"; 164 KMP_MB(); /* in case another processor initialized lock */ 165 if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) && 166 __kmp_is_tas_lock_nestable(lck)) { 167 KMP_FATAL(LockNestableUsedAsSimple, func); 168 } 169 if (__kmp_get_tas_lock_owner(lck) == -1) { 170 KMP_FATAL(LockUnsettingFree, func); 171 } 172 if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) && 173 (__kmp_get_tas_lock_owner(lck) != gtid)) { 174 KMP_FATAL(LockUnsettingSetByAnother, func); 175 } 176 return __kmp_release_tas_lock(lck, gtid); 177 } 178 179 void __kmp_init_tas_lock(kmp_tas_lock_t *lck) { 180 lck->lk.poll = KMP_LOCK_FREE(tas); 181 } 182 183 void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; } 184 185 static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) { 186 char const *const func = "omp_destroy_lock"; 187 if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) && 188 __kmp_is_tas_lock_nestable(lck)) { 189 KMP_FATAL(LockNestableUsedAsSimple, func); 190 } 191 if (__kmp_get_tas_lock_owner(lck) != -1) { 192 KMP_FATAL(LockStillOwned, func); 193 } 194 __kmp_destroy_tas_lock(lck); 195 } 196 197 // nested test and set locks 198 199 int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) { 200 KMP_DEBUG_ASSERT(gtid >= 0); 201 202 if (__kmp_get_tas_lock_owner(lck) == gtid) { 203 lck->lk.depth_locked += 1; 204 return KMP_LOCK_ACQUIRED_NEXT; 205 } else { 206 __kmp_acquire_tas_lock_timed_template(lck, gtid); 207 lck->lk.depth_locked = 1; 208 return KMP_LOCK_ACQUIRED_FIRST; 209 } 210 } 211 212 static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck, 213 kmp_int32 gtid) { 214 char const *const func = "omp_set_nest_lock"; 215 if (!__kmp_is_tas_lock_nestable(lck)) { 216 KMP_FATAL(LockSimpleUsedAsNestable, func); 217 } 218 return __kmp_acquire_nested_tas_lock(lck, gtid); 219 } 220 221 int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) { 222 int retval; 223 224 KMP_DEBUG_ASSERT(gtid >= 0); 225 226 if (__kmp_get_tas_lock_owner(lck) == gtid) { 227 retval = ++lck->lk.depth_locked; 228 } else if (!__kmp_test_tas_lock(lck, gtid)) { 229 retval = 0; 230 } else { 231 KMP_MB(); 232 retval = lck->lk.depth_locked = 1; 233 } 234 return retval; 235 } 236 237 static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck, 238 kmp_int32 gtid) { 239 char const *const func = "omp_test_nest_lock"; 240 if (!__kmp_is_tas_lock_nestable(lck)) { 241 KMP_FATAL(LockSimpleUsedAsNestable, func); 242 } 243 return __kmp_test_nested_tas_lock(lck, gtid); 244 } 245 246 int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) { 247 KMP_DEBUG_ASSERT(gtid >= 0); 248 249 KMP_MB(); 250 if (--(lck->lk.depth_locked) == 0) { 251 __kmp_release_tas_lock(lck, gtid); 252 return KMP_LOCK_RELEASED; 253 } 254 return KMP_LOCK_STILL_HELD; 255 } 256 257 static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck, 258 kmp_int32 gtid) { 259 char const *const func = "omp_unset_nest_lock"; 260 KMP_MB(); /* in case another processor initialized lock */ 261 if (!__kmp_is_tas_lock_nestable(lck)) { 262 KMP_FATAL(LockSimpleUsedAsNestable, func); 263 } 264 if (__kmp_get_tas_lock_owner(lck) == -1) { 265 KMP_FATAL(LockUnsettingFree, func); 266 } 267 if (__kmp_get_tas_lock_owner(lck) != gtid) { 268 KMP_FATAL(LockUnsettingSetByAnother, func); 269 } 270 return __kmp_release_nested_tas_lock(lck, gtid); 271 } 272 273 void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) { 274 __kmp_init_tas_lock(lck); 275 lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks 276 } 277 278 void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) { 279 __kmp_destroy_tas_lock(lck); 280 lck->lk.depth_locked = 0; 281 } 282 283 static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) { 284 char const *const func = "omp_destroy_nest_lock"; 285 if (!__kmp_is_tas_lock_nestable(lck)) { 286 KMP_FATAL(LockSimpleUsedAsNestable, func); 287 } 288 if (__kmp_get_tas_lock_owner(lck) != -1) { 289 KMP_FATAL(LockStillOwned, func); 290 } 291 __kmp_destroy_nested_tas_lock(lck); 292 } 293 294 #if KMP_USE_FUTEX 295 296 /* ------------------------------------------------------------------------ */ 297 /* futex locks */ 298 299 // futex locks are really just test and set locks, with a different method 300 // of handling contention. They take the same amount of space as test and 301 // set locks, and are allocated the same way (i.e. use the area allocated by 302 // the compiler for non-nested locks / allocate nested locks on the heap). 303 304 static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) { 305 return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1; 306 } 307 308 static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) { 309 return lck->lk.depth_locked != -1; 310 } 311 312 __forceinline static int 313 __kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) { 314 kmp_int32 gtid_code = (gtid + 1) << 1; 315 316 KMP_MB(); 317 318 #ifdef USE_LOCK_PROFILE 319 kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll)); 320 if ((curr != 0) && (curr != gtid_code)) 321 __kmp_printf("LOCK CONTENTION: %p\n", lck); 322 /* else __kmp_printf( "." );*/ 323 #endif /* USE_LOCK_PROFILE */ 324 325 KMP_FSYNC_PREPARE(lck); 326 KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n", 327 lck, lck->lk.poll, gtid)); 328 329 kmp_int32 poll_val; 330 331 while ((poll_val = KMP_COMPARE_AND_STORE_RET32( 332 &(lck->lk.poll), KMP_LOCK_FREE(futex), 333 KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) { 334 335 kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1; 336 KA_TRACE( 337 1000, 338 ("__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n", 339 lck, gtid, poll_val, cond)); 340 341 // NOTE: if you try to use the following condition for this branch 342 // 343 // if ( poll_val & 1 == 0 ) 344 // 345 // Then the 12.0 compiler has a bug where the following block will 346 // always be skipped, regardless of the value of the LSB of poll_val. 347 if (!cond) { 348 // Try to set the lsb in the poll to indicate to the owner 349 // thread that they need to wake this thread up. 350 if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val, 351 poll_val | KMP_LOCK_BUSY(1, futex))) { 352 KA_TRACE( 353 1000, 354 ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n", 355 lck, lck->lk.poll, gtid)); 356 continue; 357 } 358 poll_val |= KMP_LOCK_BUSY(1, futex); 359 360 KA_TRACE(1000, 361 ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck, 362 lck->lk.poll, gtid)); 363 } 364 365 KA_TRACE( 366 1000, 367 ("__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n", 368 lck, gtid, poll_val)); 369 370 long rc; 371 if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL, 372 NULL, 0)) != 0) { 373 KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) " 374 "failed (rc=%ld errno=%d)\n", 375 lck, gtid, poll_val, rc, errno)); 376 continue; 377 } 378 379 KA_TRACE(1000, 380 ("__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n", 381 lck, gtid, poll_val)); 382 // This thread has now done a successful futex wait call and was entered on 383 // the OS futex queue. We must now perform a futex wake call when releasing 384 // the lock, as we have no idea how many other threads are in the queue. 385 gtid_code |= 1; 386 } 387 388 KMP_FSYNC_ACQUIRED(lck); 389 KA_TRACE(1000, ("__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck, 390 lck->lk.poll, gtid)); 391 return KMP_LOCK_ACQUIRED_FIRST; 392 } 393 394 int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) { 395 int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid); 396 return retval; 397 } 398 399 static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck, 400 kmp_int32 gtid) { 401 char const *const func = "omp_set_lock"; 402 if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) && 403 __kmp_is_futex_lock_nestable(lck)) { 404 KMP_FATAL(LockNestableUsedAsSimple, func); 405 } 406 if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) { 407 KMP_FATAL(LockIsAlreadyOwned, func); 408 } 409 return __kmp_acquire_futex_lock(lck, gtid); 410 } 411 412 int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) { 413 if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex), 414 KMP_LOCK_BUSY((gtid + 1) << 1, futex))) { 415 KMP_FSYNC_ACQUIRED(lck); 416 return TRUE; 417 } 418 return FALSE; 419 } 420 421 static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck, 422 kmp_int32 gtid) { 423 char const *const func = "omp_test_lock"; 424 if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) && 425 __kmp_is_futex_lock_nestable(lck)) { 426 KMP_FATAL(LockNestableUsedAsSimple, func); 427 } 428 return __kmp_test_futex_lock(lck, gtid); 429 } 430 431 int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) { 432 KMP_MB(); /* Flush all pending memory write invalidates. */ 433 434 KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n", 435 lck, lck->lk.poll, gtid)); 436 437 KMP_FSYNC_RELEASING(lck); 438 439 kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex)); 440 441 KA_TRACE(1000, 442 ("__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n", 443 lck, gtid, poll_val)); 444 445 if (KMP_LOCK_STRIP(poll_val) & 1) { 446 KA_TRACE(1000, 447 ("__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n", 448 lck, gtid)); 449 syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex), 450 NULL, NULL, 0); 451 } 452 453 KMP_MB(); /* Flush all pending memory write invalidates. */ 454 455 KA_TRACE(1000, ("__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck, 456 lck->lk.poll, gtid)); 457 458 KMP_YIELD_OVERSUB(); 459 return KMP_LOCK_RELEASED; 460 } 461 462 static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck, 463 kmp_int32 gtid) { 464 char const *const func = "omp_unset_lock"; 465 KMP_MB(); /* in case another processor initialized lock */ 466 if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) && 467 __kmp_is_futex_lock_nestable(lck)) { 468 KMP_FATAL(LockNestableUsedAsSimple, func); 469 } 470 if (__kmp_get_futex_lock_owner(lck) == -1) { 471 KMP_FATAL(LockUnsettingFree, func); 472 } 473 if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) && 474 (__kmp_get_futex_lock_owner(lck) != gtid)) { 475 KMP_FATAL(LockUnsettingSetByAnother, func); 476 } 477 return __kmp_release_futex_lock(lck, gtid); 478 } 479 480 void __kmp_init_futex_lock(kmp_futex_lock_t *lck) { 481 TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex)); 482 } 483 484 void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; } 485 486 static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) { 487 char const *const func = "omp_destroy_lock"; 488 if ((sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) && 489 __kmp_is_futex_lock_nestable(lck)) { 490 KMP_FATAL(LockNestableUsedAsSimple, func); 491 } 492 if (__kmp_get_futex_lock_owner(lck) != -1) { 493 KMP_FATAL(LockStillOwned, func); 494 } 495 __kmp_destroy_futex_lock(lck); 496 } 497 498 // nested futex locks 499 500 int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) { 501 KMP_DEBUG_ASSERT(gtid >= 0); 502 503 if (__kmp_get_futex_lock_owner(lck) == gtid) { 504 lck->lk.depth_locked += 1; 505 return KMP_LOCK_ACQUIRED_NEXT; 506 } else { 507 __kmp_acquire_futex_lock_timed_template(lck, gtid); 508 lck->lk.depth_locked = 1; 509 return KMP_LOCK_ACQUIRED_FIRST; 510 } 511 } 512 513 static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck, 514 kmp_int32 gtid) { 515 char const *const func = "omp_set_nest_lock"; 516 if (!__kmp_is_futex_lock_nestable(lck)) { 517 KMP_FATAL(LockSimpleUsedAsNestable, func); 518 } 519 return __kmp_acquire_nested_futex_lock(lck, gtid); 520 } 521 522 int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) { 523 int retval; 524 525 KMP_DEBUG_ASSERT(gtid >= 0); 526 527 if (__kmp_get_futex_lock_owner(lck) == gtid) { 528 retval = ++lck->lk.depth_locked; 529 } else if (!__kmp_test_futex_lock(lck, gtid)) { 530 retval = 0; 531 } else { 532 KMP_MB(); 533 retval = lck->lk.depth_locked = 1; 534 } 535 return retval; 536 } 537 538 static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck, 539 kmp_int32 gtid) { 540 char const *const func = "omp_test_nest_lock"; 541 if (!__kmp_is_futex_lock_nestable(lck)) { 542 KMP_FATAL(LockSimpleUsedAsNestable, func); 543 } 544 return __kmp_test_nested_futex_lock(lck, gtid); 545 } 546 547 int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) { 548 KMP_DEBUG_ASSERT(gtid >= 0); 549 550 KMP_MB(); 551 if (--(lck->lk.depth_locked) == 0) { 552 __kmp_release_futex_lock(lck, gtid); 553 return KMP_LOCK_RELEASED; 554 } 555 return KMP_LOCK_STILL_HELD; 556 } 557 558 static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck, 559 kmp_int32 gtid) { 560 char const *const func = "omp_unset_nest_lock"; 561 KMP_MB(); /* in case another processor initialized lock */ 562 if (!__kmp_is_futex_lock_nestable(lck)) { 563 KMP_FATAL(LockSimpleUsedAsNestable, func); 564 } 565 if (__kmp_get_futex_lock_owner(lck) == -1) { 566 KMP_FATAL(LockUnsettingFree, func); 567 } 568 if (__kmp_get_futex_lock_owner(lck) != gtid) { 569 KMP_FATAL(LockUnsettingSetByAnother, func); 570 } 571 return __kmp_release_nested_futex_lock(lck, gtid); 572 } 573 574 void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) { 575 __kmp_init_futex_lock(lck); 576 lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks 577 } 578 579 void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) { 580 __kmp_destroy_futex_lock(lck); 581 lck->lk.depth_locked = 0; 582 } 583 584 static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) { 585 char const *const func = "omp_destroy_nest_lock"; 586 if (!__kmp_is_futex_lock_nestable(lck)) { 587 KMP_FATAL(LockSimpleUsedAsNestable, func); 588 } 589 if (__kmp_get_futex_lock_owner(lck) != -1) { 590 KMP_FATAL(LockStillOwned, func); 591 } 592 __kmp_destroy_nested_futex_lock(lck); 593 } 594 595 #endif // KMP_USE_FUTEX 596 597 /* ------------------------------------------------------------------------ */ 598 /* ticket (bakery) locks */ 599 600 static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) { 601 return std::atomic_load_explicit(&lck->lk.owner_id, 602 std::memory_order_relaxed) - 603 1; 604 } 605 606 static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) { 607 return std::atomic_load_explicit(&lck->lk.depth_locked, 608 std::memory_order_relaxed) != -1; 609 } 610 611 static kmp_uint32 __kmp_bakery_check(void *now_serving, kmp_uint32 my_ticket) { 612 return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving, 613 std::memory_order_acquire) == my_ticket; 614 } 615 616 __forceinline static int 617 __kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck, 618 kmp_int32 gtid) { 619 kmp_uint32 my_ticket = std::atomic_fetch_add_explicit( 620 &lck->lk.next_ticket, 1U, std::memory_order_relaxed); 621 622 #ifdef USE_LOCK_PROFILE 623 if (std::atomic_load_explicit(&lck->lk.now_serving, 624 std::memory_order_relaxed) != my_ticket) 625 __kmp_printf("LOCK CONTENTION: %p\n", lck); 626 /* else __kmp_printf( "." );*/ 627 #endif /* USE_LOCK_PROFILE */ 628 629 if (std::atomic_load_explicit(&lck->lk.now_serving, 630 std::memory_order_acquire) == my_ticket) { 631 return KMP_LOCK_ACQUIRED_FIRST; 632 } 633 KMP_WAIT_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck); 634 return KMP_LOCK_ACQUIRED_FIRST; 635 } 636 637 int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) { 638 int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid); 639 return retval; 640 } 641 642 static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck, 643 kmp_int32 gtid) { 644 char const *const func = "omp_set_lock"; 645 646 if (!std::atomic_load_explicit(&lck->lk.initialized, 647 std::memory_order_relaxed)) { 648 KMP_FATAL(LockIsUninitialized, func); 649 } 650 if (lck->lk.self != lck) { 651 KMP_FATAL(LockIsUninitialized, func); 652 } 653 if (__kmp_is_ticket_lock_nestable(lck)) { 654 KMP_FATAL(LockNestableUsedAsSimple, func); 655 } 656 if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) { 657 KMP_FATAL(LockIsAlreadyOwned, func); 658 } 659 660 __kmp_acquire_ticket_lock(lck, gtid); 661 662 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1, 663 std::memory_order_relaxed); 664 return KMP_LOCK_ACQUIRED_FIRST; 665 } 666 667 int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) { 668 kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket, 669 std::memory_order_relaxed); 670 671 if (std::atomic_load_explicit(&lck->lk.now_serving, 672 std::memory_order_relaxed) == my_ticket) { 673 kmp_uint32 next_ticket = my_ticket + 1; 674 if (std::atomic_compare_exchange_strong_explicit( 675 &lck->lk.next_ticket, &my_ticket, next_ticket, 676 std::memory_order_acquire, std::memory_order_acquire)) { 677 return TRUE; 678 } 679 } 680 return FALSE; 681 } 682 683 static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck, 