21#include "kmp_wait_release.h"
22#include "kmp_wrapper_getpid.h"
25#include <sys/syscall.h>
43void __kmp_validate_locks(
void) {
48 x = ~((kmp_uint32)0) - 2;
51 for (i = 0; i < 8; ++i, ++x, ++y) {
52 kmp_uint32 z = (x - y);
56 KMP_ASSERT(offsetof(kmp_base_queuing_lock, tail_id) % 8 == 0);
70static 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;
74static inline bool __kmp_is_tas_lock_nestable(kmp_tas_lock_t *lck) {
75 return lck->lk.depth_locked != -1;
78__forceinline
static int
79__kmp_acquire_tas_lock_timed_template(kmp_tas_lock_t *lck, kmp_int32 gtid) {
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);
89 kmp_int32 tas_free = KMP_LOCK_FREE(tas);
90 kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
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;
100 KMP_FSYNC_PREPARE(lck);
101 KMP_INIT_YIELD(spins);
102 KMP_INIT_BACKOFF(time);
103 kmp_backoff_t backoff = __kmp_spin_backoff_params;
106 __kmp_spin_backoff(&backoff);
108 if (!__kmp_tpause_enabled)
109 __kmp_spin_backoff(&backoff);
111 KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time);
112 }
while (KMP_ATOMIC_LD_RLX(&lck->lk.poll) != tas_free ||
113 !__kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy));
114 KMP_FSYNC_ACQUIRED(lck);
115 return KMP_LOCK_ACQUIRED_FIRST;
118int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
119 int retval = __kmp_acquire_tas_lock_timed_template(lck, gtid);
123static int __kmp_acquire_tas_lock_with_checks(kmp_tas_lock_t *lck,
125 char const *
const func =
"omp_set_lock";
126 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
127 __kmp_is_tas_lock_nestable(lck)) {
128 KMP_FATAL(LockNestableUsedAsSimple, func);
130 if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) == gtid)) {
131 KMP_FATAL(LockIsAlreadyOwned, func);
133 return __kmp_acquire_tas_lock(lck, gtid);
136int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
137 kmp_int32 tas_free = KMP_LOCK_FREE(tas);
138 kmp_int32 tas_busy = KMP_LOCK_BUSY(gtid + 1, tas);
139 if (KMP_ATOMIC_LD_RLX(&lck->lk.poll) == tas_free &&
140 __kmp_atomic_compare_store_acq(&lck->lk.poll, tas_free, tas_busy)) {
141 KMP_FSYNC_ACQUIRED(lck);
147static int __kmp_test_tas_lock_with_checks(kmp_tas_lock_t *lck,
149 char const *
const func =
"omp_test_lock";
150 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
151 __kmp_is_tas_lock_nestable(lck)) {
152 KMP_FATAL(LockNestableUsedAsSimple, func);
154 return __kmp_test_tas_lock(lck, gtid);
157int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
160 KMP_FSYNC_RELEASING(lck);
161 KMP_ATOMIC_ST_REL(&lck->lk.poll, KMP_LOCK_FREE(tas));
165 return KMP_LOCK_RELEASED;
168static int __kmp_release_tas_lock_with_checks(kmp_tas_lock_t *lck,
170 char const *
const func =
"omp_unset_lock";
172 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
173 __kmp_is_tas_lock_nestable(lck)) {
174 KMP_FATAL(LockNestableUsedAsSimple, func);
176 if (__kmp_get_tas_lock_owner(lck) == -1) {
177 KMP_FATAL(LockUnsettingFree, func);
179 if ((gtid >= 0) && (__kmp_get_tas_lock_owner(lck) >= 0) &&
180 (__kmp_get_tas_lock_owner(lck) != gtid)) {
181 KMP_FATAL(LockUnsettingSetByAnother, func);