684 kmp_int32 gtid) { 685 char const *const func = "omp_test_lock"; 686 687 if (!std::atomic_load_explicit(&lck->lk.initialized, 688 std::memory_order_relaxed)) { 689 KMP_FATAL(LockIsUninitialized, func); 690 } 691 if (lck->lk.self != lck) { 692 KMP_FATAL(LockIsUninitialized, func); 693 } 694 if (__kmp_is_ticket_lock_nestable(lck)) { 695 KMP_FATAL(LockNestableUsedAsSimple, func); 696 } 697 698 int retval = __kmp_test_ticket_lock(lck, gtid); 699 700 if (retval) { 701 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1, 702 std::memory_order_relaxed); 703 } 704 return retval; 705 } 706 707 int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) { 708 kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket, 709 std::memory_order_relaxed) - 710 std::atomic_load_explicit(&lck->lk.now_serving, 711 std::memory_order_relaxed); 712 713 std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U, 714 std::memory_order_release); 715 716 KMP_YIELD(distance > 717 (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)); 718 return KMP_LOCK_RELEASED; 719 } 720 721 static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck, 722 kmp_int32 gtid) { 723 char const *const func = "omp_unset_lock"; 724 725 if (!std::atomic_load_explicit(&lck->lk.initialized, 726 std::memory_order_relaxed)) { 727 KMP_FATAL(LockIsUninitialized, func); 728 } 729 if (lck->lk.self != lck) { 730 KMP_FATAL(LockIsUninitialized, func); 731 } 732 if (__kmp_is_ticket_lock_nestable(lck)) { 733 KMP_FATAL(LockNestableUsedAsSimple, func); 734 } 735 if (__kmp_get_ticket_lock_owner(lck) == -1) { 736 KMP_FATAL(LockUnsettingFree, func); 737 } 738 if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) && 739 (__kmp_get_ticket_lock_owner(lck) != gtid)) { 740 KMP_FATAL(LockUnsettingSetByAnother, func); 741 } 742 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed); 743 return __kmp_release_ticket_lock(lck, gtid); 744 } 745 746 void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) { 747 lck->lk.location = NULL; 748 lck->lk.self = lck; 749 std::atomic_store_explicit(&lck->lk.next_ticket, 0U, 750 std::memory_order_relaxed); 751 std::atomic_store_explicit(&lck->lk.now_serving, 0U, 752 std::memory_order_relaxed); 753 std::atomic_store_explicit( 754 &lck->lk.owner_id, 0, 755 std::memory_order_relaxed); // no thread owns the lock. 756 std::atomic_store_explicit( 757 &lck->lk.depth_locked, -1, 758 std::memory_order_relaxed); // -1 => not a nested lock. 759 std::atomic_store_explicit(&lck->lk.initialized, true, 760 std::memory_order_release); 761 } 762 763 void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) { 764 std::atomic_store_explicit(&lck->lk.initialized, false, 765 std::memory_order_release); 766 lck->lk.self = NULL; 767 lck->lk.location = NULL; 768 std::atomic_store_explicit(&lck->lk.next_ticket, 0U, 769 std::memory_order_relaxed); 770 std::atomic_store_explicit(&lck->lk.now_serving, 0U, 771 std::memory_order_relaxed); 772 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed); 773 std::atomic_store_explicit(&lck->lk.depth_locked, -1, 774 std::memory_order_relaxed); 775 } 776 777 static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) { 778 char const *const func = "omp_destroy_lock"; 779 780 if (!std::atomic_load_explicit(&lck->lk.initialized, 781 std::memory_order_relaxed)) { 782 KMP_FATAL(LockIsUninitialized, func); 783 } 784 if (lck->lk.self != lck) { 785 KMP_FATAL(LockIsUninitialized, func); 786 } 787 if (__kmp_is_ticket_lock_nestable(lck)) { 788 KMP_FATAL(LockNestableUsedAsSimple, func); 789 } 790 if (__kmp_get_ticket_lock_owner(lck) != -1) { 791 KMP_FATAL(LockStillOwned, func); 792 } 793 __kmp_destroy_ticket_lock(lck); 794 } 795 796 // nested ticket locks 797 798 int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) { 799 KMP_DEBUG_ASSERT(gtid >= 0); 800 801 if (__kmp_get_ticket_lock_owner(lck) == gtid) { 802 std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1, 803 std::memory_order_relaxed); 804 return KMP_LOCK_ACQUIRED_NEXT; 805 } else { 806 __kmp_acquire_ticket_lock_timed_template(lck, gtid); 807 std::atomic_store_explicit(&lck->lk.depth_locked, 1, 808 std::memory_order_relaxed); 809 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1, 810 std::memory_order_relaxed); 811 return KMP_LOCK_ACQUIRED_FIRST; 812 } 813 } 814 815 static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck, 816 kmp_int32 gtid) { 817 char const *const func = "omp_set_nest_lock"; 818 819 if (!std::atomic_load_explicit(&lck->lk.initialized, 820 std::memory_order_relaxed)) { 821 KMP_FATAL(LockIsUninitialized, func); 822 } 823 if (lck->lk.self != lck) { 824 KMP_FATAL(LockIsUninitialized, func); 825 } 826 if (!__kmp_is_ticket_lock_nestable(lck)) { 827 KMP_FATAL(LockSimpleUsedAsNestable, func); 828 } 829 return __kmp_acquire_nested_ticket_lock(lck, gtid); 830 } 831 832 int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) { 833 int retval; 834 835 KMP_DEBUG_ASSERT(gtid >= 0); 836 837 if (__kmp_get_ticket_lock_owner(lck) == gtid) { 838 retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1, 839 std::memory_order_relaxed) + 840 1; 841 } else if (!__kmp_test_ticket_lock(lck, gtid)) { 842 retval = 0; 843 } else { 844 std::atomic_store_explicit(&lck->lk.depth_locked, 1, 845 std::memory_order_relaxed); 846 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1, 847 std::memory_order_relaxed); 848 retval = 1; 849 } 850 return retval; 851 } 852 853 static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck, 854 kmp_int32 gtid) { 855 char const *const func = "omp_test_nest_lock"; 856 857 if (!std::atomic_load_explicit(&lck->lk.initialized, 858 std::memory_order_relaxed)) { 859 KMP_FATAL(LockIsUninitialized, func); 860 } 861 if (lck->lk.self != lck) { 862 KMP_FATAL(LockIsUninitialized, func); 863 } 864 if (!__kmp_is_ticket_lock_nestable(lck)) { 865 KMP_FATAL(LockSimpleUsedAsNestable, func); 866 } 867 return __kmp_test_nested_ticket_lock(lck, gtid); 868 } 869 870 int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) { 871 KMP_DEBUG_ASSERT(gtid >= 0); 872 873 if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1, 874 std::memory_order_relaxed) - 875 1) == 0) { 876 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed); 877 __kmp_release_ticket_lock(lck, gtid); 878 return KMP_LOCK_RELEASED; 879 } 880 return KMP_LOCK_STILL_HELD; 881 } 882 883 static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck, 884 kmp_int32 gtid) { 885 char const *const func = "omp_unset_nest_lock"; 886 887 if (!std::atomic_load_explicit(&lck->lk.initialized, 888 std::memory_order_relaxed)) { 889 KMP_FATAL(LockIsUninitialized, func); 890 } 891 if (lck->lk.self != lck) { 892 KMP_FATAL(LockIsUninitialized, func); 893 } 894 if (!__kmp_is_ticket_lock_nestable(lck)) { 895 KMP_FATAL(LockSimpleUsedAsNestable, func); 896 } 897 if (__kmp_get_ticket_lock_owner(lck) == -1) { 898 KMP_FATAL(LockUnsettingFree, func); 899 } 900 if (__kmp_get_ticket_lock_owner(lck) != gtid) { 901 KMP_FATAL(LockUnsettingSetByAnother, func); 902 } 903 return __kmp_release_nested_ticket_lock(lck, gtid); 904 } 905 906 void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) { 907 __kmp_init_ticket_lock(lck); 908 std::atomic_store_explicit(&lck->lk.depth_locked, 0, 909 std::memory_order_relaxed); 910 // >= 0 for nestable locks, -1 for simple locks 911 } 912 913 void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) { 914 __kmp_destroy_ticket_lock(lck); 915 std::atomic_store_explicit(&lck->lk.depth_locked, 0, 916 std::memory_order_relaxed); 917 } 918 919 static void 920 __kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) { 921 char const *const func = "omp_destroy_nest_lock"; 922 923 if (!std::atomic_load_explicit(&lck->lk.initialized, 924 std::memory_order_relaxed)) { 925 KMP_FATAL(LockIsUninitialized, func); 926 } 927 if (lck->lk.self != lck) { 928 KMP_FATAL(LockIsUninitialized, func); 929 } 930 if (!__kmp_is_ticket_lock_nestable(lck)) { 931 KMP_FATAL(LockSimpleUsedAsNestable, func); 932 } 933 if (__kmp_get_ticket_lock_owner(lck) != -1) { 934 KMP_FATAL(LockStillOwned, func); 935 } 936 __kmp_destroy_nested_ticket_lock(lck); 937 } 938 939 // access functions to fields which don't exist for all lock kinds. 940 941 static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) { 942 return lck->lk.location; 943 } 944 945 static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck, 946 const ident_t *loc) { 947 lck->lk.location = loc; 948 } 949 950 static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) { 951 return lck->lk.flags; 952 } 953 954 static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck, 955 kmp_lock_flags_t flags) { 956 lck->lk.flags = flags; 957 } 958 959 /* ------------------------------------------------------------------------ */ 960 /* queuing locks */ 961 962 /* First the states 963 (head,tail) = 0, 0 means lock is unheld, nobody on queue 964 UINT_MAX or -1, 0 means lock is held, nobody on queue 965 h, h means lock held or about to transition, 966 1 element on queue 967 h, t h <> t, means lock is held or about to 968 transition, >1 elements on queue 969 970 Now the transitions 971 Acquire(0,0) = -1 ,0 972 Release(0,0) = Error 973 Acquire(-1,0) = h ,h h > 0 974 Release(-1,0) = 0 ,0 975 Acquire(h,h) = h ,t h > 0, t > 0, h <> t 976 Release(h,h) = -1 ,0 h > 0 977 Acquire(h,t) = h ,t' h > 0, t > 0, t' > 0, h <> t, h <> t', t <> t' 978 Release(h,t) = h',t h > 0, t > 0, h <> t, h <> h', h' maybe = t 979 980 And pictorially 981 982 +-----+ 983 | 0, 0|------- release -------> Error 984 +-----+ 985 | ^ 986 acquire| |release 987 | | 988 | | 989 v | 990 +-----+ 991 |-1, 0| 992 +-----+ 993 | ^ 994 acquire| |release 995 | | 996 | | 997 v | 998 +-----+ 999 | h, h| 1000 +-----+ 1001 | ^ 1002 acquire| |release 1003 | | 1004 | | 1005 v | 1006 +-----+ 1007 | h, t|----- acquire, release loopback ---+ 1008 +-----+ | 1009 ^ | 1010 | | 1011 +------------------------------------+ 1012 */ 1013 1014 #ifdef DEBUG_QUEUING_LOCKS 1015 1016 /* Stuff for circular trace buffer */ 1017 #define TRACE_BUF_ELE 1024 1018 static char traces[TRACE_BUF_ELE][128] = {0}; 1019 static int tc = 0; 1020 #define TRACE_LOCK(X, Y) \ 1021 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y); 1022 #define TRACE_LOCK_T(X, Y, Z) \ 1023 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z); 1024 #define TRACE_LOCK_HT(X, Y, Z, Q) \ 1025 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y, \ 1026 Z, Q); 1027 1028 static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid, 1029 kmp_queuing_lock_t *lck, kmp_int32 head_id, 1030 kmp_int32 tail_id) { 1031 kmp_int32 t, i; 1032 1033 __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n"); 1034 1035 i = tc % TRACE_BUF_ELE; 1036 __kmp_printf_no_lock("%s\n", traces[i]); 1037 i = (i + 1) % TRACE_BUF_ELE; 1038 while (i != (tc % TRACE_BUF_ELE)) { 1039 __kmp_printf_no_lock("%s", traces[i]); 1040 i = (i + 1) % TRACE_BUF_ELE; 1041 } 1042 __kmp_printf_no_lock("\n"); 1043 1044 __kmp_printf_no_lock("\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, " 1045 "next_wait:%d, head_id:%d, tail_id:%d\n", 1046 gtid + 1, this_thr->th.th_spin_here, 1047 this_thr->th.th_next_waiting, head_id, tail_id); 1048 1049 __kmp_printf_no_lock("\t\thead: %d ", lck->lk.head_id); 1050 1051 if (lck->lk.head_id >= 1) { 1052 t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting; 1053 while (t > 0) { 1054 __kmp_printf_no_lock("-> %d ", t); 1055 t = __kmp_threads[t - 1]->th.th_next_waiting; 1056 } 1057 } 1058 __kmp_printf_no_lock("; tail: %d ", lck->lk.tail_id); 1059 __kmp_printf_no_lock("\n\n"); 1060 } 1061 1062 #endif /* DEBUG_QUEUING_LOCKS */ 1063 1064 static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) { 1065 return TCR_4(lck->lk.owner_id) - 1; 1066 } 1067 1068 static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) { 1069 return lck->lk.depth_locked != -1; 1070 } 1071 1072 /* Acquire a lock using a the queuing lock implementation */ 1073 template <bool takeTime> 1074 /* [TLW] The unused template above is left behind because of what BEB believes 1075 is a potential compiler problem with __forceinline. */ 1076 __forceinline static int 1077 __kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck, 1078 kmp_int32 gtid) { 1079 kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid); 1080 volatile kmp_int32 *head_id_p = &lck->lk.head_id; 1081 volatile kmp_int32 *tail_id_p = &lck->lk.tail_id; 1082 volatile kmp_uint32 *spin_here_p; 1083 1084 #if OMPT_SUPPORT 1085 ompt_state_t prev_state = ompt_state_undefined; 1086 #endif 1087 1088 KA_TRACE(1000, 1089 ("__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid)); 1090 1091 KMP_FSYNC_PREPARE(lck); 1092 KMP_DEBUG_ASSERT(this_thr != NULL); 1093 spin_here_p = &this_thr->th.th_spin_here; 1094 1095 #ifdef DEBUG_QUEUING_LOCKS 1096 TRACE_LOCK(gtid + 1, "acq ent"); 1097 if (*spin_here_p) 1098 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p); 1099 if (this_thr->th.th_next_waiting != 0) 1100 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p); 1101 #endif 1102 KMP_DEBUG_ASSERT(!*spin_here_p); 1103 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0); 1104 1105 /* The following st.rel to spin_here_p needs to precede the cmpxchg.acq to 1106 head_id_p that may follow, not just in execution order, but also in 1107 visibility order. This way, when a releasing thread observes the changes to 1108 the queue by this thread, it can rightly assume that spin_here_p has 1109 already been set to TRUE, so that when it sets spin_here_p to FALSE, it is 1110 not premature. If the releasing thread sets spin_here_p to FALSE before 1111 this thread sets it to TRUE, this thread will hang. */ 1112 *spin_here_p = TRUE; /* before enqueuing to prevent race */ 1113 1114 while (1) { 1115 kmp_int32 enqueued; 1116 kmp_int32 head; 1117 kmp_int32 tail; 1118 1119 head = *head_id_p; 1120 1121 switch (head) { 1122 1123 case -1: { 1124 #ifdef DEBUG_QUEUING_LOCKS 1125 tail = *tail_id_p; 1126 TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail); 1127 #endif 1128 tail = 0; /* to make sure next link asynchronously read is not set 1129 accidentally; this assignment prevents us from entering the 1130 if ( t > 0 ) condition in the enqueued case below, which is not 1131 necessary for this state transition */ 1132 1133 /* try (-1,0)->(tid,tid) */ 1134 enqueued = KMP_COMPARE_AND_STORE_ACQ64((volatile kmp_int64 *)tail_id_p, 1135 KMP_PACK_64(-1, 0), 1136 KMP_PACK_64(gtid + 1, gtid + 1)); 1137 #ifdef DEBUG_QUEUING_LOCKS 1138 if (enqueued) 1139 TRACE_LOCK(gtid + 1, "acq enq: (-1,0)->(tid,tid)"); 1140 #endif 1141 } break; 1142 1143 default: { 1144 tail = *tail_id_p; 1145 KMP_DEBUG_ASSERT(tail != gtid + 1); 1146 1147 #ifdef DEBUG_QUEUING_LOCKS 1148 TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail); 1149 #endif 1150 1151 if (tail == 0) { 1152 enqueued = FALSE; 1153 } else { 1154 /* try (h,t) or (h,h)->(h,tid) */ 1155 enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1); 1156 1157 #ifdef DEBUG_QUEUING_LOCKS 1158 if (enqueued) 1159 TRACE_LOCK(gtid + 1, "acq enq: (h,t)->(h,tid)"); 1160 #endif 1161 } 1162 } break; 1163 1164 case 0: /* empty queue */ 1165 { 1166 kmp_int32 grabbed_lock; 1167 1168 #ifdef DEBUG_QUEUING_LOCKS 1169 tail = *tail_id_p; 1170 TRACE_LOCK_HT(gtid + 1, "acq read: ", head, tail); 1171 #endif 1172 /* try (0,0)->(-1,0) */ 1173 1174 /* only legal transition out of head = 0 is head = -1 with no change to 1175 * tail */ 1176 grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1); 1177 1178 if (grabbed_lock) { 1179 1180 *spin_here_p = FALSE; 1181 1182 KA_TRACE( 1183 1000, 1184 ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n", 1185 lck, gtid)); 1186 #ifdef DEBUG_QUEUING_LOCKS 1187 TRACE_LOCK_HT(gtid + 1, "acq exit: ", head, 0); 1188 #endif 1189 1190 #if OMPT_SUPPORT 1191 if (ompt_enabled.enabled && prev_state != ompt_state_undefined) { 1192 /* change the state before clearing wait_id */ 1193 this_thr->th.ompt_thread_info.state = prev_state; 1194 this_thr->th.ompt_thread_info.wait_id = 0; 1195 } 1196 #endif 1197 1198 KMP_FSYNC_ACQUIRED(lck); 1199 return KMP_LOCK_ACQUIRED_FIRST; /* lock holder cannot be on queue */ 1200 } 1201 enqueued = FALSE; 1202 } break; 1203 } 1204 1205 #if OMPT_SUPPORT 1206 if (ompt_enabled.enabled && prev_state == ompt_state_undefined) { 1207 /* this thread will spin; set wait_id before entering wait state */ 1208 prev_state = this_thr->th.ompt_thread_info.state; 1209 this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck; 1210 this_thr->th.ompt_thread_info.state = ompt_state_wait_lock; 1211 } 1212 #endif 1213 1214 if (enqueued) { 1215 if (tail > 0) { 1216 kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1); 1217 KMP_ASSERT(tail_thr != NULL); 1218 tail_thr->th.th_next_waiting = gtid + 1; 1219 /* corresponding wait for this write in release code */ 1220 } 1221 KA_TRACE(1000, 1222 ("__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n", 1223 lck, gtid)); 1224 1225 KMP_MB(); 1226 // ToDo: Use __kmp_wait_sleep or similar when blocktime != inf 1227 KMP_WAIT(spin_here_p, FALSE, KMP_EQ, lck); 1228 // Synchronize writes to both runtime thread structures 1229 // and writes in user code. 1230 KMP_MB(); 1231 1232 #ifdef DEBUG_QUEUING_LOCKS 1233 TRACE_LOCK(gtid + 1, "acq spin"); 1234 1235 if (this_thr->th.th_next_waiting != 0) 1236 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p); 1237 #endif 1238 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0); 1239 KA_TRACE(1000, ("__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after " 1240 "waiting on queue\n", 1241 lck, gtid)); 1242 1243 #ifdef DEBUG_QUEUING_LOCKS 1244 TRACE_LOCK(gtid + 1, "acq exit 2"); 1245 #endif 1246 1247 #if OMPT_SUPPORT 1248 /* change the state before clearing wait_id */ 1249 this_thr->th.ompt_thread_info.state = prev_state; 1250 this_thr->th.ompt_thread_info.wait_id = 0; 1251 #endif 1252 1253 /* got lock, we were dequeued by the thread that released lock */ 1254 return KMP_LOCK_ACQUIRED_FIRST; 1255 } 1256 1257 /* Yield if number of threads > number of logical processors */ 1258 /* ToDo: Not sure why this should only be in oversubscription case, 1259 maybe should be traditional YIELD_INIT/YIELD_WHEN loop */ 1260 KMP_YIELD_OVERSUB(); 1261 1262 #ifdef DEBUG_QUEUING_LOCKS 1263 TRACE_LOCK(gtid + 1, "acq retry"); 1264 #endif 1265 } 1266 KMP_ASSERT2(0, "should not get here"); 1267 return KMP_LOCK_ACQUIRED_FIRST; 1268 } 1269 1270 int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) { 1271 KMP_DEBUG_ASSERT(gtid >= 0); 1272 1273 int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid); 1274 return retval; 1275 } 1276 1277 static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 1278 kmp_int32 gtid) { 1279 char const *const func = "omp_set_lock"; 1280 if (lck->lk.initialized != lck) { 1281 KMP_FATAL(LockIsUninitialized, func); 1282 } 1283 if (__kmp_is_queuing_lock_nestable(lck)) { 1284 KMP_FATAL(LockNestableUsedAsSimple, func); 1285 } 1286 if (__kmp_get_queuing_lock_owner(lck) == gtid) { 1287 KMP_FATAL(LockIsAlreadyOwned, func); 1288 } 1289 1290 __kmp_acquire_queuing_lock(lck, gtid); 1291 1292 lck->lk.owner_id = gtid + 1; 1293 return KMP_LOCK_ACQUIRED_FIRST; 1294 } 1295 1296 int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) { 1297 volatile kmp_int32 *head_id_p = &lck->lk.head_id; 1298 kmp_int32 head; 1299 #ifdef KMP_DEBUG 1300 kmp_info_t *this_thr; 1301 #endif 1302 1303 KA_TRACE(1000, ("__kmp_test_queuing_lock: T#%d entering\n", gtid)); 1304 KMP_DEBUG_ASSERT(gtid >= 0); 1305 #ifdef KMP_DEBUG 1306 this_thr = __kmp_thread_from_gtid(gtid); 1307 KMP_DEBUG_ASSERT(this_thr != NULL); 1308 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here); 1309 #endif 1310 1311 head = *head_id_p; 1312 1313 if (head == 0) { /* nobody on queue, nobody holding */ 1314 /* try (0,0)->(-1,0) */ 1315 if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) { 1316 KA_TRACE(1000, 1317 ("__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid)); 1318 KMP_FSYNC_ACQUIRED(lck); 1319 return TRUE; 1320 } 1321 } 1322 1323 KA_TRACE(1000, 1324 ("__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid)); 1325 return FALSE; 1326 } 1327 1328 static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 1329 kmp_int32 gtid) { 1330 char const *const func = "omp_test_lock"; 1331 if (lck->lk.initialized != lck) { 1332 KMP_FATAL(LockIsUninitialized, func); 1333 } 1334 if (__kmp_is_queuing_lock_nestable(lck)) { 1335 KMP_FATAL(LockNestableUsedAsSimple, func); 1336 } 1337 1338 int retval = __kmp_test_queuing_lock(lck, gtid); 1339 1340 if (retval) { 1341 lck->lk.owner_id = gtid + 1; 1342 } 1343 return retval; 1344 } 1345 1346 int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) { 1347 kmp_info_t *this_thr; 1348 volatile kmp_int32 *head_id_p = &lck->lk.head_id; 1349 volatile kmp_int32 *tail_id_p = &lck->lk.tail_id; 1350 1351 KA_TRACE(1000, 1352 ("__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid)); 1353 KMP_DEBUG_ASSERT(gtid >= 0); 1354 this_thr = __kmp_thread_from_gtid(gtid); 1355 KMP_DEBUG_ASSERT(this_thr != NULL); 1356 #ifdef DEBUG_QUEUING_LOCKS 1357 TRACE_LOCK(gtid + 1, "rel ent"); 1358 1359 if (this_thr->th.th_spin_here) 1360 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p); 1361 if (this_thr->th.th_next_waiting != 0) 1362 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p); 1363 #endif 1364 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here); 1365 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0); 1366 1367 KMP_FSYNC_RELEASING(lck); 1368 1369 while (1) { 1370 kmp_int32 dequeued; 1371 kmp_int32 head; 1372 kmp_int32 tail; 1373 1374 head = *head_id_p; 1375 1376 #ifdef DEBUG_QUEUING_LOCKS 1377 tail = *tail_id_p; 1378 TRACE_LOCK_HT(gtid + 1, "rel read: ", head, tail); 1379 if (head == 0) 1380 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail); 1381 #endif 1382 KMP_DEBUG_ASSERT(head != 1383 0); /* holding the lock, head must be -1 or queue head */ 1384 1385 if (head == -1) { /* nobody on queue */ 1386 /* try (-1,0)->(0,0) */ 1387 if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) { 1388 KA_TRACE( 1389 1000, 1390 ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n", 1391 lck, gtid)); 1392 #ifdef DEBUG_QUEUING_LOCKS 1393 TRACE_LOCK_HT(gtid + 1, "rel exit: ", 0, 0); 1394 #endif 1395 1396 #if OMPT_SUPPORT 1397 /* nothing to do - no other thread is trying to shift blame */ 1398 #endif 1399 return KMP_LOCK_RELEASED; 1400 } 1401 dequeued = FALSE; 1402 } else { 1403 KMP_MB(); 1404 tail = *tail_id_p; 1405 if (head == tail) { /* only one thread on the queue */ 1406 #ifdef DEBUG_QUEUING_LOCKS 1407 if (head <= 0) 1408 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail); 1409 #endif 1410 KMP_DEBUG_ASSERT(head > 0); 1411 1412 /* try (h,h)->(-1,0) */ 1413 dequeued = KMP_COMPARE_AND_STORE_REL64( 1414 RCAST(volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head), 1415 KMP_PACK_64(-1, 0)); 1416 #ifdef DEBUG_QUEUING_LOCKS 1417 TRACE_LOCK(gtid + 1, "rel deq: (h,h)->(-1,0)"); 1418 #endif 1419 1420 } else { 1421 volatile kmp_int32 *waiting_id_p; 1422 kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1); 1423 KMP_DEBUG_ASSERT(head_thr != NULL); 1424 waiting_id_p = &head_thr->th.th_next_waiting; 1425 1426 /* Does this require synchronous reads? */ 1427 #ifdef DEBUG_QUEUING_LOCKS 1428 if (head <= 0 || tail <= 0) 1429 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail); 1430 #endif 1431 KMP_DEBUG_ASSERT(head > 0 && tail > 0); 1432 1433 /* try (h,t)->(h',t) or (t,t) */ 1434 KMP_MB(); 1435 /* make sure enqueuing thread has time to update next waiting thread 1436 * field */ 1437 *head_id_p = 1438 KMP_WAIT((volatile kmp_uint32 *)waiting_id_p, 0, KMP_NEQ, NULL); 1439 #ifdef DEBUG_QUEUING_LOCKS 1440 TRACE_LOCK(gtid + 1, "rel deq: (h,t)->(h',t)"); 1441 #endif 1442 dequeued = TRUE; 1443 } 1444 } 1445 1446 if (dequeued) { 1447 kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1); 1448 KMP_DEBUG_ASSERT(head_thr != NULL); 1449 1450 /* Does this require synchronous reads? */ 1451 #ifdef DEBUG_QUEUING_LOCKS 1452 if (head <= 0 || tail <= 0) 1453 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail); 1454 #endif 1455 KMP_DEBUG_ASSERT(head > 0 && tail > 0); 1456 1457 /* For clean code only. Thread not released until next statement prevents 1458 race with acquire code. */ 1459 head_thr->th.th_next_waiting = 0; 1460 #ifdef DEBUG_QUEUING_LOCKS 1461 TRACE_LOCK_T(gtid + 1, "rel nw=0 for t=", head); 1462 #endif 1463 1464 KMP_MB(); 1465 /* reset spin value */ 1466 head_thr->th.th_spin_here = FALSE; 1467 1468 KA_TRACE(1000, ("__kmp_release_queuing_lock: lck:%p, T#%d exiting: after " 1469 "dequeuing\n", 1470 lck, gtid)); 1471 #ifdef DEBUG_QUEUING_LOCKS 1472 TRACE_LOCK(gtid + 1, "rel exit 2"); 1473 #endif 1474 return KMP_LOCK_RELEASED; 1475 } 1476 /* KMP_CPU_PAUSE(); don't want to make releasing thread hold up acquiring 1477 threads */ 1478 1479 #ifdef DEBUG_QUEUING_LOCKS 1480 TRACE_LOCK(gtid + 1, "rel retry"); 1481 #endif 1482 1483 } /* while */ 1484 KMP_ASSERT2(0, "should not get here"); 1485 return KMP_LOCK_RELEASED; 1486 } 1487 1488 static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 1489 kmp_int32 gtid) { 1490 char const *const func = "omp_unset_lock"; 1491 KMP_MB(); /* in case another processor initialized lock */ 1492 if (lck->lk.initialized != lck) { 1493 KMP_FATAL(LockIsUninitialized, func); 1494 } 1495 if (__kmp_is_queuing_lock_nestable(lck)) { 1496 KMP_FATAL(LockNestableUsedAsSimple, func); 1497 } 1498 if (__kmp_get_queuing_lock_owner(lck) == -1) { 1499 KMP_FATAL(LockUnsettingFree, func); 1500 } 1501 if (__kmp_get_queuing_lock_owner(lck) != gtid) { 1502 KMP_FATAL(LockUnsettingSetByAnother, func); 1503 } 1504 lck->lk.owner_id = 0; 1505 return __kmp_release_queuing_lock(lck, gtid); 1506 } 1507 1508 void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) { 1509 lck->lk.location = NULL; 1510 lck->lk.head_id = 0; 1511 lck->lk.tail_id = 0; 1512 lck->lk.next_ticket = 0; 1513 lck->lk.now_serving = 0; 1514 lck->lk.owner_id = 0; // no thread owns the lock. 1515 lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks. 1516 lck->lk.initialized = lck; 1517 1518 KA_TRACE(1000, ("__kmp_init_queuing_lock: lock %p initialized\n", lck)); 1519 } 1520 1521 void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) { 1522 lck->lk.initialized = NULL; 1523 lck->lk.location = NULL; 1524 lck->lk.head_id = 0; 1525 lck->lk.tail_id = 0; 1526 lck->lk.next_ticket = 0; 1527 lck->lk.now_serving = 0; 1528 lck->lk.owner_id = 0; 1529 lck->lk.depth_locked = -1; 1530 } 1531 1532 static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) { 1533 char const *const func = "omp_destroy_lock"; 1534 if (lck->lk.initialized != lck) { 1535 KMP_FATAL(LockIsUninitialized, func); 1536 } 1537 if (__kmp_is_queuing_lock_nestable(lck)) { 1538 KMP_FATAL(LockNestableUsedAsSimple, func); 1539 } 1540 if (__kmp_get_queuing_lock_owner(lck) != -1) { 1541 KMP_FATAL(LockStillOwned, func); 1542 } 1543 __kmp_destroy_queuing_lock(lck); 1544 } 1545 1546 // nested queuing locks 1547 1548 int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) { 1549 KMP_DEBUG_ASSERT(gtid >= 0); 1550 1551 if (__kmp_get_queuing_lock_owner(lck) == gtid) { 1552 lck->lk.depth_locked += 1; 1553 return KMP_LOCK_ACQUIRED_NEXT; 1554 } else { 1555 __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid); 1556 KMP_MB(); 1557 lck->lk.depth_locked = 1; 1558 KMP_MB(); 1559 lck->lk.owner_id = gtid + 1; 1560 return KMP_LOCK_ACQUIRED_FIRST; 1561 } 1562 } 1563 1564 static int 1565 __kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 1566 kmp_int32 gtid) { 1567 char const *const func = "omp_set_nest_lock"; 1568 if (lck->lk.initialized != lck) { 1569 KMP_FATAL(LockIsUninitialized, func); 1570 } 1571 if (!__kmp_is_queuing_lock_nestable(lck)) { 1572 KMP_FATAL(LockSimpleUsedAsNestable, func); 1573 } 1574 return __kmp_acquire_nested_queuing_lock(lck, gtid); 1575 } 1576 1577 int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) { 1578 int retval; 1579 1580 KMP_DEBUG_ASSERT(gtid >= 0); 1581 1582 if (__kmp_get_queuing_lock_owner(lck) == gtid) { 1583 retval = ++lck->lk.depth_locked; 1584 } else if (!__kmp_test_queuing_lock(lck, gtid)) { 1585 retval = 0; 1586 } else { 1587 KMP_MB(); 1588 retval = lck->lk.depth_locked = 1; 1589 KMP_MB(); 1590 lck->lk.owner_id = gtid + 1; 1591 } 1592 return retval; 1593 } 1594 1595 static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 1596 kmp_int32 gtid) { 1597 char const *const func = "omp_test_nest_lock"; 1598 if (lck->lk.initialized != lck) { 1599 KMP_FATAL(LockIsUninitialized, func); 1600 } 1601 if (!__kmp_is_queuing_lock_nestable(lck)) { 1602 KMP_FATAL(LockSimpleUsedAsNestable, func); 1603 } 1604 return __kmp_test_nested_queuing_lock(lck, gtid); 1605 } 1606 1607 int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) { 1608 KMP_DEBUG_ASSERT(gtid >= 0); 1609 1610 KMP_MB(); 1611 if (--(lck->lk.depth_locked) == 0) { 1612 KMP_MB(); 1613 lck->lk.owner_id = 0; 1614 __kmp_release_queuing_lock(lck, gtid); 1615 return KMP_LOCK_RELEASED; 1616 } 1617 return KMP_LOCK_STILL_HELD; 1618 } 1619 1620 static int 1621 __kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 1622 kmp_int32 gtid) { 1623 char const *const func = "omp_unset_nest_lock"; 1624 KMP_MB(); /* in case another processor initialized lock */ 1625 if (lck->lk.initialized != lck) { 1626 KMP_FATAL(LockIsUninitialized, func); 1627 } 1628 if (!__kmp_is_queuing_lock_nestable(lck)) { 1629 KMP_FATAL(LockSimpleUsedAsNestable, func); 1630 } 1631 if (__kmp_get_queuing_lock_owner(lck) == -1) { 1632 KMP_FATAL(LockUnsettingFree, func); 1633 } 1634 if (__kmp_get_queuing_lock_owner(lck) != gtid) { 1635 KMP_FATAL(LockUnsettingSetByAnother, func); 1636 } 1637 return __kmp_release_nested_queuing_lock(lck, gtid); 1638 } 1639 1640 void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) { 1641 __kmp_init_queuing_lock(lck); 1642 lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks 1643 } 1644 1645 void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) { 1646 __kmp_destroy_queuing_lock(lck); 1647 lck->lk.depth_locked = 0; 1648 } 1649 1650 static void 1651 __kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) { 1652 char const *const func = "omp_destroy_nest_lock"; 1653 if (lck->lk.initialized != lck) { 1654 KMP_FATAL(LockIsUninitialized, func); 1655 } 1656 if (!__kmp_is_queuing_lock_nestable(lck)) { 1657 KMP_FATAL(LockSimpleUsedAsNestable, func); 1658 } 1659 if (__kmp_get_queuing_lock_owner(lck) != -1) { 1660 KMP_FATAL(LockStillOwned, func); 1661 } 1662 __kmp_destroy_nested_queuing_lock(lck); 1663 } 1664 1665 // access functions to fields which don't exist for all lock kinds. 1666 1667 static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) { 1668 return lck->lk.location; 1669 } 1670 1671 static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck, 1672 const ident_t *loc) { 1673 lck->lk.location = loc; 1674 } 1675 1676 static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) { 1677 return lck->lk.flags; 1678 } 1679 1680 static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck, 1681 kmp_lock_flags_t flags) { 1682 lck->lk.flags = flags; 1683 } 1684 1685 #if KMP_USE_ADAPTIVE_LOCKS 1686 1687 /* RTM Adaptive locks */ 1688 1689 #if KMP_HAVE_RTM_INTRINSICS 1690 #include <immintrin.h> 1691 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT) 1692 1693 #else 1694 1695 // Values from the status register after failed speculation. 1696 #define _XBEGIN_STARTED (~0u) 1697 #define _XABORT_EXPLICIT (1 << 0) 1698 #define _XABORT_RETRY (1 << 1) 1699 #define _XABORT_CONFLICT (1 << 2) 1700 #define _XABORT_CAPACITY (1 << 3) 1701 #define _XABORT_DEBUG (1 << 4) 1702 #define _XABORT_NESTED (1 << 5) 1703 #define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF)) 1704 1705 // Aborts for which it's worth trying again immediately 1706 #define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT) 1707 1708 #define STRINGIZE_INTERNAL(arg) #arg 1709 #define STRINGIZE(arg) STRINGIZE_INTERNAL(arg) 1710 1711 // Access to RTM instructions 1712 /*A version of XBegin which returns -1 on speculation, and the value of EAX on 1713 an abort. This is the same definition as the compiler intrinsic that will be 1714 supported at some point. */ 1715 static __inline int _xbegin() { 1716 int res = -1; 1717 1718 #if KMP_OS_WINDOWS 1719 #if KMP_ARCH_X86_64 1720 _asm { 1721 _emit 0xC7 1722 _emit 0xF8 1723 _emit 2 1724 _emit 0 1725 _emit 0 1726 _emit 0 1727 jmp L2 1728 mov res, eax 1729 L2: 1730 } 1731 #else /* IA32 */ 1732 _asm { 1733 _emit 0xC7 1734 _emit 0xF8 1735 _emit 2 1736 _emit 0 1737 _emit 0 1738 _emit 0 1739 jmp L2 1740 mov res, eax 1741 L2: 1742 } 1743 #endif // KMP_ARCH_X86_64 1744 #else 1745 /* Note that %eax must be noted as killed (clobbered), because the XSR is 1746 returned in %eax(%rax) on abort. Other register values are restored, so 1747 don't need to be killed. 