183 return __kmp_release_tas_lock(lck, gtid);
186void __kmp_init_tas_lock(kmp_tas_lock_t *lck) {
187 lck->lk.poll = KMP_LOCK_FREE(tas);
190void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck) { lck->lk.poll = 0; }
192static void __kmp_destroy_tas_lock_with_checks(kmp_tas_lock_t *lck) {
193 char const *
const func =
"omp_destroy_lock";
194 if ((
sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
195 __kmp_is_tas_lock_nestable(lck)) {
196 KMP_FATAL(LockNestableUsedAsSimple, func);
198 if (__kmp_get_tas_lock_owner(lck) != -1) {
199 KMP_FATAL(LockStillOwned, func);
201 __kmp_destroy_tas_lock(lck);
206int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
207 KMP_DEBUG_ASSERT(gtid >= 0);
209 if (__kmp_get_tas_lock_owner(lck) == gtid) {
210 lck->lk.depth_locked += 1;
211 return KMP_LOCK_ACQUIRED_NEXT;
213 __kmp_acquire_tas_lock_timed_template(lck, gtid);
214 lck->lk.depth_locked = 1;
215 return KMP_LOCK_ACQUIRED_FIRST;
219static int __kmp_acquire_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
221 char const *
const func =
"omp_set_nest_lock";
222 if (!__kmp_is_tas_lock_nestable(lck)) {
223 KMP_FATAL(LockSimpleUsedAsNestable, func);
225 return __kmp_acquire_nested_tas_lock(lck, gtid);
228int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
231 KMP_DEBUG_ASSERT(gtid >= 0);
233 if (__kmp_get_tas_lock_owner(lck) == gtid) {
234 retval = ++lck->lk.depth_locked;
235 }
else if (!__kmp_test_tas_lock(lck, gtid)) {
239 retval = lck->lk.depth_locked = 1;
244static int __kmp_test_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
246 char const *
const func =
"omp_test_nest_lock";
247 if (!__kmp_is_tas_lock_nestable(lck)) {
248 KMP_FATAL(LockSimpleUsedAsNestable, func);
250 return __kmp_test_nested_tas_lock(lck, gtid);
253int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid) {
254 KMP_DEBUG_ASSERT(gtid >= 0);
257 if (--(lck->lk.depth_locked) == 0) {
258 __kmp_release_tas_lock(lck, gtid);
259 return KMP_LOCK_RELEASED;
261 return KMP_LOCK_STILL_HELD;
264static int __kmp_release_nested_tas_lock_with_checks(kmp_tas_lock_t *lck,
266 char const *
const func =
"omp_unset_nest_lock";
268 if (!__kmp_is_tas_lock_nestable(lck)) {
269 KMP_FATAL(LockSimpleUsedAsNestable, func);
271 if (__kmp_get_tas_lock_owner(lck) == -1) {
272 KMP_FATAL(LockUnsettingFree, func);
274 if (__kmp_get_tas_lock_owner(lck) != gtid) {
275 KMP_FATAL(LockUnsettingSetByAnother, func);
277 return __kmp_release_nested_tas_lock(lck, gtid);
280void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck) {
281 __kmp_init_tas_lock(lck);
282 lck->lk.depth_locked = 0;
285void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck) {
286 __kmp_destroy_tas_lock(lck);
287 lck->lk.depth_locked = 0;
290static void __kmp_destroy_nested_tas_lock_with_checks(kmp_tas_lock_t *lck) {
291 char const *
const func =
"omp_destroy_nest_lock";
292 if (!__kmp_is_tas_lock_nestable(lck)) {
293 KMP_FATAL(LockSimpleUsedAsNestable, func);
295 if (__kmp_get_tas_lock_owner(lck) != -1) {
296 KMP_FATAL(LockStillOwned, func);
298 __kmp_destroy_nested_tas_lock(lck);
311static kmp_int32 __kmp_get_futex_lock_owner(kmp_futex_lock_t *lck) {
312 return KMP_LOCK_STRIP((TCR_4(lck->lk.poll) >> 1)) - 1;
315static inline bool __kmp_is_futex_lock_nestable(kmp_futex_lock_t *lck) {
316 return lck->lk.depth_locked != -1;
319__forceinline
static int
320__kmp_acquire_futex_lock_timed_template(kmp_futex_lock_t *lck, kmp_int32 gtid) {
321 kmp_int32 gtid_code = (gtid + 1) << 1;
325#ifdef USE_LOCK_PROFILE
326 kmp_uint32 curr = KMP_LOCK_STRIP(TCR_4(lck->lk.poll));
327 if ((curr != 0) && (curr != gtid_code))
328 __kmp_printf(
"LOCK CONTENTION: %p\n", lck);
332 KMP_FSYNC_PREPARE(lck);
333 KA_TRACE(1000, (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d entering\n",
334 lck, lck->lk.poll, gtid));
338 while ((poll_val = KMP_COMPARE_AND_STORE_RET32(
339 &(lck->lk.poll), KMP_LOCK_FREE(futex),
340 KMP_LOCK_BUSY(gtid_code, futex))) != KMP_LOCK_FREE(futex)) {
342 kmp_int32 cond = KMP_LOCK_STRIP(poll_val) & 1;
345 (
"__kmp_acquire_futex_lock: lck:%p, T#%d poll_val = 0x%x cond = 0x%x\n",
346 lck, gtid, poll_val, cond));
357 if (!KMP_COMPARE_AND_STORE_REL32(&(lck->lk.poll), poll_val,
358 poll_val | KMP_LOCK_BUSY(1, futex))) {
361 (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d can't set bit 0\n",
362 lck, lck->lk.poll, gtid));
365 poll_val |= KMP_LOCK_BUSY(1, futex);
368 (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d bit 0 set\n", lck,
369 lck->lk.poll, gtid));
374 (
"__kmp_acquire_futex_lock: lck:%p, T#%d before futex_wait(0x%x)\n",
375 lck, gtid, poll_val));
378 if ((rc = syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAIT, poll_val, NULL,
380 KA_TRACE(1000, (
"__kmp_acquire_futex_lock: lck:%p, T#%d futex_wait(0x%x) "
381 "failed (rc=%ld errno=%d)\n",
382 lck, gtid, poll_val, rc, errno));
387 (
"__kmp_acquire_futex_lock: lck:%p, T#%d after futex_wait(0x%x)\n",
388 lck, gtid, poll_val));
395 KMP_FSYNC_ACQUIRED(lck);
396 KA_TRACE(1000, (
"__kmp_acquire_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
397 lck->lk.poll, gtid));
398 return KMP_LOCK_ACQUIRED_FIRST;
401int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
402 int retval = __kmp_acquire_futex_lock_timed_template(lck, gtid);
406static int __kmp_acquire_futex_lock_with_checks(kmp_futex_lock_t *lck,
408 char const *
const func =
"omp_set_lock";
409 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
410 __kmp_is_futex_lock_nestable(lck)) {
411 KMP_FATAL(LockNestableUsedAsSimple, func);
413 if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) == gtid)) {
414 KMP_FATAL(LockIsAlreadyOwned, func);
416 return __kmp_acquire_futex_lock(lck, gtid);
419int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
420 if (KMP_COMPARE_AND_STORE_ACQ32(&(lck->lk.poll), KMP_LOCK_FREE(futex),
421 KMP_LOCK_BUSY((gtid + 1) << 1, futex))) {
422 KMP_FSYNC_ACQUIRED(lck);
428static int __kmp_test_futex_lock_with_checks(kmp_futex_lock_t *lck,
430 char const *
const func =
"omp_test_lock";
431 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
432 __kmp_is_futex_lock_nestable(lck)) {
433 KMP_FATAL(LockNestableUsedAsSimple, func);
435 return __kmp_test_futex_lock(lck, gtid);
438int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
441 KA_TRACE(1000, (
"__kmp_release_futex_lock: lck:%p(0x%x), T#%d entering\n",