1748 1749 We must also mark 'res' as an input and an output, since otherwise 1750 'res=-1' may be dropped as being dead, whereas we do need the assignment on 1751 the successful (i.e., non-abort) path. */ 1752 __asm__ volatile("1: .byte 0xC7; .byte 0xF8;\n" 1753 " .long 1f-1b-6\n" 1754 " jmp 2f\n" 1755 "1: movl %%eax,%0\n" 1756 "2:" 1757 : "+r"(res)::"memory", "%eax"); 1758 #endif // KMP_OS_WINDOWS 1759 return res; 1760 } 1761 1762 /* Transaction end */ 1763 static __inline void _xend() { 1764 #if KMP_OS_WINDOWS 1765 __asm { 1766 _emit 0x0f 1767 _emit 0x01 1768 _emit 0xd5 1769 } 1770 #else 1771 __asm__ volatile(".byte 0x0f; .byte 0x01; .byte 0xd5" ::: "memory"); 1772 #endif 1773 } 1774 1775 /* This is a macro, the argument must be a single byte constant which can be 1776 evaluated by the inline assembler, since it is emitted as a byte into the 1777 assembly code. */ 1778 // clang-format off 1779 #if KMP_OS_WINDOWS 1780 #define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG 1781 #else 1782 #define _xabort(ARG) \ 1783 __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory"); 1784 #endif 1785 // clang-format on 1786 #endif // KMP_COMPILER_ICC && __INTEL_COMPILER >= 1300 1787 1788 // Statistics is collected for testing purpose 1789 #if KMP_DEBUG_ADAPTIVE_LOCKS 1790 1791 // We accumulate speculative lock statistics when the lock is destroyed. We 1792 // keep locks that haven't been destroyed in the liveLocks list so that we can 1793 // grab their statistics too. 1794 static kmp_adaptive_lock_statistics_t destroyedStats; 1795 1796 // To hold the list of live locks. 1797 static kmp_adaptive_lock_info_t liveLocks; 1798 1799 // A lock so we can safely update the list of locks. 1800 static kmp_bootstrap_lock_t chain_lock = 1801 KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock); 1802 1803 // Initialize the list of stats. 1804 void __kmp_init_speculative_stats() { 1805 kmp_adaptive_lock_info_t *lck = &liveLocks; 1806 1807 memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0, 1808 sizeof(lck->stats)); 1809 lck->stats.next = lck; 1810 lck->stats.prev = lck; 1811 1812 KMP_ASSERT(lck->stats.next->stats.prev == lck); 1813 KMP_ASSERT(lck->stats.prev->stats.next == lck); 1814 1815 __kmp_init_bootstrap_lock(&chain_lock); 1816 } 1817 1818 // Insert the lock into the circular list 1819 static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) { 1820 __kmp_acquire_bootstrap_lock(&chain_lock); 1821 1822 lck->stats.next = liveLocks.stats.next; 1823 lck->stats.prev = &liveLocks; 1824 1825 liveLocks.stats.next = lck; 1826 lck->stats.next->stats.prev = lck; 1827 1828 KMP_ASSERT(lck->stats.next->stats.prev == lck); 1829 KMP_ASSERT(lck->stats.prev->stats.next == lck); 1830 1831 __kmp_release_bootstrap_lock(&chain_lock); 1832 } 1833 1834 static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) { 1835 KMP_ASSERT(lck->stats.next->stats.prev == lck); 1836 KMP_ASSERT(lck->stats.prev->stats.next == lck); 1837 1838 kmp_adaptive_lock_info_t *n = lck->stats.next; 1839 kmp_adaptive_lock_info_t *p = lck->stats.prev; 1840 1841 n->stats.prev = p; 1842 p->stats.next = n; 1843 } 1844 1845 static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) { 1846 memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0, 1847 sizeof(lck->stats)); 1848 __kmp_remember_lock(lck); 1849 } 1850 1851 static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t, 1852 kmp_adaptive_lock_info_t *lck) { 1853 kmp_adaptive_lock_statistics_t volatile *s = &lck->stats; 1854 1855 t->nonSpeculativeAcquireAttempts += lck->acquire_attempts; 1856 t->successfulSpeculations += s->successfulSpeculations; 1857 t->hardFailedSpeculations += s->hardFailedSpeculations; 1858 t->softFailedSpeculations += s->softFailedSpeculations; 1859 t->nonSpeculativeAcquires += s->nonSpeculativeAcquires; 1860 t->lemmingYields += s->lemmingYields; 1861 } 1862 1863 static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) { 1864 __kmp_acquire_bootstrap_lock(&chain_lock); 1865 1866 __kmp_add_stats(&destroyedStats, lck); 1867 __kmp_forget_lock(lck); 1868 1869 __kmp_release_bootstrap_lock(&chain_lock); 1870 } 1871 1872 static float percent(kmp_uint32 count, kmp_uint32 total) { 1873 return (total == 0) ? 0.0 : (100.0 * count) / total; 1874 } 1875 1876 void __kmp_print_speculative_stats() { 1877 kmp_adaptive_lock_statistics_t total = destroyedStats; 1878 kmp_adaptive_lock_info_t *lck; 1879 1880 for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) { 1881 __kmp_add_stats(&total, lck); 1882 } 1883 kmp_adaptive_lock_statistics_t *t = &total; 1884 kmp_uint32 totalSections = 1885 t->nonSpeculativeAcquires + t->successfulSpeculations; 1886 kmp_uint32 totalSpeculations = t->successfulSpeculations + 1887 t->hardFailedSpeculations + 1888 t->softFailedSpeculations; 1889 if (totalSections <= 0) 1890 return; 1891 1892 kmp_safe_raii_file_t statsFile; 1893 if (strcmp(__kmp_speculative_statsfile, "-") == 0) { 1894 statsFile.set_stdout(); 1895 } else { 1896 size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20; 1897 char buffer[buffLen]; 1898 KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile, 1899 (kmp_int32)getpid()); 1900 statsFile.open(buffer, "w"); 1901 } 1902 1903 fprintf(statsFile, "Speculative lock statistics (all approximate!)\n"); 1904 fprintf(statsFile, 1905 " Lock parameters: \n" 1906 " max_soft_retries : %10d\n" 1907 " max_badness : %10d\n", 1908 __kmp_adaptive_backoff_params.max_soft_retries, 1909 __kmp_adaptive_backoff_params.max_badness); 1910 fprintf(statsFile, " Non-speculative acquire attempts : %10d\n", 1911 t->nonSpeculativeAcquireAttempts); 1912 fprintf(statsFile, " Total critical sections : %10d\n", 1913 totalSections); 1914 fprintf(statsFile, " Successful speculations : %10d (%5.1f%%)\n", 1915 t->successfulSpeculations, 1916 percent(t->successfulSpeculations, totalSections)); 1917 fprintf(statsFile, " Non-speculative acquires : %10d (%5.1f%%)\n", 1918 t->nonSpeculativeAcquires, 1919 percent(t->nonSpeculativeAcquires, totalSections)); 1920 fprintf(statsFile, " Lemming yields : %10d\n\n", 1921 t->lemmingYields); 1922 1923 fprintf(statsFile, " Speculative acquire attempts : %10d\n", 1924 totalSpeculations); 1925 fprintf(statsFile, " Successes : %10d (%5.1f%%)\n", 1926 t->successfulSpeculations, 1927 percent(t->successfulSpeculations, totalSpeculations)); 1928 fprintf(statsFile, " Soft failures : %10d (%5.1f%%)\n", 1929 t->softFailedSpeculations, 1930 percent(t->softFailedSpeculations, totalSpeculations)); 1931 fprintf(statsFile, " Hard failures : %10d (%5.1f%%)\n", 1932 t->hardFailedSpeculations, 1933 percent(t->hardFailedSpeculations, totalSpeculations)); 1934 } 1935 1936 #define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++) 1937 #else 1938 #define KMP_INC_STAT(lck, stat) 1939 1940 #endif // KMP_DEBUG_ADAPTIVE_LOCKS 1941 1942 static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) { 1943 // It is enough to check that the head_id is zero. 1944 // We don't also need to check the tail. 1945 bool res = lck->lk.head_id == 0; 1946 1947 // We need a fence here, since we must ensure that no memory operations 1948 // from later in this thread float above that read. 1949 #if KMP_COMPILER_ICC 1950 _mm_mfence(); 1951 #else 1952 __sync_synchronize(); 1953 #endif 1954 1955 return res; 1956 } 1957 1958 // Functions for manipulating the badness 1959 static __inline void 1960 __kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) { 1961 // Reset the badness to zero so we eagerly try to speculate again 1962 lck->lk.adaptive.badness = 0; 1963 KMP_INC_STAT(lck, successfulSpeculations); 1964 } 1965 1966 // Create a bit mask with one more set bit. 1967 static __inline void __kmp_step_badness(kmp_adaptive_lock_t *lck) { 1968 kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1; 1969 if (newBadness > lck->lk.adaptive.max_badness) { 1970 return; 1971 } else { 1972 lck->lk.adaptive.badness = newBadness; 1973 } 1974 } 1975 1976 // Check whether speculation should be attempted. 1977 KMP_ATTRIBUTE_TARGET_RTM 1978 static __inline int __kmp_should_speculate(kmp_adaptive_lock_t *lck, 1979 kmp_int32 gtid) { 1980 kmp_uint32 badness = lck->lk.adaptive.badness; 1981 kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts; 1982 int res = (attempts & badness) == 0; 1983 return res; 1984 } 1985 1986 // Attempt to acquire only the speculative lock. 1987 // Does not back off to the non-speculative lock. 1988 KMP_ATTRIBUTE_TARGET_RTM 1989 static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck, 1990 kmp_int32 gtid) { 1991 int retries = lck->lk.adaptive.max_soft_retries; 1992 1993 // We don't explicitly count the start of speculation, rather we record the 1994 // results (success, hard fail, soft fail). The sum of all of those is the 1995 // total number of times we started speculation since all speculations must 1996 // end one of those ways. 1997 do { 1998 kmp_uint32 status = _xbegin(); 1999 // Switch this in to disable actual speculation but exercise at least some 2000 // of the rest of the code. Useful for debugging... 2001 // kmp_uint32 status = _XABORT_NESTED; 2002 2003 if (status == _XBEGIN_STARTED) { 2004 /* We have successfully started speculation. Check that no-one acquired 2005 the lock for real between when we last looked and now. This also gets 2006 the lock cache line into our read-set, which we need so that we'll 2007 abort if anyone later claims it for real. */ 2008 if (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) { 2009 // Lock is now visibly acquired, so someone beat us to it. Abort the 2010 // transaction so we'll restart from _xbegin with the failure status. 2011 _xabort(0x01); 2012 KMP_ASSERT2(0, "should not get here"); 2013 } 2014 return 1; // Lock has been acquired (speculatively) 2015 } else { 2016 // We have aborted, update the statistics 2017 if (status & SOFT_ABORT_MASK) { 2018 KMP_INC_STAT(lck, softFailedSpeculations); 2019 // and loop round to retry. 2020 } else { 2021 KMP_INC_STAT(lck, hardFailedSpeculations); 2022 // Give up if we had a hard failure. 2023 break; 2024 } 2025 } 2026 } while (retries--); // Loop while we have retries, and didn't fail hard. 2027 2028 // Either we had a hard failure or we didn't succeed softly after 2029 // the full set of attempts, so back off the badness. 2030 __kmp_step_badness(lck); 2031 return 0; 2032 } 2033 2034 // Attempt to acquire the speculative lock, or back off to the non-speculative 2035 // one if the speculative lock cannot be acquired. 2036 // We can succeed speculatively, non-speculatively, or fail. 2037 static int __kmp_test_adaptive_lock(kmp_adaptive_lock_t *lck, kmp_int32 gtid) { 2038 // First try to acquire the lock speculatively 2039 if (__kmp_should_speculate(lck, gtid) && 2040 __kmp_test_adaptive_lock_only(lck, gtid)) 2041 return 1; 2042 2043 // Speculative acquisition failed, so try to acquire it non-speculatively. 2044 // Count the non-speculative acquire attempt 2045 lck->lk.adaptive.acquire_attempts++; 2046 2047 // Use base, non-speculative lock. 2048 if (__kmp_test_queuing_lock(GET_QLK_PTR(lck), gtid)) { 2049 KMP_INC_STAT(lck, nonSpeculativeAcquires); 2050 return 1; // Lock is acquired (non-speculatively) 2051 } else { 2052 return 0; // Failed to acquire the lock, it's already visibly locked. 2053 } 2054 } 2055 2056 static int __kmp_test_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck, 2057 kmp_int32 gtid) { 2058 char const *const func = "omp_test_lock"; 2059 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) { 2060 KMP_FATAL(LockIsUninitialized, func); 2061 } 2062 2063 int retval = __kmp_test_adaptive_lock(lck, gtid); 2064 2065 if (retval) { 2066 lck->lk.qlk.owner_id = gtid + 1; 2067 } 2068 return retval; 2069 } 2070 2071 // Block until we can acquire a speculative, adaptive lock. We check whether we 2072 // should be trying to speculate. If we should be, we check the real lock to see 2073 // if it is free, and, if not, pause without attempting to acquire it until it 2074 // is. Then we try the speculative acquire. This means that although we suffer 2075 // from lemmings a little (because all we can't acquire the lock speculatively 2076 // until the queue of threads waiting has cleared), we don't get into a state 2077 // where we can never acquire the lock speculatively (because we force the queue 2078 // to clear by preventing new arrivals from entering the queue). This does mean 2079 // that when we're trying to break lemmings, the lock is no longer fair. However 2080 // OpenMP makes no guarantee that its locks are fair, so this isn't a real 2081 // problem. 2082 static void __kmp_acquire_adaptive_lock(kmp_adaptive_lock_t *lck, 2083 kmp_int32 gtid) { 2084 if (__kmp_should_speculate(lck, gtid)) { 2085 if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) { 2086 if (__kmp_test_adaptive_lock_only(lck, gtid)) 2087 return; 2088 // We tried speculation and failed, so give up. 2089 } else { 2090 // We can't try speculation until the lock is free, so we pause here 2091 // (without suspending on the queueing lock, to allow it to drain, then 2092 // try again. All other threads will also see the same result for 2093 // shouldSpeculate, so will be doing the same if they try to claim the 2094 // lock from now on. 2095 while (!__kmp_is_unlocked_queuing_lock(GET_QLK_PTR(lck))) { 2096 KMP_INC_STAT(lck, lemmingYields); 2097 KMP_YIELD(TRUE); 2098 } 2099 2100 if (__kmp_test_adaptive_lock_only(lck, gtid)) 2101 return; 2102 } 2103 } 2104 2105 // Speculative acquisition failed, so acquire it non-speculatively. 2106 // Count the non-speculative acquire attempt 2107 lck->lk.adaptive.acquire_attempts++; 2108 2109 __kmp_acquire_queuing_lock_timed_template<FALSE>(GET_QLK_PTR(lck), gtid); 2110 // We have acquired the base lock, so count that. 2111 KMP_INC_STAT(lck, nonSpeculativeAcquires); 2112 } 2113 2114 static void __kmp_acquire_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck, 2115 kmp_int32 gtid) { 2116 char const *const func = "omp_set_lock"; 2117 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) { 2118 KMP_FATAL(LockIsUninitialized, func); 2119 } 2120 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == gtid) { 2121 KMP_FATAL(LockIsAlreadyOwned, func); 2122 } 2123 2124 __kmp_acquire_adaptive_lock(lck, gtid); 2125 2126 lck->lk.qlk.owner_id = gtid + 1; 2127 } 2128 2129 KMP_ATTRIBUTE_TARGET_RTM 2130 static int __kmp_release_adaptive_lock(kmp_adaptive_lock_t *lck, 2131 kmp_int32 gtid) { 2132 if (__kmp_is_unlocked_queuing_lock(GET_QLK_PTR( 2133 lck))) { // If the lock doesn't look claimed we must be speculating. 2134 // (Or the user's code is buggy and they're releasing without locking; 2135 // if we had XTEST we'd be able to check that case...) 2136 _xend(); // Exit speculation 2137 __kmp_update_badness_after_success(lck); 2138 } else { // Since the lock *is* visibly locked we're not speculating, 2139 // so should use the underlying lock's release scheme. 2140 __kmp_release_queuing_lock(GET_QLK_PTR(lck), gtid); 2141 } 2142 return KMP_LOCK_RELEASED; 2143 } 2144 2145 static int __kmp_release_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck, 2146 kmp_int32 gtid) { 2147 char const *const func = "omp_unset_lock"; 2148 KMP_MB(); /* in case another processor initialized lock */ 2149 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) { 2150 KMP_FATAL(LockIsUninitialized, func); 2151 } 2152 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) == -1) { 2153 KMP_FATAL(LockUnsettingFree, func); 2154 } 2155 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != gtid) { 2156 KMP_FATAL(LockUnsettingSetByAnother, func); 2157 } 2158 lck->lk.qlk.owner_id = 0; 2159 __kmp_release_adaptive_lock(lck, gtid); 2160 return KMP_LOCK_RELEASED; 2161 } 2162 2163 static void __kmp_init_adaptive_lock(kmp_adaptive_lock_t *lck) { 2164 __kmp_init_queuing_lock(GET_QLK_PTR(lck)); 2165 lck->lk.adaptive.badness = 0; 2166 lck->lk.adaptive.acquire_attempts = 0; // nonSpeculativeAcquireAttempts = 0; 2167 lck->lk.adaptive.max_soft_retries = 2168 __kmp_adaptive_backoff_params.max_soft_retries; 2169 lck->lk.adaptive.max_badness = __kmp_adaptive_backoff_params.max_badness; 2170 #if KMP_DEBUG_ADAPTIVE_LOCKS 2171 __kmp_zero_speculative_stats(&lck->lk.adaptive); 2172 #endif 2173 KA_TRACE(1000, ("__kmp_init_adaptive_lock: lock %p initialized\n", lck)); 2174 } 2175 2176 static void __kmp_destroy_adaptive_lock(kmp_adaptive_lock_t *lck) { 2177 #if KMP_DEBUG_ADAPTIVE_LOCKS 2178 __kmp_accumulate_speculative_stats(&lck->lk.adaptive); 2179 #endif 2180 __kmp_destroy_queuing_lock(GET_QLK_PTR(lck)); 2181 // Nothing needed for the speculative part. 2182 } 2183 2184 static void __kmp_destroy_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) { 2185 char const *const func = "omp_destroy_lock"; 2186 if (lck->lk.qlk.initialized != GET_QLK_PTR(lck)) { 2187 KMP_FATAL(LockIsUninitialized, func); 2188 } 2189 if (__kmp_get_queuing_lock_owner(GET_QLK_PTR(lck)) != -1) { 2190 KMP_FATAL(LockStillOwned, func); 2191 } 2192 __kmp_destroy_adaptive_lock(lck); 2193 } 2194 2195 #endif // KMP_USE_ADAPTIVE_LOCKS 2196 2197 /* ------------------------------------------------------------------------ */ 2198 /* DRDPA ticket locks */ 2199 /* "DRDPA" means Dynamically Reconfigurable Distributed Polling Area */ 2200 2201 static kmp_int32 __kmp_get_drdpa_lock_owner(kmp_drdpa_lock_t *lck) { 2202 return lck->lk.owner_id - 1; 2203 } 2204 2205 static inline bool __kmp_is_drdpa_lock_nestable(kmp_drdpa_lock_t *lck) { 2206 return lck->lk.depth_locked != -1; 2207 } 2208 2209 __forceinline static int 2210 __kmp_acquire_drdpa_lock_timed_template(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2211 kmp_uint64 ticket = KMP_ATOMIC_INC(&lck->lk.next_ticket); 2212 kmp_uint64 mask = lck->lk.mask; // atomic load 2213 std::atomic<kmp_uint64> *polls = lck->lk.polls; 2214 2215 #ifdef USE_LOCK_PROFILE 2216 if (polls[ticket & mask] != ticket) 2217 __kmp_printf("LOCK CONTENTION: %p\n", lck); 2218 /* else __kmp_printf( "." );*/ 2219 #endif /* USE_LOCK_PROFILE */ 2220 2221 // Now spin-wait, but reload the polls pointer and mask, in case the 2222 // polling area has been reconfigured. Unless it is reconfigured, the 2223 // reloads stay in L1 cache and are cheap. 2224 // 2225 // Keep this code in sync with KMP_WAIT, in kmp_dispatch.cpp !!! 2226 // The current implementation of KMP_WAIT doesn't allow for mask 2227 // and poll to be re-read every spin iteration. 2228 kmp_uint32 spins; 2229 KMP_FSYNC_PREPARE(lck); 2230 KMP_INIT_YIELD(spins); 2231 while (polls[ticket & mask] < ticket) { // atomic load 2232 KMP_YIELD_OVERSUB_ELSE_SPIN(spins); 2233 // Re-read the mask and the poll pointer from the lock structure. 2234 // 2235 // Make certain that "mask" is read before "polls" !!! 2236 // 2237 // If another thread picks reconfigures the polling area and updates their 2238 // values, and we get the new value of mask and the old polls pointer, we 2239 // could access memory beyond the end of the old polling area. 2240 mask = lck->lk.mask; // atomic load 2241 polls = lck->lk.polls; // atomic load 2242 } 2243 2244 // Critical section starts here 2245 KMP_FSYNC_ACQUIRED(lck); 2246 KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld acquired lock %p\n", 2247 ticket, lck)); 2248 lck->lk.now_serving = ticket; // non-volatile store 2249 2250 // Deallocate a garbage polling area if we know that we are the last 2251 // thread that could possibly access it. 2252 // 2253 // The >= check is in case __kmp_test_drdpa_lock() allocated the cleanup 2254 // ticket. 2255 if ((lck->lk.old_polls != NULL) && (ticket >= lck->lk.cleanup_ticket)) { 2256 __kmp_free(lck->lk.old_polls); 2257 lck->lk.old_polls = NULL; 2258 lck->lk.cleanup_ticket = 0; 2259 } 2260 2261 // Check to see if we should reconfigure the polling area. 2262 // If there is still a garbage polling area to be deallocated from a 2263 // previous reconfiguration, let a later thread reconfigure it. 2264 if (lck->lk.old_polls == NULL) { 2265 bool reconfigure = false; 2266 std::atomic<kmp_uint64> *old_polls = polls; 2267 kmp_uint32 num_polls = TCR_4(lck->lk.num_polls); 2268 2269 if (TCR_4(__kmp_nth) > 2270 (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) { 2271 // We are in oversubscription mode. Contract the polling area 2272 // down to a single location, if that hasn't been done already. 2273 if (num_polls > 1) { 2274 reconfigure = true; 2275 num_polls = TCR_4(lck->lk.num_polls); 2276 mask = 0; 2277 num_polls = 1; 2278 polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls * 2279 sizeof(*polls)); 2280 polls[0] = ticket; 2281 } 2282 } else { 2283 // We are in under/fully subscribed mode. Check the number of 2284 // threads waiting on the lock. The size of the polling area 2285 // should be at least the number of threads waiting. 2286 kmp_uint64 num_waiting = TCR_8(lck->lk.next_ticket) - ticket - 1; 2287 if (num_waiting > num_polls) { 2288 kmp_uint32 old_num_polls = num_polls; 2289 reconfigure = true; 2290 do { 2291 mask = (mask << 1) | 1; 2292 num_polls *= 2; 2293 } while (num_polls <= num_waiting); 2294 2295 // Allocate the new polling area, and copy the relevant portion 2296 // of the old polling area to the new area. __kmp_allocate() 2297 // zeroes the memory it allocates, and most of the old area is 2298 // just zero padding, so we only copy the release counters. 2299 polls = (std::atomic<kmp_uint64> *)__kmp_allocate(num_polls * 2300 sizeof(*polls)); 2301 kmp_uint32 i; 2302 for (i = 0; i < old_num_polls; i++) { 2303 polls[i].store(old_polls[i]); 2304 } 2305 } 2306 } 2307 2308 if (reconfigure) { 2309 // Now write the updated fields back to the lock structure. 2310 // 2311 // Make certain that "polls" is written before "mask" !!! 2312 // 2313 // If another thread picks up the new value of mask and the old polls 2314 // pointer , it could access memory beyond the end of the old polling 2315 // area. 2316 // 2317 // On x86, we need memory fences. 2318 KA_TRACE(1000, ("__kmp_acquire_drdpa_lock: ticket #%lld reconfiguring " 2319 "lock %p to %d polls\n", 2320 ticket, lck, num_polls)); 2321 2322 lck->lk.old_polls = old_polls; 2323 lck->lk.polls = polls; // atomic store 2324 2325 KMP_MB(); 2326 2327 lck->lk.num_polls = num_polls; 2328 lck->lk.mask = mask; // atomic store 2329 2330 KMP_MB(); 2331 2332 // Only after the new polling area and mask have been flushed 2333 // to main memory can we update the cleanup ticket field. 2334 // 2335 // volatile load / non-volatile store 2336 lck->lk.cleanup_ticket = lck->lk.next_ticket; 2337 } 2338 } 2339 return KMP_LOCK_ACQUIRED_FIRST; 2340 } 2341 2342 int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2343 int retval = __kmp_acquire_drdpa_lock_timed_template(lck, gtid); 2344 return retval; 2345 } 2346 2347 static int __kmp_acquire_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck, 2348 kmp_int32 gtid) { 2349 char const *const func = "omp_set_lock"; 2350 if (lck->lk.initialized != lck) { 2351 KMP_FATAL(LockIsUninitialized, func); 2352 } 2353 if (__kmp_is_drdpa_lock_nestable(lck)) { 2354 KMP_FATAL(LockNestableUsedAsSimple, func); 2355 } 2356 if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) == gtid)) { 2357 KMP_FATAL(LockIsAlreadyOwned, func); 2358 } 2359 2360 __kmp_acquire_drdpa_lock(lck, gtid); 2361 2362 lck->lk.owner_id = gtid + 1; 2363 return KMP_LOCK_ACQUIRED_FIRST; 2364 } 2365 2366 int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2367 // First get a ticket, then read the polls pointer and the mask. 2368 // The polls pointer must be read before the mask!!! (See above) 2369 kmp_uint64 ticket = lck->lk.next_ticket; // atomic load 2370 std::atomic<kmp_uint64> *polls = lck->lk.polls; 2371 kmp_uint64 mask = lck->lk.mask; // atomic load 2372 if (polls[ticket & mask] == ticket) { 2373 kmp_uint64 next_ticket = ticket + 1; 2374 if (__kmp_atomic_compare_store_acq(&lck->lk.next_ticket, ticket, 2375 next_ticket)) { 2376 KMP_FSYNC_ACQUIRED(lck); 2377 KA_TRACE(1000, ("__kmp_test_drdpa_lock: ticket #%lld acquired lock %p\n", 2378 ticket, lck)); 2379 lck->lk.now_serving = ticket; // non-volatile store 2380 2381 // Since no threads are waiting, there is no possibility that we would 2382 // want to reconfigure the polling area. We might have the cleanup ticket 2383 // value (which says that it is now safe to deallocate old_polls), but 2384 // we'll let a later thread which calls __kmp_acquire_lock do that - this 2385 // routine isn't supposed to block, and we would risk blocks if we called 2386 // __kmp_free() to do the deallocation. 2387 return TRUE; 2388 } 2389 } 2390 return FALSE; 2391 } 2392 2393 static int __kmp_test_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck, 2394 kmp_int32 gtid) { 2395 char const *const func = "omp_test_lock"; 2396 if (lck->lk.initialized != lck) { 2397 KMP_FATAL(LockIsUninitialized, func); 2398 } 2399 if (__kmp_is_drdpa_lock_nestable(lck)) { 2400 KMP_FATAL(LockNestableUsedAsSimple, func); 2401 } 2402 2403 int retval = __kmp_test_drdpa_lock(lck, gtid); 2404 2405 if (retval) { 2406 lck->lk.owner_id = gtid + 1; 2407 } 2408 return retval; 2409 } 2410 2411 int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2412 // Read the ticket value from the lock data struct, then the polls pointer and 2413 // the mask. The polls pointer must be read before the mask!!! (See above) 2414 kmp_uint64 ticket = lck->lk.now_serving + 1; // non-atomic load 2415 std::atomic<kmp_uint64> *polls = lck->lk.polls; // atomic load 2416 kmp_uint64 mask = lck->lk.mask; // atomic load 2417 KA_TRACE(1000, ("__kmp_release_drdpa_lock: ticket #%lld released lock %p\n", 2418 ticket - 1, lck)); 2419 KMP_FSYNC_RELEASING(lck); 2420 polls[ticket & mask] = ticket; // atomic store 2421 return KMP_LOCK_RELEASED; 2422 } 2423 2424 static int __kmp_release_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck, 2425 kmp_int32 gtid) { 2426 char const *const func = "omp_unset_lock"; 2427 KMP_MB(); /* in case another processor initialized lock */ 2428 if (lck->lk.initialized != lck) { 2429 KMP_FATAL(LockIsUninitialized, func); 2430 } 2431 if (__kmp_is_drdpa_lock_nestable(lck)) { 2432 KMP_FATAL(LockNestableUsedAsSimple, func); 2433 } 2434 if (__kmp_get_drdpa_lock_owner(lck) == -1) { 2435 KMP_FATAL(LockUnsettingFree, func); 2436 } 2437 if ((gtid >= 0) && (__kmp_get_drdpa_lock_owner(lck) >= 0) && 2438 (__kmp_get_drdpa_lock_owner(lck) != gtid)) { 2439 KMP_FATAL(LockUnsettingSetByAnother, func); 2440 } 2441 lck->lk.owner_id = 0; 2442 return __kmp_release_drdpa_lock(lck, gtid); 2443 } 2444 2445 void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck) { 2446 lck->lk.location = NULL; 2447 lck->lk.mask = 0; 2448 lck->lk.num_polls = 1; 2449 lck->lk.polls = (std::atomic<kmp_uint64> *)__kmp_allocate( 2450 lck->lk.num_polls * sizeof(*(lck->lk.polls))); 2451 lck->lk.cleanup_ticket = 0; 2452 lck->lk.old_polls = NULL; 2453 lck->lk.next_ticket = 0; 2454 lck->lk.now_serving = 0; 2455 lck->lk.owner_id = 0; // no thread owns the lock. 2456 lck->lk.depth_locked = -1; // >= 0 for nestable locks, -1 for simple locks. 2457 lck->lk.initialized = lck; 2458 2459 KA_TRACE(1000, ("__kmp_init_drdpa_lock: lock %p initialized\n", lck)); 2460 } 2461 2462 void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck) { 2463 lck->lk.initialized = NULL; 2464 lck->lk.location = NULL; 2465 if (lck->lk.polls.load() != NULL) { 2466 __kmp_free(lck->lk.polls.load()); 2467 lck->lk.polls = NULL; 2468 } 2469 if (lck->lk.old_polls != NULL) { 2470 __kmp_free(lck->lk.old_polls); 2471 lck->lk.old_polls = NULL; 2472 } 2473 lck->lk.mask = 0; 2474 lck->lk.num_polls = 0; 2475 lck->lk.cleanup_ticket = 0; 2476 lck->lk.next_ticket = 0; 2477 lck->lk.now_serving = 0; 2478 lck->lk.owner_id = 0; 2479 lck->lk.depth_locked = -1; 2480 } 2481 2482 static void __kmp_destroy_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) { 2483 char const *const func = "omp_destroy_lock"; 2484 if (lck->lk.initialized != lck) { 2485 KMP_FATAL(LockIsUninitialized, func); 2486 } 2487 if (__kmp_is_drdpa_lock_nestable(lck)) { 2488 KMP_FATAL(LockNestableUsedAsSimple, func); 2489 } 2490 if (__kmp_get_drdpa_lock_owner(lck) != -1) { 2491 KMP_FATAL(LockStillOwned, func); 2492 } 2493 __kmp_destroy_drdpa_lock(lck); 2494 } 2495 2496 // nested drdpa ticket locks 2497 2498 int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2499 KMP_DEBUG_ASSERT(gtid >= 0); 2500 2501 if (__kmp_get_drdpa_lock_owner(lck) == gtid) { 2502 lck->lk.depth_locked += 1; 2503 return KMP_LOCK_ACQUIRED_NEXT; 2504 } else { 2505 __kmp_acquire_drdpa_lock_timed_template(lck, gtid); 2506 KMP_MB(); 2507 lck->lk.depth_locked = 1; 2508 KMP_MB(); 2509 lck->lk.owner_id = gtid + 1; 2510 return KMP_LOCK_ACQUIRED_FIRST; 2511 } 2512 } 2513 2514 static void __kmp_acquire_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck, 2515 kmp_int32 gtid) { 2516 char const *const func = "omp_set_nest_lock"; 2517 if (lck->lk.initialized != lck) { 2518 KMP_FATAL(LockIsUninitialized, func); 2519 } 2520 if (!__kmp_is_drdpa_lock_nestable(lck)) { 2521 KMP_FATAL(LockSimpleUsedAsNestable, func); 2522 } 2523 __kmp_acquire_nested_drdpa_lock(lck, gtid); 2524 } 2525 2526 int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2527 int retval; 2528 2529 KMP_DEBUG_ASSERT(gtid >= 0); 2530 2531 if (__kmp_get_drdpa_lock_owner(lck) == gtid) { 2532 retval = ++lck->lk.depth_locked; 2533 } else if (!__kmp_test_drdpa_lock(lck, gtid)) { 2534 retval = 0; 2535 } else { 2536 KMP_MB(); 2537 retval = lck->lk.depth_locked = 1; 2538 KMP_MB(); 2539 lck->lk.owner_id = gtid + 1; 2540 } 2541 return retval; 2542 } 2543 2544 static int __kmp_test_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck, 2545 kmp_int32 gtid) { 2546 char const *const func = "omp_test_nest_lock"; 2547 if (lck->lk.initialized != lck) { 2548 KMP_FATAL(LockIsUninitialized, func); 2549 } 2550 if (!__kmp_is_drdpa_lock_nestable(lck)) { 2551 KMP_FATAL(LockSimpleUsedAsNestable, func); 2552 } 2553 return __kmp_test_nested_drdpa_lock(lck, gtid); 2554 } 2555 2556 int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid) { 2557 KMP_DEBUG_ASSERT(gtid >= 0); 2558 2559 KMP_MB(); 2560 if (--(lck->lk.depth_locked) == 0) { 2561 KMP_MB(); 2562 lck->lk.owner_id = 0; 2563 __kmp_release_drdpa_lock(lck, gtid); 2564 return KMP_LOCK_RELEASED; 2565 } 2566 return KMP_LOCK_STILL_HELD; 2567 } 2568 2569 static int __kmp_release_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck, 2570 kmp_int32 gtid) { 2571 char const *const func = "omp_unset_nest_lock"; 2572 KMP_MB(); /* in case another processor initialized lock */ 2573 if (lck->lk.initialized != lck) { 2574 KMP_FATAL(LockIsUninitialized, func); 2575 } 2576 if (!__kmp_is_drdpa_lock_nestable(lck)) { 2577 KMP_FATAL(LockSimpleUsedAsNestable, func); 2578 } 2579 if (__kmp_get_drdpa_lock_owner(lck) == -1) { 2580 KMP_FATAL(LockUnsettingFree, func); 2581 } 2582 if (__kmp_get_drdpa_lock_owner(lck) != gtid) { 2583 KMP_FATAL(LockUnsettingSetByAnother, func); 2584 } 2585 return __kmp_release_nested_drdpa_lock(lck, gtid); 2586 } 2587 2588 void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck) { 2589 __kmp_init_drdpa_lock(lck); 2590 lck->lk.depth_locked = 0; // >= 0 for nestable locks, -1 for simple locks 2591 } 2592 2593 void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck) { 2594 __kmp_destroy_drdpa_lock(lck); 2595 lck->lk.depth_locked = 0; 2596 } 2597 2598 static void __kmp_destroy_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) { 2599 char const *const func = "omp_destroy_nest_lock"; 2600 if (lck->lk.initialized != lck) { 2601 KMP_FATAL(LockIsUninitialized, func); 2602 } 2603 if (!__kmp_is_drdpa_lock_nestable(lck)) { 2604 KMP_FATAL(LockSimpleUsedAsNestable, func); 2605 } 2606 if (__kmp_get_drdpa_lock_owner(lck) != -1) { 2607 KMP_FATAL(LockStillOwned, func); 2608 } 2609 __kmp_destroy_nested_drdpa_lock(lck); 2610 } 2611 2612 // access functions to fields which don't exist for all lock kinds. 2613 2614 static const ident_t *__kmp_get_drdpa_lock_location(kmp_drdpa_lock_t *lck) { 2615 return lck->lk.location; 2616 } 2617 2618 static void __kmp_set_drdpa_lock_location(kmp_drdpa_lock_t *lck, 2619 const ident_t *loc) { 2620 lck->lk.location = loc; 2621 } 2622 2623 static kmp_lock_flags_t __kmp_get_drdpa_lock_flags(kmp_drdpa_lock_t *lck) { 2624 return lck->lk.flags; 2625 } 2626 2627 static void __kmp_set_drdpa_lock_flags(kmp_drdpa_lock_t *lck, 2628 kmp_lock_flags_t flags) { 2629 lck->lk.flags = flags; 2630 } 2631 2632 // Time stamp counter 2633 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2634 #define __kmp_tsc() __kmp_hardware_timestamp() 2635 // Runtime's default backoff parameters 2636 kmp_backoff_t __kmp_spin_backoff_params = {1, 4096, 100}; 2637 #else 2638 // Use nanoseconds for other platforms 2639 extern kmp_uint64 __kmp_now_nsec(); 2640 kmp_backoff_t __kmp_spin_backoff_params = {1, 256, 100}; 2641 #define __kmp_tsc() __kmp_now_nsec() 2642 #endif 2643 2644 // A useful predicate for dealing with timestamps that may wrap. 2645 // Is a before b? Since the timestamps may wrap, this is asking whether it's 2646 // shorter to go clockwise from a to b around the clock-face, or anti-clockwise. 2647 // Times where going clockwise is less distance than going anti-clockwise 2648 // are in the future, others are in the past. e.g. a = MAX-1, b = MAX+1 (=0), 2649 // then a > b (true) does not mean a reached b; whereas signed(a) = -2, 2650 // signed(b) = 0 captures the actual difference 2651 static inline bool before(kmp_uint64 a, kmp_uint64 b) { 2652 return ((kmp_int64)b - (kmp_int64)a) > 0; 2653 } 2654 2655 // Truncated binary exponential backoff function 2656 void __kmp_spin_backoff(kmp_backoff_t *boff) { 2657 // We could flatten this loop, but making it a nested loop gives better result 2658 kmp_uint32 i; 2659 for (i = boff->step; i > 0; i--) { 2660 kmp_uint64 goal = __kmp_tsc() + boff->min_tick; 2661 do { 2662 KMP_CPU_PAUSE(); 2663 } while (before(__kmp_tsc(), goal)); 2664 } 2665 boff->step = (boff->step << 1 | 1) & (boff->max_backoff - 1); 2666 } 2667 2668 #if KMP_USE_DYNAMIC_LOCK 2669 2670 // Direct lock initializers. It simply writes a tag to the low 8 bits of the 2671 // lock word. 2672 static void __kmp_init_direct_lock(kmp_dyna_lock_t *lck, 2673 kmp_dyna_lockseq_t seq) { 2674 TCW_4(*lck, KMP_GET_D_TAG(seq)); 2675 KA_TRACE( 2676 20, 2677 ("__kmp_init_direct_lock: initialized direct lock with type#%d\n", seq)); 2678 } 2679 2680 #if KMP_USE_TSX 2681 2682 // HLE lock functions - imported from the testbed runtime. 