442 lck, lck->lk.poll, gtid));
444 KMP_FSYNC_RELEASING(lck);
446 kmp_int32 poll_val = KMP_XCHG_FIXED32(&(lck->lk.poll), KMP_LOCK_FREE(futex));
449 (
"__kmp_release_futex_lock: lck:%p, T#%d released poll_val = 0x%x\n",
450 lck, gtid, poll_val));
452 if (KMP_LOCK_STRIP(poll_val) & 1) {
454 (
"__kmp_release_futex_lock: lck:%p, T#%d futex_wake 1 thread\n",
456 syscall(__NR_futex, &(lck->lk.poll), FUTEX_WAKE, KMP_LOCK_BUSY(1, futex),
462 KA_TRACE(1000, (
"__kmp_release_futex_lock: lck:%p(0x%x), T#%d exiting\n", lck,
463 lck->lk.poll, gtid));
466 return KMP_LOCK_RELEASED;
469static int __kmp_release_futex_lock_with_checks(kmp_futex_lock_t *lck,
471 char const *
const func =
"omp_unset_lock";
473 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
474 __kmp_is_futex_lock_nestable(lck)) {
475 KMP_FATAL(LockNestableUsedAsSimple, func);
477 if (__kmp_get_futex_lock_owner(lck) == -1) {
478 KMP_FATAL(LockUnsettingFree, func);
480 if ((gtid >= 0) && (__kmp_get_futex_lock_owner(lck) >= 0) &&
481 (__kmp_get_futex_lock_owner(lck) != gtid)) {
482 KMP_FATAL(LockUnsettingSetByAnother, func);
484 return __kmp_release_futex_lock(lck, gtid);
487void __kmp_init_futex_lock(kmp_futex_lock_t *lck) {
488 TCW_4(lck->lk.poll, KMP_LOCK_FREE(futex));
491void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck) { lck->lk.poll = 0; }
493static void __kmp_destroy_futex_lock_with_checks(kmp_futex_lock_t *lck) {
494 char const *
const func =
"omp_destroy_lock";
495 if ((
sizeof(kmp_futex_lock_t) <= OMP_LOCK_T_SIZE) &&
496 __kmp_is_futex_lock_nestable(lck)) {
497 KMP_FATAL(LockNestableUsedAsSimple, func);
499 if (__kmp_get_futex_lock_owner(lck) != -1) {
500 KMP_FATAL(LockStillOwned, func);
502 __kmp_destroy_futex_lock(lck);
507int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
508 KMP_DEBUG_ASSERT(gtid >= 0);
510 if (__kmp_get_futex_lock_owner(lck) == gtid) {
511 lck->lk.depth_locked += 1;
512 return KMP_LOCK_ACQUIRED_NEXT;
514 __kmp_acquire_futex_lock_timed_template(lck, gtid);
515 lck->lk.depth_locked = 1;
516 return KMP_LOCK_ACQUIRED_FIRST;
520static int __kmp_acquire_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
522 char const *
const func =
"omp_set_nest_lock";
523 if (!__kmp_is_futex_lock_nestable(lck)) {
524 KMP_FATAL(LockSimpleUsedAsNestable, func);
526 return __kmp_acquire_nested_futex_lock(lck, gtid);
529int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
532 KMP_DEBUG_ASSERT(gtid >= 0);
534 if (__kmp_get_futex_lock_owner(lck) == gtid) {
535 retval = ++lck->lk.depth_locked;
536 }
else if (!__kmp_test_futex_lock(lck, gtid)) {
540 retval = lck->lk.depth_locked = 1;
545static int __kmp_test_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
547 char const *
const func =
"omp_test_nest_lock";
548 if (!__kmp_is_futex_lock_nestable(lck)) {
549 KMP_FATAL(LockSimpleUsedAsNestable, func);
551 return __kmp_test_nested_futex_lock(lck, gtid);
554int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid) {
555 KMP_DEBUG_ASSERT(gtid >= 0);
558 if (--(lck->lk.depth_locked) == 0) {
559 __kmp_release_futex_lock(lck, gtid);
560 return KMP_LOCK_RELEASED;
562 return KMP_LOCK_STILL_HELD;