2683 #define HLE_ACQUIRE ".byte 0xf2;" 2684 #define HLE_RELEASE ".byte 0xf3;" 2685 2686 static inline kmp_uint32 swap4(kmp_uint32 volatile *p, kmp_uint32 v) { 2687 __asm__ volatile(HLE_ACQUIRE "xchg %1,%0" : "+r"(v), "+m"(*p) : : "memory"); 2688 return v; 2689 } 2690 2691 static void __kmp_destroy_hle_lock(kmp_dyna_lock_t *lck) { TCW_4(*lck, 0); } 2692 2693 static void __kmp_destroy_hle_lock_with_checks(kmp_dyna_lock_t *lck) { 2694 TCW_4(*lck, 0); 2695 } 2696 2697 static void __kmp_acquire_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) { 2698 // Use gtid for KMP_LOCK_BUSY if necessary 2699 if (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)) { 2700 int delay = 1; 2701 do { 2702 while (*(kmp_uint32 volatile *)lck != KMP_LOCK_FREE(hle)) { 2703 for (int i = delay; i != 0; --i) 2704 KMP_CPU_PAUSE(); 2705 delay = ((delay << 1) | 1) & 7; 2706 } 2707 } while (swap4(lck, KMP_LOCK_BUSY(1, hle)) != KMP_LOCK_FREE(hle)); 2708 } 2709 } 2710 2711 static void __kmp_acquire_hle_lock_with_checks(kmp_dyna_lock_t *lck, 2712 kmp_int32 gtid) { 2713 __kmp_acquire_hle_lock(lck, gtid); // TODO: add checks 2714 } 2715 2716 static int __kmp_release_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) { 2717 __asm__ volatile(HLE_RELEASE "movl %1,%0" 2718 : "=m"(*lck) 2719 : "r"(KMP_LOCK_FREE(hle)) 2720 : "memory"); 2721 return KMP_LOCK_RELEASED; 2722 } 2723 2724 static int __kmp_release_hle_lock_with_checks(kmp_dyna_lock_t *lck, 2725 kmp_int32 gtid) { 2726 return __kmp_release_hle_lock(lck, gtid); // TODO: add checks 2727 } 2728 2729 static int __kmp_test_hle_lock(kmp_dyna_lock_t *lck, kmp_int32 gtid) { 2730 return swap4(lck, KMP_LOCK_BUSY(1, hle)) == KMP_LOCK_FREE(hle); 2731 } 2732 2733 static int __kmp_test_hle_lock_with_checks(kmp_dyna_lock_t *lck, 2734 kmp_int32 gtid) { 2735 return __kmp_test_hle_lock(lck, gtid); // TODO: add checks 2736 } 2737 2738 static void __kmp_init_rtm_queuing_lock(kmp_queuing_lock_t *lck) { 2739 __kmp_init_queuing_lock(lck); 2740 } 2741 2742 static void __kmp_destroy_rtm_queuing_lock(kmp_queuing_lock_t *lck) { 2743 __kmp_destroy_queuing_lock(lck); 2744 } 2745 2746 static void 2747 __kmp_destroy_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck) { 2748 __kmp_destroy_queuing_lock_with_checks(lck); 2749 } 2750 2751 KMP_ATTRIBUTE_TARGET_RTM 2752 static void __kmp_acquire_rtm_queuing_lock(kmp_queuing_lock_t *lck, 2753 kmp_int32 gtid) { 2754 unsigned retries = 3, status; 2755 do { 2756 status = _xbegin(); 2757 if (status == _XBEGIN_STARTED) { 2758 if (__kmp_is_unlocked_queuing_lock(lck)) 2759 return; 2760 _xabort(0xff); 2761 } 2762 if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) { 2763 // Wait until lock becomes free 2764 while (!__kmp_is_unlocked_queuing_lock(lck)) { 2765 KMP_YIELD(TRUE); 2766 } 2767 } else if (!(status & _XABORT_RETRY)) 2768 break; 2769 } while (retries--); 2770 2771 // Fall-back non-speculative lock (xchg) 2772 __kmp_acquire_queuing_lock(lck, gtid); 2773 } 2774 2775 static void __kmp_acquire_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 2776 kmp_int32 gtid) { 2777 __kmp_acquire_rtm_queuing_lock(lck, gtid); 2778 } 2779 2780 KMP_ATTRIBUTE_TARGET_RTM 2781 static int __kmp_release_rtm_queuing_lock(kmp_queuing_lock_t *lck, 2782 kmp_int32 gtid) { 2783 if (__kmp_is_unlocked_queuing_lock(lck)) { 2784 // Releasing from speculation 2785 _xend(); 2786 } else { 2787 // Releasing from a real lock 2788 __kmp_release_queuing_lock(lck, gtid); 2789 } 2790 return KMP_LOCK_RELEASED; 2791 } 2792 2793 static int __kmp_release_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 2794 kmp_int32 gtid) { 2795 return __kmp_release_rtm_queuing_lock(lck, gtid); 2796 } 2797 2798 KMP_ATTRIBUTE_TARGET_RTM 2799 static int __kmp_test_rtm_queuing_lock(kmp_queuing_lock_t *lck, 2800 kmp_int32 gtid) { 2801 unsigned retries = 3, status; 2802 do { 2803 status = _xbegin(); 2804 if (status == _XBEGIN_STARTED && __kmp_is_unlocked_queuing_lock(lck)) { 2805 return 1; 2806 } 2807 if (!(status & _XABORT_RETRY)) 2808 break; 2809 } while (retries--); 2810 2811 return __kmp_test_queuing_lock(lck, gtid); 2812 } 2813 2814 static int __kmp_test_rtm_queuing_lock_with_checks(kmp_queuing_lock_t *lck, 2815 kmp_int32 gtid) { 2816 return __kmp_test_rtm_queuing_lock(lck, gtid); 2817 } 2818 2819 // Reuse kmp_tas_lock_t for TSX lock which use RTM with fall-back spin lock. 2820 typedef kmp_tas_lock_t kmp_rtm_spin_lock_t; 2821 2822 static void __kmp_destroy_rtm_spin_lock(kmp_rtm_spin_lock_t *lck) { 2823 KMP_ATOMIC_ST_REL(&lck->lk.poll, 0); 2824 } 2825 2826 static void __kmp_destroy_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck) { 2827 __kmp_destroy_rtm_spin_lock(lck); 2828 } 2829 2830 KMP_ATTRIBUTE_TARGET_RTM 2831 static int __kmp_acquire_rtm_spin_lock(kmp_rtm_spin_lock_t *lck, 2832 kmp_int32 gtid) { 2833 unsigned retries = 3, status; 2834 kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin); 2835 kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin); 2836 do { 2837 status = _xbegin(); 2838 if (status == _XBEGIN_STARTED) { 2839 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free) 2840 return KMP_LOCK_ACQUIRED_FIRST; 2841 _xabort(0xff); 2842 } 2843 if ((status & _XABORT_EXPLICIT) && _XABORT_CODE(status) == 0xff) { 2844 // Wait until lock becomes free 2845 while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free) { 2846 KMP_YIELD(TRUE); 2847 } 2848 } else if (!(status & _XABORT_RETRY)) 2849 break; 2850 } while (retries--); 2851 2852 // Fall-back spin lock 2853 KMP_FSYNC_PREPARE(lck); 2854 kmp_backoff_t backoff = __kmp_spin_backoff_params; 2855 while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != lock_free || 2856 !__kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) { 2857 __kmp_spin_backoff(&backoff); 2858 } 2859 KMP_FSYNC_ACQUIRED(lck); 2860 return KMP_LOCK_ACQUIRED_FIRST; 2861 } 2862 2863 static int __kmp_acquire_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck, 2864 kmp_int32 gtid) { 2865 return __kmp_acquire_rtm_spin_lock(lck, gtid); 2866 } 2867 2868 KMP_ATTRIBUTE_TARGET_RTM 2869 static int __kmp_release_rtm_spin_lock(kmp_rtm_spin_lock_t *lck, 2870 kmp_int32 gtid) { 2871 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == KMP_LOCK_FREE(rtm_spin)) { 2872 // Releasing from speculation 2873 _xend(); 2874 } else { 2875 // Releasing from a real lock 2876 KMP_FSYNC_RELEASING(lck); 2877 KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(rtm_spin)); 2878 } 2879 return KMP_LOCK_RELEASED; 2880 } 2881 2882 static int __kmp_release_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck, 2883 kmp_int32 gtid) { 2884 return __kmp_release_rtm_spin_lock(lck, gtid); 2885 } 2886 2887 KMP_ATTRIBUTE_TARGET_RTM 2888 static int __kmp_test_rtm_spin_lock(kmp_rtm_spin_lock_t *lck, kmp_int32 gtid) { 2889 unsigned retries = 3, status; 2890 kmp_int32 lock_free = KMP_LOCK_FREE(rtm_spin); 2891 kmp_int32 lock_busy = KMP_LOCK_BUSY(1, rtm_spin); 2892 do { 2893 status = _xbegin(); 2894 if (status == _XBEGIN_STARTED && 2895 KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free) { 2896 return TRUE; 2897 } 2898 if (!(status & _XABORT_RETRY)) 2899 break; 2900 } while (retries--); 2901 2902 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == lock_free && 2903 __kmp_atomic_compare_store_acq(&lck->lk.poll, lock_free, lock_busy)) { 2904 KMP_FSYNC_ACQUIRED(lck); 2905 return TRUE; 2906 } 2907 return FALSE; 2908 } 2909 2910 static int __kmp_test_rtm_spin_lock_with_checks(kmp_rtm_spin_lock_t *lck, 2911 kmp_int32 gtid) { 2912 return __kmp_test_rtm_spin_lock(lck, gtid); 2913 } 2914 2915 #endif // KMP_USE_TSX 2916 2917 // Entry functions for indirect locks (first element of direct lock jump tables) 2918 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *l, 2919 kmp_dyna_lockseq_t tag); 2920 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock); 2921 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32); 2922 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32); 2923 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32); 2924 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock, 2925 kmp_int32); 2926 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock, 2927 kmp_int32); 2928 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock, 2929 kmp_int32); 2930 2931 // Lock function definitions for the union parameter type 2932 #define KMP_FOREACH_LOCK_KIND(m, a) m(ticket, a) m(queuing, a) m(drdpa, a) 2933 2934 #define expand1(lk, op) \ 2935 static void __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock) { \ 2936 __kmp_##op##_##lk##_##lock(&lock->lk); \ 2937 } 2938 #define expand2(lk, op) \ 2939 static int __kmp_##op##_##lk##_##lock(kmp_user_lock_p lock, \ 2940 kmp_int32 gtid) { \ 2941 return __kmp_##op##_##lk##_##lock(&lock->lk, gtid); \ 2942 } 2943 #define expand3(lk, op) \ 2944 static void __kmp_set_##lk##_##lock_flags(kmp_user_lock_p lock, \ 2945 kmp_lock_flags_t flags) { \ 2946 __kmp_set_##lk##_lock_flags(&lock->lk, flags); \ 2947 } 2948 #define expand4(lk, op) \ 2949 static void __kmp_set_##lk##_##lock_location(kmp_user_lock_p lock, \ 2950 const ident_t *loc) { \ 2951 __kmp_set_##lk##_lock_location(&lock->lk, loc); \ 2952 } 2953 2954 KMP_FOREACH_LOCK_KIND(expand1, init) 2955 KMP_FOREACH_LOCK_KIND(expand1, init_nested) 2956 KMP_FOREACH_LOCK_KIND(expand1, destroy) 2957 KMP_FOREACH_LOCK_KIND(expand1, destroy_nested) 2958 KMP_FOREACH_LOCK_KIND(expand2, acquire) 2959 KMP_FOREACH_LOCK_KIND(expand2, acquire_nested) 2960 KMP_FOREACH_LOCK_KIND(expand2, release) 2961 KMP_FOREACH_LOCK_KIND(expand2, release_nested) 2962 KMP_FOREACH_LOCK_KIND(expand2, test) 2963 KMP_FOREACH_LOCK_KIND(expand2, test_nested) 2964 KMP_FOREACH_LOCK_KIND(expand3, ) 2965 KMP_FOREACH_LOCK_KIND(expand4, ) 2966 2967 #undef expand1 2968 #undef expand2 2969 #undef expand3 2970 #undef expand4 2971 2972 // Jump tables for the indirect lock functions 2973 // Only fill in the odd entries, that avoids the need to shift out the low bit 2974 2975 // init functions 2976 #define expand(l, op) 0, __kmp_init_direct_lock, 2977 void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t) = { 2978 __kmp_init_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, init)}; 2979 #undef expand 2980 2981 // destroy functions 2982 #define expand(l, op) 0, (void (*)(kmp_dyna_lock_t *))__kmp_##op##_##l##_lock, 2983 static void (*direct_destroy[])(kmp_dyna_lock_t *) = { 2984 __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)}; 2985 #undef expand 2986 #define expand(l, op) \ 2987 0, (void (*)(kmp_dyna_lock_t *))__kmp_destroy_##l##_lock_with_checks, 2988 static void (*direct_destroy_check[])(kmp_dyna_lock_t *) = { 2989 __kmp_destroy_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, destroy)}; 2990 #undef expand 2991 2992 // set/acquire functions 2993 #define expand(l, op) \ 2994 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock, 2995 static int (*direct_set[])(kmp_dyna_lock_t *, kmp_int32) = { 2996 __kmp_set_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, acquire)}; 2997 #undef expand 2998 #define expand(l, op) \ 2999 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks, 3000 static int (*direct_set_check[])(kmp_dyna_lock_t *, kmp_int32) = { 3001 __kmp_set_indirect_lock_with_checks, 0, 3002 KMP_FOREACH_D_LOCK(expand, acquire)}; 3003 #undef expand 3004 3005 // unset/release and test functions 3006 #define expand(l, op) \ 3007 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock, 3008 static int (*direct_unset[])(kmp_dyna_lock_t *, kmp_int32) = { 3009 __kmp_unset_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, release)}; 3010 static int (*direct_test[])(kmp_dyna_lock_t *, kmp_int32) = { 3011 __kmp_test_indirect_lock, 0, KMP_FOREACH_D_LOCK(expand, test)}; 3012 #undef expand 3013 #define expand(l, op) \ 3014 0, (int (*)(kmp_dyna_lock_t *, kmp_int32))__kmp_##op##_##l##_lock_with_checks, 3015 static int (*direct_unset_check[])(kmp_dyna_lock_t *, kmp_int32) = { 3016 __kmp_unset_indirect_lock_with_checks, 0, 3017 KMP_FOREACH_D_LOCK(expand, release)}; 3018 static int (*direct_test_check[])(kmp_dyna_lock_t *, kmp_int32) = { 3019 __kmp_test_indirect_lock_with_checks, 0, KMP_FOREACH_D_LOCK(expand, test)}; 3020 #undef expand 3021 3022 // Exposes only one set of jump tables (*lock or *lock_with_checks). 3023 void (**__kmp_direct_destroy)(kmp_dyna_lock_t *) = 0; 3024 int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32) = 0; 3025 int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32) = 0; 3026 int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32) = 0; 3027 3028 // Jump tables for the indirect lock functions 3029 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock, 3030 void (*__kmp_indirect_init[])(kmp_user_lock_p) = { 3031 KMP_FOREACH_I_LOCK(expand, init)}; 3032 #undef expand 3033 3034 #define expand(l, op) (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock, 3035 static void (*indirect_destroy[])(kmp_user_lock_p) = { 3036 KMP_FOREACH_I_LOCK(expand, destroy)}; 3037 #undef expand 3038 #define expand(l, op) \ 3039 (void (*)(kmp_user_lock_p)) __kmp_##op##_##l##_##lock_with_checks, 3040 static void (*indirect_destroy_check[])(kmp_user_lock_p) = { 3041 KMP_FOREACH_I_LOCK(expand, destroy)}; 3042 #undef expand 3043 3044 // set/acquire functions 3045 #define expand(l, op) \ 3046 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock, 3047 static int (*indirect_set[])(kmp_user_lock_p, 3048 kmp_int32) = {KMP_FOREACH_I_LOCK(expand, acquire)}; 3049 #undef expand 3050 #define expand(l, op) \ 3051 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks, 3052 static int (*indirect_set_check[])(kmp_user_lock_p, kmp_int32) = { 3053 KMP_FOREACH_I_LOCK(expand, acquire)}; 3054 #undef expand 3055 3056 // unset/release and test functions 3057 #define expand(l, op) \ 3058 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock, 3059 static int (*indirect_unset[])(kmp_user_lock_p, kmp_int32) = { 3060 KMP_FOREACH_I_LOCK(expand, release)}; 3061 static int (*indirect_test[])(kmp_user_lock_p, 3062 kmp_int32) = {KMP_FOREACH_I_LOCK(expand, test)}; 3063 #undef expand 3064 #define expand(l, op) \ 3065 (int (*)(kmp_user_lock_p, kmp_int32)) __kmp_##op##_##l##_##lock_with_checks, 3066 static int (*indirect_unset_check[])(kmp_user_lock_p, kmp_int32) = { 3067 KMP_FOREACH_I_LOCK(expand, release)}; 3068 static int (*indirect_test_check[])(kmp_user_lock_p, kmp_int32) = { 3069 KMP_FOREACH_I_LOCK(expand, test)}; 3070 #undef expand 3071 3072 // Exposes only one jump tables (*lock or *lock_with_checks). 3073 void (**__kmp_indirect_destroy)(kmp_user_lock_p) = 0; 3074 int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32) = 0; 3075 int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32) = 0; 3076 int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32) = 0; 3077 3078 // Lock index table. 3079 kmp_indirect_lock_table_t __kmp_i_lock_table; 3080 3081 // Size of indirect locks. 3082 static kmp_uint32 __kmp_indirect_lock_size[KMP_NUM_I_LOCKS] = {0}; 3083 3084 // Jump tables for lock accessor/modifier. 3085 void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p, 3086 const ident_t *) = {0}; 3087 void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p, 3088 kmp_lock_flags_t) = {0}; 3089 const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])( 3090 kmp_user_lock_p) = {0}; 3091 kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])( 3092 kmp_user_lock_p) = {0}; 3093 3094 // Use different lock pools for different lock types. 