565static int __kmp_release_nested_futex_lock_with_checks(kmp_futex_lock_t *lck,
567 char const *
const func =
"omp_unset_nest_lock";
569 if (!__kmp_is_futex_lock_nestable(lck)) {
570 KMP_FATAL(LockSimpleUsedAsNestable, func);
572 if (__kmp_get_futex_lock_owner(lck) == -1) {
573 KMP_FATAL(LockUnsettingFree, func);
575 if (__kmp_get_futex_lock_owner(lck) != gtid) {
576 KMP_FATAL(LockUnsettingSetByAnother, func);
578 return __kmp_release_nested_futex_lock(lck, gtid);
581void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck) {
582 __kmp_init_futex_lock(lck);
583 lck->lk.depth_locked = 0;
586void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck) {
587 __kmp_destroy_futex_lock(lck);
588 lck->lk.depth_locked = 0;
591static void __kmp_destroy_nested_futex_lock_with_checks(kmp_futex_lock_t *lck) {
592 char const *
const func =
"omp_destroy_nest_lock";
593 if (!__kmp_is_futex_lock_nestable(lck)) {
594 KMP_FATAL(LockSimpleUsedAsNestable, func);
596 if (__kmp_get_futex_lock_owner(lck) != -1) {
597 KMP_FATAL(LockStillOwned, func);
599 __kmp_destroy_nested_futex_lock(lck);
607static kmp_int32 __kmp_get_ticket_lock_owner(kmp_ticket_lock_t *lck) {
608 return std::atomic_load_explicit(&lck->lk.owner_id,
609 std::memory_order_relaxed) -
613static inline bool __kmp_is_ticket_lock_nestable(kmp_ticket_lock_t *lck) {
614 return std::atomic_load_explicit(&lck->lk.depth_locked,
615 std::memory_order_relaxed) != -1;
618static kmp_uint32 __kmp_bakery_check(
void *now_serving, kmp_uint32 my_ticket) {
619 return std::atomic_load_explicit((std::atomic<unsigned> *)now_serving,
620 std::memory_order_acquire) == my_ticket;
623__forceinline
static int
624__kmp_acquire_ticket_lock_timed_template(kmp_ticket_lock_t *lck,
626 kmp_uint32 my_ticket = std::atomic_fetch_add_explicit(
627 &lck->lk.next_ticket, 1U, std::memory_order_relaxed);
629#ifdef USE_LOCK_PROFILE
630 if (std::atomic_load_explicit(&lck->lk.now_serving,
631 std::memory_order_relaxed) != my_ticket)
632 __kmp_printf(
"LOCK CONTENTION: %p\n", lck);
636 if (std::atomic_load_explicit(&lck->lk.now_serving,
637 std::memory_order_acquire) == my_ticket) {
638 return KMP_LOCK_ACQUIRED_FIRST;
640 KMP_WAIT_PTR(&lck->lk.now_serving, my_ticket, __kmp_bakery_check, lck);
641 return KMP_LOCK_ACQUIRED_FIRST;
644int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
645 int retval = __kmp_acquire_ticket_lock_timed_template(lck, gtid);
649static int __kmp_acquire_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
651 char const *
const func =
"omp_set_lock";
653 if (!std::atomic_load_explicit(&lck->lk.initialized,
654 std::memory_order_relaxed)) {
655 KMP_FATAL(LockIsUninitialized, func);
657 if (lck->lk.self != lck) {
658 KMP_FATAL(LockIsUninitialized, func);
660 if (__kmp_is_ticket_lock_nestable(lck)) {
661 KMP_FATAL(LockNestableUsedAsSimple, func);
663 if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) == gtid)) {
664 KMP_FATAL(LockIsAlreadyOwned, func);
667 __kmp_acquire_ticket_lock(lck, gtid);
669 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
670 std::memory_order_relaxed);
671 return KMP_LOCK_ACQUIRED_FIRST;
674int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
675 kmp_uint32 my_ticket = std::atomic_load_explicit(&lck->lk.next_ticket,
676 std::memory_order_relaxed);
678 if (std::atomic_load_explicit(&lck->lk.now_serving,
679 std::memory_order_relaxed) == my_ticket) {
680 kmp_uint32 next_ticket = my_ticket + 1;
681 if (std::atomic_compare_exchange_strong_explicit(
682 &lck->lk.next_ticket, &my_ticket, next_ticket,
683 std::memory_order_acquire, std::memory_order_acquire)) {
690static int __kmp_test_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
692 char const *
const func =
"omp_test_lock";
694 if (!std::atomic_load_explicit(&lck->lk.initialized,
695 std::memory_order_relaxed)) {
696 KMP_FATAL(LockIsUninitialized, func);
698 if (lck->lk.self != lck) {
699 KMP_FATAL(LockIsUninitialized, func);
701 if (__kmp_is_ticket_lock_nestable(lck)) {
702 KMP_FATAL(LockNestableUsedAsSimple, func);
705 int retval = __kmp_test_ticket_lock(lck, gtid);
708 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
709 std::memory_order_relaxed);
714int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
715 kmp_uint32 distance = std::atomic_load_explicit(&lck->lk.next_ticket,
716 std::memory_order_relaxed) -
717 std::atomic_load_explicit(&lck->lk.now_serving,
718 std::memory_order_relaxed);
720 std::atomic_fetch_add_explicit(&lck->lk.now_serving, 1U,
721 std::memory_order_release);
724 (kmp_uint32)(__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc));
725 return KMP_LOCK_RELEASED;
728static int __kmp_release_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
730 char const *
const func =
"omp_unset_lock";
732 if (!std::atomic_load_explicit(&lck->lk.initialized,
733 std::memory_order_relaxed)) {
734 KMP_FATAL(LockIsUninitialized, func);
736 if (lck->lk.self != lck) {
737 KMP_FATAL(LockIsUninitialized, func);
739 if (__kmp_is_ticket_lock_nestable(lck)) {
740 KMP_FATAL(LockNestableUsedAsSimple, func);
742 if (__kmp_get_ticket_lock_owner(lck) == -1) {
743 KMP_FATAL(LockUnsettingFree, func);
745 if ((gtid >= 0) && (__kmp_get_ticket_lock_owner(lck) >= 0) &&
746 (__kmp_get_ticket_lock_owner(lck) != gtid)) {
747 KMP_FATAL(LockUnsettingSetByAnother, func);
749 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
750 return __kmp_release_ticket_lock(lck, gtid);
753void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck) {
754 lck->lk.location = NULL;
756 std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
757 std::memory_order_relaxed);
758 std::atomic_store_explicit(&lck->lk.now_serving, 0U,
759 std::memory_order_relaxed);
760 std::atomic_store_explicit(
761 &lck->lk.owner_id, 0,
762 std::memory_order_relaxed);
763 std::atomic_store_explicit(
764 &lck->lk.depth_locked, -1,
765 std::memory_order_relaxed);
766 std::atomic_store_explicit(&lck->lk.initialized,
true,
767 std::memory_order_release);
770void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck) {
771 std::atomic_store_explicit(&lck->lk.initialized,
false,
772 std::memory_order_release);
774 lck->lk.location = NULL;
775 std::atomic_store_explicit(&lck->lk.next_ticket, 0U,
776 std::memory_order_relaxed);
777 std::atomic_store_explicit(&lck->lk.now_serving, 0U,
778 std::memory_order_relaxed);