3095 static kmp_indirect_lock_t *__kmp_indirect_lock_pool[KMP_NUM_I_LOCKS] = {0}; 3096 3097 // User lock allocator for dynamically dispatched indirect locks. Every entry of 3098 // the indirect lock table holds the address and type of the allocated indirect 3099 // lock (kmp_indirect_lock_t), and the size of the table doubles when it is 3100 // full. A destroyed indirect lock object is returned to the reusable pool of 3101 // locks, unique to each lock type. 3102 kmp_indirect_lock_t *__kmp_allocate_indirect_lock(void **user_lock, 3103 kmp_int32 gtid, 3104 kmp_indirect_locktag_t tag) { 3105 kmp_indirect_lock_t *lck; 3106 kmp_lock_index_t idx; 3107 3108 __kmp_acquire_lock(&__kmp_global_lock, gtid); 3109 3110 if (__kmp_indirect_lock_pool[tag] != NULL) { 3111 // Reuse the allocated and destroyed lock object 3112 lck = __kmp_indirect_lock_pool[tag]; 3113 if (OMP_LOCK_T_SIZE < sizeof(void *)) 3114 idx = lck->lock->pool.index; 3115 __kmp_indirect_lock_pool[tag] = (kmp_indirect_lock_t *)lck->lock->pool.next; 3116 KA_TRACE(20, ("__kmp_allocate_indirect_lock: reusing an existing lock %p\n", 3117 lck)); 3118 } else { 3119 idx = __kmp_i_lock_table.next; 3120 // Check capacity and double the size if it is full 3121 if (idx == __kmp_i_lock_table.size) { 3122 // Double up the space for block pointers 3123 int row = __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; 3124 kmp_indirect_lock_t **new_table = (kmp_indirect_lock_t **)__kmp_allocate( 3125 2 * row * sizeof(kmp_indirect_lock_t *)); 3126 KMP_MEMCPY(new_table, __kmp_i_lock_table.table, 3127 row * sizeof(kmp_indirect_lock_t *)); 3128 kmp_indirect_lock_t **old_table = __kmp_i_lock_table.table; 3129 __kmp_i_lock_table.table = new_table; 3130 __kmp_free(old_table); 3131 // Allocate new objects in the new blocks 3132 for (int i = row; i < 2 * row; ++i) 3133 *(__kmp_i_lock_table.table + i) = (kmp_indirect_lock_t *)__kmp_allocate( 3134 KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t)); 3135 __kmp_i_lock_table.size = 2 * idx; 3136 } 3137 __kmp_i_lock_table.next++; 3138 lck = KMP_GET_I_LOCK(idx); 3139 // Allocate a new base lock object 3140 lck->lock = (kmp_user_lock_p)__kmp_allocate(__kmp_indirect_lock_size[tag]); 3141 KA_TRACE(20, 3142 ("__kmp_allocate_indirect_lock: allocated a new lock %p\n", lck)); 3143 } 3144 3145 __kmp_release_lock(&__kmp_global_lock, gtid); 3146 3147 lck->type = tag; 3148 3149 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3150 *((kmp_lock_index_t *)user_lock) = idx 3151 << 1; // indirect lock word must be even 3152 } else { 3153 *((kmp_indirect_lock_t **)user_lock) = lck; 3154 } 3155 3156 return lck; 3157 } 3158 3159 // User lock lookup for dynamically dispatched locks. 3160 static __forceinline kmp_indirect_lock_t * 3161 __kmp_lookup_indirect_lock(void **user_lock, const char *func) { 3162 if (__kmp_env_consistency_check) { 3163 kmp_indirect_lock_t *lck = NULL; 3164 if (user_lock == NULL) { 3165 KMP_FATAL(LockIsUninitialized, func); 3166 } 3167 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3168 kmp_lock_index_t idx = KMP_EXTRACT_I_INDEX(user_lock); 3169 if (idx >= __kmp_i_lock_table.size) { 3170 KMP_FATAL(LockIsUninitialized, func); 3171 } 3172 lck = KMP_GET_I_LOCK(idx); 3173 } else { 3174 lck = *((kmp_indirect_lock_t **)user_lock); 3175 } 3176 if (lck == NULL) { 3177 KMP_FATAL(LockIsUninitialized, func); 3178 } 3179 return lck; 3180 } else { 3181 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3182 return KMP_GET_I_LOCK(KMP_EXTRACT_I_INDEX(user_lock)); 3183 } else { 3184 return *((kmp_indirect_lock_t **)user_lock); 3185 } 3186 } 3187 } 3188 3189 static void __kmp_init_indirect_lock(kmp_dyna_lock_t *lock, 3190 kmp_dyna_lockseq_t seq) { 3191 #if KMP_USE_ADAPTIVE_LOCKS 3192 if (seq == lockseq_adaptive && !__kmp_cpuinfo.rtm) { 3193 KMP_WARNING(AdaptiveNotSupported, "kmp_lockseq_t", "adaptive"); 3194 seq = lockseq_queuing; 3195 } 3196 #endif 3197 #if KMP_USE_TSX 3198 if (seq == lockseq_rtm_queuing && !__kmp_cpuinfo.rtm) { 3199 seq = lockseq_queuing; 3200 } 3201 #endif 3202 kmp_indirect_locktag_t tag = KMP_GET_I_TAG(seq); 3203 kmp_indirect_lock_t *l = 3204 __kmp_allocate_indirect_lock((void **)lock, __kmp_entry_gtid(), tag); 3205 KMP_I_LOCK_FUNC(l, init)(l->lock); 3206 KA_TRACE( 3207 20, ("__kmp_init_indirect_lock: initialized indirect lock with type#%d\n", 3208 seq)); 3209 } 3210 3211 static void __kmp_destroy_indirect_lock(kmp_dyna_lock_t *lock) { 3212 kmp_uint32 gtid = __kmp_entry_gtid(); 3213 kmp_indirect_lock_t *l = 3214 __kmp_lookup_indirect_lock((void **)lock, "omp_destroy_lock"); 3215 KMP_I_LOCK_FUNC(l, destroy)(l->lock); 3216 kmp_indirect_locktag_t tag = l->type; 3217 3218 __kmp_acquire_lock(&__kmp_global_lock, gtid); 3219 3220 // Use the base lock's space to keep the pool chain. 3221 l->lock->pool.next = (kmp_user_lock_p)__kmp_indirect_lock_pool[tag]; 3222 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3223 l->lock->pool.index = KMP_EXTRACT_I_INDEX(lock); 3224 } 3225 __kmp_indirect_lock_pool[tag] = l; 3226 3227 __kmp_release_lock(&__kmp_global_lock, gtid); 3228 } 3229 3230 static int __kmp_set_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) { 3231 kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock); 3232 return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid); 3233 } 3234 3235 static int __kmp_unset_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) { 3236 kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock); 3237 return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid); 3238 } 3239 3240 static int __kmp_test_indirect_lock(kmp_dyna_lock_t *lock, kmp_int32 gtid) { 3241 kmp_indirect_lock_t *l = KMP_LOOKUP_I_LOCK(lock); 3242 return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid); 3243 } 3244 3245 static int __kmp_set_indirect_lock_with_checks(kmp_dyna_lock_t *lock, 3246 kmp_int32 gtid) { 3247 kmp_indirect_lock_t *l = 3248 __kmp_lookup_indirect_lock((void **)lock, "omp_set_lock"); 3249 return KMP_I_LOCK_FUNC(l, set)(l->lock, gtid); 3250 } 3251 3252 static int __kmp_unset_indirect_lock_with_checks(kmp_dyna_lock_t *lock, 3253 kmp_int32 gtid) { 3254 kmp_indirect_lock_t *l = 3255 __kmp_lookup_indirect_lock((void **)lock, "omp_unset_lock"); 3256 return KMP_I_LOCK_FUNC(l, unset)(l->lock, gtid); 3257 } 3258 3259 static int __kmp_test_indirect_lock_with_checks(kmp_dyna_lock_t *lock, 3260 kmp_int32 gtid) { 3261 kmp_indirect_lock_t *l = 3262 __kmp_lookup_indirect_lock((void **)lock, "omp_test_lock"); 3263 return KMP_I_LOCK_FUNC(l, test)(l->lock, gtid); 3264 } 3265 3266 kmp_dyna_lockseq_t __kmp_user_lock_seq = lockseq_queuing; 3267 3268 // This is used only in kmp_error.cpp when consistency checking is on. 3269 kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck, kmp_uint32 seq) { 3270 switch (seq) { 3271 case lockseq_tas: 3272 case lockseq_nested_tas: 3273 return __kmp_get_tas_lock_owner((kmp_tas_lock_t *)lck); 3274 #if KMP_USE_FUTEX 3275 case lockseq_futex: 3276 case lockseq_nested_futex: 3277 return __kmp_get_futex_lock_owner((kmp_futex_lock_t *)lck); 3278 #endif 3279 case lockseq_ticket: 3280 case lockseq_nested_ticket: 3281 return __kmp_get_ticket_lock_owner((kmp_ticket_lock_t *)lck); 3282 case lockseq_queuing: 3283 case lockseq_nested_queuing: 3284 #if KMP_USE_ADAPTIVE_LOCKS 3285 case lockseq_adaptive: 3286 #endif 3287 return __kmp_get_queuing_lock_owner((kmp_queuing_lock_t *)lck); 3288 case lockseq_drdpa: 3289 case lockseq_nested_drdpa: 3290 return __kmp_get_drdpa_lock_owner((kmp_drdpa_lock_t *)lck); 3291 default: 3292 return 0; 3293 } 3294 } 3295 3296 // Initializes data for dynamic user locks. 3297 void __kmp_init_dynamic_user_locks() { 3298 // Initialize jump table for the lock functions 3299 if (__kmp_env_consistency_check) { 3300 __kmp_direct_set = direct_set_check; 3301 __kmp_direct_unset = direct_unset_check; 3302 __kmp_direct_test = direct_test_check; 3303 __kmp_direct_destroy = direct_destroy_check; 3304 __kmp_indirect_set = indirect_set_check; 3305 __kmp_indirect_unset = indirect_unset_check; 3306 __kmp_indirect_test = indirect_test_check; 3307 __kmp_indirect_destroy = indirect_destroy_check; 3308 } else { 3309 __kmp_direct_set = direct_set; 3310 __kmp_direct_unset = direct_unset; 3311 __kmp_direct_test = direct_test; 3312 __kmp_direct_destroy = direct_destroy; 3313 __kmp_indirect_set = indirect_set; 3314 __kmp_indirect_unset = indirect_unset; 3315 __kmp_indirect_test = indirect_test; 3316 __kmp_indirect_destroy = indirect_destroy; 3317 } 3318 // If the user locks have already been initialized, then return. Allow the 3319 // switch between different KMP_CONSISTENCY_CHECK values, but do not allocate 3320 // new lock tables if they have already been allocated. 3321 if (__kmp_init_user_locks) 3322 return; 3323 3324 // Initialize lock index table 3325 __kmp_i_lock_table.size = KMP_I_LOCK_CHUNK; 3326 __kmp_i_lock_table.table = 3327 (kmp_indirect_lock_t **)__kmp_allocate(sizeof(kmp_indirect_lock_t *)); 3328 *(__kmp_i_lock_table.table) = (kmp_indirect_lock_t *)__kmp_allocate( 3329 KMP_I_LOCK_CHUNK * sizeof(kmp_indirect_lock_t)); 3330 __kmp_i_lock_table.next = 0; 3331 3332 // Indirect lock size 3333 __kmp_indirect_lock_size[locktag_ticket] = sizeof(kmp_ticket_lock_t); 3334 __kmp_indirect_lock_size[locktag_queuing] = sizeof(kmp_queuing_lock_t); 3335 #if KMP_USE_ADAPTIVE_LOCKS 3336 __kmp_indirect_lock_size[locktag_adaptive] = sizeof(kmp_adaptive_lock_t); 3337 #endif 3338 __kmp_indirect_lock_size[locktag_drdpa] = sizeof(kmp_drdpa_lock_t); 3339 #if KMP_USE_TSX 3340 __kmp_indirect_lock_size[locktag_rtm_queuing] = sizeof(kmp_queuing_lock_t); 3341 #endif 3342 __kmp_indirect_lock_size[locktag_nested_tas] = sizeof(kmp_tas_lock_t); 3343 #if KMP_USE_FUTEX 3344 __kmp_indirect_lock_size[locktag_nested_futex] = sizeof(kmp_futex_lock_t); 3345 #endif 3346 __kmp_indirect_lock_size[locktag_nested_ticket] = sizeof(kmp_ticket_lock_t); 3347 __kmp_indirect_lock_size[locktag_nested_queuing] = sizeof(kmp_queuing_lock_t); 3348 __kmp_indirect_lock_size[locktag_nested_drdpa] = sizeof(kmp_drdpa_lock_t); 3349 3350 // Initialize lock accessor/modifier 3351 #define fill_jumps(table, expand, sep) \ 3352 { \ 3353 table[locktag##sep##ticket] = expand(ticket); \ 3354 table[locktag##sep##queuing] = expand(queuing); \ 3355 table[locktag##sep##drdpa] = expand(drdpa); \ 3356 } 3357 3358 #if KMP_USE_ADAPTIVE_LOCKS 3359 #define fill_table(table, expand) \ 3360 { \ 3361 fill_jumps(table, expand, _); \ 3362 table[locktag_adaptive] = expand(queuing); \ 3363 fill_jumps(table, expand, _nested_); \ 3364 } 3365 #else 3366 #define fill_table(table, expand) \ 3367 { \ 3368 fill_jumps(table, expand, _); \ 3369 fill_jumps(table, expand, _nested_); \ 3370 } 3371 #endif // KMP_USE_ADAPTIVE_LOCKS 3372 3373 #define expand(l) \ 3374 (void (*)(kmp_user_lock_p, const ident_t *)) __kmp_set_##l##_lock_location 3375 fill_table(__kmp_indirect_set_location, expand); 3376 #undef expand 3377 #define expand(l) \ 3378 (void (*)(kmp_user_lock_p, kmp_lock_flags_t)) __kmp_set_##l##_lock_flags 3379 fill_table(__kmp_indirect_set_flags, expand); 3380 #undef expand 3381 #define expand(l) \ 3382 (const ident_t *(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_location 3383 fill_table(__kmp_indirect_get_location, expand); 3384 #undef expand 3385 #define expand(l) \ 3386 (kmp_lock_flags_t(*)(kmp_user_lock_p)) __kmp_get_##l##_lock_flags 3387 fill_table(__kmp_indirect_get_flags, expand); 3388 #undef expand 3389 3390 __kmp_init_user_locks = TRUE; 3391 } 3392 3393 // Clean up the lock table. 3394 void __kmp_cleanup_indirect_user_locks() { 3395 kmp_lock_index_t i; 3396 int k; 3397 3398 // Clean up locks in the pools first (they were already destroyed before going 3399 // into the pools). 3400 for (k = 0; k < KMP_NUM_I_LOCKS; ++k) { 3401 kmp_indirect_lock_t *l = __kmp_indirect_lock_pool[k]; 3402 while (l != NULL) { 3403 kmp_indirect_lock_t *ll = l; 3404 l = (kmp_indirect_lock_t *)l->lock->pool.next; 3405 KA_TRACE(20, ("__kmp_cleanup_indirect_user_locks: freeing %p from pool\n", 3406 ll)); 3407 __kmp_free(ll->lock); 3408 ll->lock = NULL; 3409 } 3410 __kmp_indirect_lock_pool[k] = NULL; 3411 } 3412 // Clean up the remaining undestroyed locks. 3413 for (i = 0; i < __kmp_i_lock_table.next; i++) { 3414 kmp_indirect_lock_t *l = KMP_GET_I_LOCK(i); 3415 if (l->lock != NULL) { 3416 // Locks not destroyed explicitly need to be destroyed here. 3417 KMP_I_LOCK_FUNC(l, destroy)(l->lock); 3418 KA_TRACE( 3419 20, 3420 ("__kmp_cleanup_indirect_user_locks: destroy/freeing %p from table\n", 3421 l)); 3422 __kmp_free(l->lock); 3423 } 3424 } 3425 // Free the table 3426 for (i = 0; i < __kmp_i_lock_table.size / KMP_I_LOCK_CHUNK; i++) 3427 __kmp_free(__kmp_i_lock_table.table[i]); 3428 __kmp_free(__kmp_i_lock_table.table); 3429 3430 __kmp_init_user_locks = FALSE; 3431 } 3432 3433 enum kmp_lock_kind __kmp_user_lock_kind = lk_default; 3434 int __kmp_num_locks_in_block = 1; // FIXME - tune this value 3435 3436 #else // KMP_USE_DYNAMIC_LOCK 3437 3438 static void __kmp_init_tas_lock_with_checks(kmp_tas_lock_t *lck) { 3439 __kmp_init_tas_lock(lck); 3440 } 3441 3442 static void __kmp_init_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) { 3443 __kmp_init_nested_tas_lock(lck); 3444 } 3445 3446 #if KMP_USE_FUTEX 3447 static void __kmp_init_futex_lock_with_checks(kmp_futex_lock_t *lck) { 3448 __kmp_init_futex_lock(lck); 3449 } 3450 3451 static void __kmp_init_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) { 3452 __kmp_init_nested_futex_lock(lck); 3453 } 3454 #endif 3455 3456 static int __kmp_is_ticket_lock_initialized(kmp_ticket_lock_t *lck) { 3457 return lck == lck->lk.self; 3458 } 3459 3460 static void __kmp_init_ticket_lock_with_checks(kmp_ticket_lock_t *lck) { 3461 __kmp_init_ticket_lock(lck); 3462 } 3463 3464 static void __kmp_init_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) { 3465 __kmp_init_nested_ticket_lock(lck); 3466 } 3467 3468 static int __kmp_is_queuing_lock_initialized(kmp_queuing_lock_t *lck) { 3469 return lck == lck->lk.initialized; 3470 } 3471 3472 static void __kmp_init_queuing_lock_with_checks(kmp_queuing_lock_t *lck) { 3473 __kmp_init_queuing_lock(lck); 3474 } 3475 3476 static void 3477 __kmp_init_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) { 3478 __kmp_init_nested_queuing_lock(lck); 3479 } 3480 3481 #if KMP_USE_ADAPTIVE_LOCKS 3482 static void __kmp_init_adaptive_lock_with_checks(kmp_adaptive_lock_t *lck) { 3483 __kmp_init_adaptive_lock(lck); 3484 } 3485 #endif 3486 3487 static int __kmp_is_drdpa_lock_initialized(kmp_drdpa_lock_t *lck) { 3488 return lck == lck->lk.initialized; 3489 } 3490 3491 static void __kmp_init_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) { 3492 __kmp_init_drdpa_lock(lck); 3493 } 3494 3495 static void __kmp_init_nested_drdpa_lock_with_checks(kmp_drdpa_lock_t *lck) { 3496 __kmp_init_nested_drdpa_lock(lck); 3497 } 3498 3499 /* user locks 3500 * They are implemented as a table of function pointers which are set to the 3501 * lock functions of the appropriate kind, once that has been determined. */ 3502 3503 enum kmp_lock_kind __kmp_user_lock_kind = lk_default; 3504 3505 size_t __kmp_base_user_lock_size = 0; 3506 size_t __kmp_user_lock_size = 0; 3507 3508 kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck) = NULL; 3509 int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck, 3510 kmp_int32 gtid) = NULL; 3511 3512 int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck, 3513 kmp_int32 gtid) = NULL; 3514 int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck, 3515 kmp_int32 gtid) = NULL; 3516 void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL; 3517 void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck) = NULL; 3518 void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL; 3519 int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck, 3520 kmp_int32 gtid) = NULL; 3521 3522 int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck, 3523 kmp_int32 gtid) = NULL; 3524 int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck, 3525 kmp_int32 gtid) = NULL; 3526 void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL; 3527 void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck) = NULL; 3528 3529 int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck) = NULL; 3530 const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck) = NULL; 3531 void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck, 3532 const ident_t *loc) = NULL; 3533 kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck) = NULL; 3534 void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck, 3535 kmp_lock_flags_t flags) = NULL; 3536 3537 void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind) { 3538 switch (user_lock_kind) { 3539 case lk_default: 3540 default: 3541 KMP_ASSERT(0); 3542 3543 case lk_tas: { 3544 __kmp_base_user_lock_size = sizeof(kmp_base_tas_lock_t); 3545 __kmp_user_lock_size = sizeof(kmp_tas_lock_t); 3546 3547 __kmp_get_user_lock_owner_ = 3548 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_tas_lock_owner); 3549 3550 if (__kmp_env_consistency_check) { 3551 KMP_BIND_USER_LOCK_WITH_CHECKS(tas); 3552 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(tas); 3553 } else { 3554 KMP_BIND_USER_LOCK(tas); 3555 