779 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
780 std::atomic_store_explicit(&lck->lk.depth_locked, -1,
781 std::memory_order_relaxed);
784static void __kmp_destroy_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
785 char const *
const func =
"omp_destroy_lock";
787 if (!std::atomic_load_explicit(&lck->lk.initialized,
788 std::memory_order_relaxed)) {
789 KMP_FATAL(LockIsUninitialized, func);
791 if (lck->lk.self != lck) {
792 KMP_FATAL(LockIsUninitialized, func);
794 if (__kmp_is_ticket_lock_nestable(lck)) {
795 KMP_FATAL(LockNestableUsedAsSimple, func);
797 if (__kmp_get_ticket_lock_owner(lck) != -1) {
798 KMP_FATAL(LockStillOwned, func);
800 __kmp_destroy_ticket_lock(lck);
805int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
806 KMP_DEBUG_ASSERT(gtid >= 0);
808 if (__kmp_get_ticket_lock_owner(lck) == gtid) {
809 std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
810 std::memory_order_relaxed);
811 return KMP_LOCK_ACQUIRED_NEXT;
813 __kmp_acquire_ticket_lock_timed_template(lck, gtid);
814 std::atomic_store_explicit(&lck->lk.depth_locked, 1,
815 std::memory_order_relaxed);
816 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
817 std::memory_order_relaxed);
818 return KMP_LOCK_ACQUIRED_FIRST;
822static int __kmp_acquire_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
824 char const *
const func =
"omp_set_nest_lock";
826 if (!std::atomic_load_explicit(&lck->lk.initialized,
827 std::memory_order_relaxed)) {
828 KMP_FATAL(LockIsUninitialized, func);
830 if (lck->lk.self != lck) {
831 KMP_FATAL(LockIsUninitialized, func);
833 if (!__kmp_is_ticket_lock_nestable(lck)) {
834 KMP_FATAL(LockSimpleUsedAsNestable, func);
836 return __kmp_acquire_nested_ticket_lock(lck, gtid);
839int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
842 KMP_DEBUG_ASSERT(gtid >= 0);
844 if (__kmp_get_ticket_lock_owner(lck) == gtid) {
845 retval = std::atomic_fetch_add_explicit(&lck->lk.depth_locked, 1,
846 std::memory_order_relaxed) +
848 }
else if (!__kmp_test_ticket_lock(lck, gtid)) {
851 std::atomic_store_explicit(&lck->lk.depth_locked, 1,
852 std::memory_order_relaxed);
853 std::atomic_store_explicit(&lck->lk.owner_id, gtid + 1,
854 std::memory_order_relaxed);
860static int __kmp_test_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
862 char const *
const func =
"omp_test_nest_lock";
864 if (!std::atomic_load_explicit(&lck->lk.initialized,
865 std::memory_order_relaxed)) {
866 KMP_FATAL(LockIsUninitialized, func);
868 if (lck->lk.self != lck) {
869 KMP_FATAL(LockIsUninitialized, func);
871 if (!__kmp_is_ticket_lock_nestable(lck)) {
872 KMP_FATAL(LockSimpleUsedAsNestable, func);
874 return __kmp_test_nested_ticket_lock(lck, gtid);
877int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid) {
878 KMP_DEBUG_ASSERT(gtid >= 0);
880 if ((std::atomic_fetch_add_explicit(&lck->lk.depth_locked, -1,
881 std::memory_order_relaxed) -
883 std::atomic_store_explicit(&lck->lk.owner_id, 0, std::memory_order_relaxed);
884 __kmp_release_ticket_lock(lck, gtid);
885 return KMP_LOCK_RELEASED;
887 return KMP_LOCK_STILL_HELD;
890static int __kmp_release_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck,