KMP_BIND_NESTED_USER_LOCK(tas); 3556 } 3557 3558 __kmp_destroy_user_lock_ = 3559 (void (*)(kmp_user_lock_p))(&__kmp_destroy_tas_lock); 3560 3561 __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL; 3562 3563 __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL; 3564 3565 __kmp_set_user_lock_location_ = 3566 (void (*)(kmp_user_lock_p, const ident_t *))NULL; 3567 3568 __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL; 3569 3570 __kmp_set_user_lock_flags_ = 3571 (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL; 3572 } break; 3573 3574 #if KMP_USE_FUTEX 3575 3576 case lk_futex: { 3577 __kmp_base_user_lock_size = sizeof(kmp_base_futex_lock_t); 3578 __kmp_user_lock_size = sizeof(kmp_futex_lock_t); 3579 3580 __kmp_get_user_lock_owner_ = 3581 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_futex_lock_owner); 3582 3583 if (__kmp_env_consistency_check) { 3584 KMP_BIND_USER_LOCK_WITH_CHECKS(futex); 3585 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(futex); 3586 } else { 3587 KMP_BIND_USER_LOCK(futex); 3588 KMP_BIND_NESTED_USER_LOCK(futex); 3589 } 3590 3591 __kmp_destroy_user_lock_ = 3592 (void (*)(kmp_user_lock_p))(&__kmp_destroy_futex_lock); 3593 3594 __kmp_is_user_lock_initialized_ = (int (*)(kmp_user_lock_p))NULL; 3595 3596 __kmp_get_user_lock_location_ = (const ident_t *(*)(kmp_user_lock_p))NULL; 3597 3598 __kmp_set_user_lock_location_ = 3599 (void (*)(kmp_user_lock_p, const ident_t *))NULL; 3600 3601 __kmp_get_user_lock_flags_ = (kmp_lock_flags_t(*)(kmp_user_lock_p))NULL; 3602 3603 __kmp_set_user_lock_flags_ = 3604 (void (*)(kmp_user_lock_p, kmp_lock_flags_t))NULL; 3605 } break; 3606 3607 #endif // KMP_USE_FUTEX 3608 3609 case lk_ticket: { 3610 __kmp_base_user_lock_size = sizeof(kmp_base_ticket_lock_t); 3611 __kmp_user_lock_size = sizeof(kmp_ticket_lock_t); 3612 3613 __kmp_get_user_lock_owner_ = 3614 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_owner); 3615 3616 if (__kmp_env_consistency_check) { 3617 KMP_BIND_USER_LOCK_WITH_CHECKS(ticket); 3618 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(ticket); 3619 } else { 3620 KMP_BIND_USER_LOCK(ticket); 3621 KMP_BIND_NESTED_USER_LOCK(ticket); 3622 } 3623 3624 __kmp_destroy_user_lock_ = 3625 (void (*)(kmp_user_lock_p))(&__kmp_destroy_ticket_lock); 3626 3627 __kmp_is_user_lock_initialized_ = 3628 (int (*)(kmp_user_lock_p))(&__kmp_is_ticket_lock_initialized); 3629 3630 __kmp_get_user_lock_location_ = 3631 (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_location); 3632 3633 __kmp_set_user_lock_location_ = (void (*)( 3634 kmp_user_lock_p, const ident_t *))(&__kmp_set_ticket_lock_location); 3635 3636 __kmp_get_user_lock_flags_ = 3637 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_ticket_lock_flags); 3638 3639 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))( 3640 &__kmp_set_ticket_lock_flags); 3641 } break; 3642 3643 case lk_queuing: { 3644 __kmp_base_user_lock_size = sizeof(kmp_base_queuing_lock_t); 3645 __kmp_user_lock_size = sizeof(kmp_queuing_lock_t); 3646 3647 __kmp_get_user_lock_owner_ = 3648 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner); 3649 3650 if (__kmp_env_consistency_check) { 3651 KMP_BIND_USER_LOCK_WITH_CHECKS(queuing); 3652 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(queuing); 3653 } else { 3654 KMP_BIND_USER_LOCK(queuing); 3655 KMP_BIND_NESTED_USER_LOCK(queuing); 3656 } 3657 3658 __kmp_destroy_user_lock_ = 3659 (void (*)(kmp_user_lock_p))(&__kmp_destroy_queuing_lock); 3660 3661 __kmp_is_user_lock_initialized_ = 3662 (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized); 3663 3664 __kmp_get_user_lock_location_ = 3665 (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location); 3666 3667 __kmp_set_user_lock_location_ = (void (*)( 3668 kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location); 3669 3670 __kmp_get_user_lock_flags_ = 3671 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags); 3672 3673 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))( 3674 &__kmp_set_queuing_lock_flags); 3675 } break; 3676 3677 #if KMP_USE_ADAPTIVE_LOCKS 3678 case lk_adaptive: { 3679 __kmp_base_user_lock_size = sizeof(kmp_base_adaptive_lock_t); 3680 __kmp_user_lock_size = sizeof(kmp_adaptive_lock_t); 3681 3682 __kmp_get_user_lock_owner_ = 3683 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_owner); 3684 3685 if (__kmp_env_consistency_check) { 3686 KMP_BIND_USER_LOCK_WITH_CHECKS(adaptive); 3687 } else { 3688 KMP_BIND_USER_LOCK(adaptive); 3689 } 3690 3691 __kmp_destroy_user_lock_ = 3692 (void (*)(kmp_user_lock_p))(&__kmp_destroy_adaptive_lock); 3693 3694 __kmp_is_user_lock_initialized_ = 3695 (int (*)(kmp_user_lock_p))(&__kmp_is_queuing_lock_initialized); 3696 3697 __kmp_get_user_lock_location_ = 3698 (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_location); 3699 3700 __kmp_set_user_lock_location_ = (void (*)( 3701 kmp_user_lock_p, const ident_t *))(&__kmp_set_queuing_lock_location); 3702 3703 __kmp_get_user_lock_flags_ = 3704 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_queuing_lock_flags); 3705 3706 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))( 3707 &__kmp_set_queuing_lock_flags); 3708 3709 } break; 3710 #endif // KMP_USE_ADAPTIVE_LOCKS 3711 3712 case lk_drdpa: { 3713 __kmp_base_user_lock_size = sizeof(kmp_base_drdpa_lock_t); 3714 __kmp_user_lock_size = sizeof(kmp_drdpa_lock_t); 3715 3716 __kmp_get_user_lock_owner_ = 3717 (kmp_int32(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_owner); 3718 3719 if (__kmp_env_consistency_check) { 3720 KMP_BIND_USER_LOCK_WITH_CHECKS(drdpa); 3721 KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(drdpa); 3722 } else { 3723 KMP_BIND_USER_LOCK(drdpa); 3724 KMP_BIND_NESTED_USER_LOCK(drdpa); 3725 } 3726 3727 __kmp_destroy_user_lock_ = 3728 (void (*)(kmp_user_lock_p))(&__kmp_destroy_drdpa_lock); 3729 3730 __kmp_is_user_lock_initialized_ = 3731 (int (*)(kmp_user_lock_p))(&__kmp_is_drdpa_lock_initialized); 3732 3733 __kmp_get_user_lock_location_ = 3734 (const ident_t *(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_location); 3735 3736 __kmp_set_user_lock_location_ = (void (*)( 3737 kmp_user_lock_p, const ident_t *))(&__kmp_set_drdpa_lock_location); 3738 3739 __kmp_get_user_lock_flags_ = 3740 (kmp_lock_flags_t(*)(kmp_user_lock_p))(&__kmp_get_drdpa_lock_flags); 3741 3742 __kmp_set_user_lock_flags_ = (void (*)(kmp_user_lock_p, kmp_lock_flags_t))( 3743 &__kmp_set_drdpa_lock_flags); 3744 } break; 3745 } 3746 } 3747 3748 // ---------------------------------------------------------------------------- 3749 // User lock table & lock allocation 3750 3751 kmp_lock_table_t __kmp_user_lock_table = {1, 0, NULL}; 3752 kmp_user_lock_p __kmp_lock_pool = NULL; 3753 3754 // Lock block-allocation support. 3755 kmp_block_of_locks *__kmp_lock_blocks = NULL; 3756 int __kmp_num_locks_in_block = 1; // FIXME - tune this value 3757 3758 static kmp_lock_index_t __kmp_lock_table_insert(kmp_user_lock_p lck) { 3759 // Assume that kmp_global_lock is held upon entry/exit. 3760 kmp_lock_index_t index; 3761 if (__kmp_user_lock_table.used >= __kmp_user_lock_table.allocated) { 3762 kmp_lock_index_t size; 3763 kmp_user_lock_p *table; 3764 // Reallocate lock table. 3765 if (__kmp_user_lock_table.allocated == 0) { 3766 size = 1024; 3767 } else { 3768 size = __kmp_user_lock_table.allocated * 2; 3769 } 3770 table = (kmp_user_lock_p *)__kmp_allocate(sizeof(kmp_user_lock_p) * size); 3771 KMP_MEMCPY(table + 1, __kmp_user_lock_table.table + 1, 3772 sizeof(kmp_user_lock_p) * (__kmp_user_lock_table.used - 1)); 3773 table[0] = (kmp_user_lock_p)__kmp_user_lock_table.table; 3774 // We cannot free the previous table now, since it may be in use by other 3775 // threads. So save the pointer to the previous table in in the first 3776 // element of the new table. All the tables will be organized into a list, 3777 // and could be freed when library shutting down. 3778 __kmp_user_lock_table.table = table; 3779 __kmp_user_lock_table.allocated = size; 3780 } 3781 KMP_DEBUG_ASSERT(__kmp_user_lock_table.used < 3782 __kmp_user_lock_table.allocated); 3783 index = __kmp_user_lock_table.used; 3784 __kmp_user_lock_table.table[index] = lck; 3785 ++__kmp_user_lock_table.used; 3786 return index; 3787 } 3788 3789 static kmp_user_lock_p __kmp_lock_block_allocate() { 3790 // Assume that kmp_global_lock is held upon entry/exit. 3791 static int last_index = 0; 3792 if ((last_index >= __kmp_num_locks_in_block) || (__kmp_lock_blocks == NULL)) { 3793 // Restart the index. 3794 last_index = 0; 3795 // Need to allocate a new block. 3796 KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0); 3797 size_t space_for_locks = __kmp_user_lock_size * __kmp_num_locks_in_block; 3798 char *buffer = 3799 (char *)__kmp_allocate(space_for_locks + sizeof(kmp_block_of_locks)); 3800 // Set up the new block. 3801 kmp_block_of_locks *new_block = 3802 (kmp_block_of_locks *)(&buffer[space_for_locks]); 3803 new_block->next_block = __kmp_lock_blocks; 3804 new_block->locks = (void *)buffer; 3805 // Publish the new block. 3806 KMP_MB(); 3807 __kmp_lock_blocks = new_block; 3808 } 3809 kmp_user_lock_p ret = (kmp_user_lock_p)(&( 3810 ((char *)(__kmp_lock_blocks->locks))[last_index * __kmp_user_lock_size])); 3811 last_index++; 3812 return ret; 3813 } 3814 3815 // Get memory for a lock. It may be freshly allocated memory or reused memory 3816 // from lock pool. 3817 kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock, kmp_int32 gtid, 3818 kmp_lock_flags_t flags) { 3819 kmp_user_lock_p lck; 3820 kmp_lock_index_t index; 3821 KMP_DEBUG_ASSERT(user_lock); 3822 3823 __kmp_acquire_lock(&__kmp_global_lock, gtid); 3824 3825 if (__kmp_lock_pool == NULL) { 3826 // Lock pool is empty. Allocate new memory. 3827 3828 if (__kmp_num_locks_in_block <= 1) { // Tune this cutoff point. 3829 lck = (kmp_user_lock_p)__kmp_allocate(__kmp_user_lock_size); 3830 } else { 3831 lck = __kmp_lock_block_allocate(); 3832 } 3833 3834 // Insert lock in the table so that it can be freed in __kmp_cleanup, 3835 // and debugger has info on all allocated locks. 3836 index = __kmp_lock_table_insert(lck); 3837 } else { 3838 // Pick up lock from pool. 3839 lck = __kmp_lock_pool; 3840 index = __kmp_lock_pool->pool.index; 3841 __kmp_lock_pool = __kmp_lock_pool->pool.next; 3842 } 3843 3844 // We could potentially differentiate between nested and regular locks 3845 // here, and do the lock table lookup for regular locks only. 3846 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3847 *((kmp_lock_index_t *)user_lock) = index; 3848 } else { 3849 *((kmp_user_lock_p *)user_lock) = lck; 3850 } 3851 3852 // mark the lock if it is critical section lock. 3853 __kmp_set_user_lock_flags(lck, flags); 3854 3855 __kmp_release_lock(&__kmp_global_lock, gtid); // AC: TODO move this line upper 3856 3857 return lck; 3858 } 3859 3860 // Put lock's memory to pool for reusing. 3861 void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid, 3862 kmp_user_lock_p lck) { 3863 KMP_DEBUG_ASSERT(user_lock != NULL); 3864 KMP_DEBUG_ASSERT(lck != NULL); 3865 3866 __kmp_acquire_lock(&__kmp_global_lock, gtid); 3867 3868 lck->pool.next = __kmp_lock_pool; 3869 __kmp_lock_pool = lck; 3870 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3871 kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock); 3872 KMP_DEBUG_ASSERT(0 < index && index <= __kmp_user_lock_table.used); 3873 lck->pool.index = index; 3874 } 3875 3876 __kmp_release_lock(&__kmp_global_lock, gtid); 3877 } 3878 3879 kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock, char const *func) { 3880 kmp_user_lock_p lck = NULL; 3881 3882 if (__kmp_env_consistency_check) { 3883 if (user_lock == NULL) { 3884 KMP_FATAL(LockIsUninitialized, func); 3885 } 3886 } 3887 3888 if (OMP_LOCK_T_SIZE < sizeof(void *)) { 3889 kmp_lock_index_t index = *((kmp_lock_index_t *)user_lock); 3890 if (__kmp_env_consistency_check) { 3891 if (!(0 < index && index < __kmp_user_lock_table.used)) { 3892 KMP_FATAL(LockIsUninitialized, func); 3893 } 3894 } 3895 KMP_DEBUG_ASSERT(0 < index && index < __kmp_user_lock_table.used); 3896 KMP_DEBUG_ASSERT(__kmp_user_lock_size > 0); 3897 lck = __kmp_user_lock_table.table[index]; 3898 } else { 3899 lck = *((kmp_user_lock_p *)user_lock); 3900 } 3901 3902 if (__kmp_env_consistency_check) { 3903 if (lck == NULL) { 3904 KMP_FATAL(LockIsUninitialized, func); 3905 } 3906 } 3907 3908 return lck; 3909 } 3910 3911 void __kmp_cleanup_user_locks(void) { 3912 // Reset lock pool. Don't worry about lock in the pool--we will free them when 3913 // iterating through lock table (it includes all the locks, dead or alive). 3914 __kmp_lock_pool = NULL; 3915 3916 #define IS_CRITICAL(lck) \ 3917 ((__kmp_get_user_lock_flags_ != NULL) && \ 3918 ((*__kmp_get_user_lock_flags_)(lck)&kmp_lf_critical_section)) 3919 3920 // Loop through lock table, free all locks. 3921 // Do not free item [0], it is reserved for lock tables list. 3922 // 3923 // FIXME - we are iterating through a list of (pointers to) objects of type 3924 // union kmp_user_lock, but we have no way of knowing whether the base type is 3925 // currently "pool" or whatever the global user lock type is. 3926 // 3927 // We are relying on the fact that for all of the user lock types 3928 // (except "tas"), the first field in the lock struct is the "initialized" 3929 // field, which is set to the address of the lock object itself when 3930 // the lock is initialized. When the union is of type "pool", the 3931 // first field is a pointer to the next object in the free list, which 3932 // will not be the same address as the object itself. 3933 // 3934 // This means that the check (*__kmp_is_user_lock_initialized_)(lck) will fail 3935 // for "pool" objects on the free list. This must happen as the "location" 3936 // field of real user locks overlaps the "index" field of "pool" objects. 3937 // 3938 // It would be better to run through the free list, and remove all "pool" 3939 // objects from the lock table before executing this loop. However, 3940 // "pool" objects do not always have their index field set (only on 3941 // lin_32e), and I don't want to search the lock table for the address 3942 // of every "pool" object on the free list. 3943 while (__kmp_user_lock_table.used > 1) { 3944 const ident *loc; 3945 3946 // reduce __kmp_user_lock_table.used before freeing the lock, 3947 // so that state of locks is consistent 3948 kmp_user_lock_p lck = 3949 __kmp_user_lock_table.table[--__kmp_user_lock_table.used]; 3950 3951 if ((__kmp_is_user_lock_initialized_ != NULL) && 3952 (*__kmp_is_user_lock_initialized_)(lck)) { 3953 // Issue a warning if: KMP_CONSISTENCY_CHECK AND lock is initialized AND 3954 // it is NOT a critical section (user is not responsible for destroying 3955 // criticals) AND we know source location to report. 3956 if (__kmp_env_consistency_check && (!IS_CRITICAL(lck)) && 3957 ((loc = __kmp_get_user_lock_location(lck)) != NULL) && 3958 (loc->psource != NULL)) { 3959 kmp_str_loc_t str_loc = __kmp_str_loc_init(loc->psource, false); 3960 KMP_WARNING(CnsLockNotDestroyed, str_loc.file, str_loc.line); 3961 __kmp_str_loc_free(&str_loc); 3962 } 3963 3964 #ifdef KMP_DEBUG 3965 if (IS_CRITICAL(lck)) { 3966 KA_TRACE( 3967 20, 3968 ("__kmp_cleanup_user_locks: free critical section lock %p (%p)\n", 3969 lck, *(void **)lck)); 3970 } else { 3971 KA_TRACE(20, ("__kmp_cleanup_user_locks: free lock %p (%p)\n", lck, 3972 *(void **)lck)); 3973 } 3974 #endif // KMP_DEBUG 3975 3976 // Cleanup internal lock dynamic resources (for drdpa locks particularly). 3977 __kmp_destroy_user_lock(lck); 3978 } 3979 3980 // Free the lock if block allocation of locks is not used. 3981 if (__kmp_lock_blocks == NULL) { 3982 __kmp_free(lck); 3983 } 3984 } 3985 3986 #undef IS_CRITICAL 3987 3988 // delete lock table(s). 3989 kmp_user_lock_p *table_ptr = __kmp_user_lock_table.table; 3990 __kmp_user_lock_table.table = NULL; 3991 __kmp_user_lock_table.allocated = 0; 3992 3993 while (table_ptr != NULL) { 3994 // In the first element we saved the pointer to the previous 3995 // (smaller) lock table. 3996 kmp_user_lock_p *next = (kmp_user_lock_p *)(table_ptr[0]); 3997 __kmp_free(table_ptr); 3998 table_ptr = next; 3999 } 4000 4001 // Free buffers allocated for blocks of locks. 4002 kmp_block_of_locks_t *block_ptr = __kmp_lock_blocks; 4003 __kmp_lock_blocks = NULL; 4004 4005 while (block_ptr != NULL) { 4006 kmp_block_of_locks_t *next = block_ptr->next_block; 4007 __kmp_free(block_ptr->locks); 4008 // *block_ptr itself was allocated at the end of the locks vector. 4009 block_ptr = next; 4010 } 4011 4012 TCW_4(__kmp_init_user_locks, FALSE); 4013 } 4014 4015 #endif // KMP_USE_DYNAMIC_LOCK 4016