892 char const *
const func =
"omp_unset_nest_lock";
894 if (!std::atomic_load_explicit(&lck->lk.initialized,
895 std::memory_order_relaxed)) {
896 KMP_FATAL(LockIsUninitialized, func);
898 if (lck->lk.self != lck) {
899 KMP_FATAL(LockIsUninitialized, func);
901 if (!__kmp_is_ticket_lock_nestable(lck)) {
902 KMP_FATAL(LockSimpleUsedAsNestable, func);
904 if (__kmp_get_ticket_lock_owner(lck) == -1) {
905 KMP_FATAL(LockUnsettingFree, func);
907 if (__kmp_get_ticket_lock_owner(lck) != gtid) {
908 KMP_FATAL(LockUnsettingSetByAnother, func);
910 return __kmp_release_nested_ticket_lock(lck, gtid);
913void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck) {
914 __kmp_init_ticket_lock(lck);
915 std::atomic_store_explicit(&lck->lk.depth_locked, 0,
916 std::memory_order_relaxed);
920void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck) {
921 __kmp_destroy_ticket_lock(lck);
922 std::atomic_store_explicit(&lck->lk.depth_locked, 0,
923 std::memory_order_relaxed);
927__kmp_destroy_nested_ticket_lock_with_checks(kmp_ticket_lock_t *lck) {
928 char const *
const func =
"omp_destroy_nest_lock";
930 if (!std::atomic_load_explicit(&lck->lk.initialized,
931 std::memory_order_relaxed)) {
932 KMP_FATAL(LockIsUninitialized, func);
934 if (lck->lk.self != lck) {
935 KMP_FATAL(LockIsUninitialized, func);
937 if (!__kmp_is_ticket_lock_nestable(lck)) {
938 KMP_FATAL(LockSimpleUsedAsNestable, func);
940 if (__kmp_get_ticket_lock_owner(lck) != -1) {
941 KMP_FATAL(LockStillOwned, func);
943 __kmp_destroy_nested_ticket_lock(lck);
948static const ident_t *__kmp_get_ticket_lock_location(kmp_ticket_lock_t *lck) {
949 return lck->lk.location;
952static void __kmp_set_ticket_lock_location(kmp_ticket_lock_t *lck,
954 lck->lk.location = loc;
957static kmp_lock_flags_t __kmp_get_ticket_lock_flags(kmp_ticket_lock_t *lck) {
958 return lck->lk.flags;
961static void __kmp_set_ticket_lock_flags(kmp_ticket_lock_t *lck,
962 kmp_lock_flags_t flags) {
963 lck->lk.flags = flags;
1021#ifdef DEBUG_QUEUING_LOCKS
1024#define TRACE_BUF_ELE 1024
1025static char traces[TRACE_BUF_ELE][128] = {0};
1027#define TRACE_LOCK(X, Y) \
1028 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s\n", X, Y);
1029#define TRACE_LOCK_T(X, Y, Z) \
1030 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s%d\n", X, Y, Z);
1031#define TRACE_LOCK_HT(X, Y, Z, Q) \
1032 KMP_SNPRINTF(traces[tc++ % TRACE_BUF_ELE], 128, "t%d at %s %d,%d\n", X, Y, \
1035static void __kmp_dump_queuing_lock(kmp_info_t *this_thr, kmp_int32 gtid,
1036 kmp_queuing_lock_t *lck, kmp_int32 head_id,
1037 kmp_int32 tail_id) {
1040 __kmp_printf_no_lock(
"\n__kmp_dump_queuing_lock: TRACE BEGINS HERE! \n");
1042 i = tc % TRACE_BUF_ELE;
1043 __kmp_printf_no_lock(
"%s\n", traces[i]);
1044 i = (i + 1) % TRACE_BUF_ELE;
1045 while (i != (tc % TRACE_BUF_ELE)) {
1046 __kmp_printf_no_lock(
"%s", traces[i]);
1047 i = (i + 1) % TRACE_BUF_ELE;
1049 __kmp_printf_no_lock(
"\n");
1051 __kmp_printf_no_lock(
"\n__kmp_dump_queuing_lock: gtid+1:%d, spin_here:%d, "
1052 "next_wait:%d, head_id:%d, tail_id:%d\n",
1053 gtid + 1, this_thr->th.th_spin_here,
1054 this_thr->th.th_next_waiting, head_id, tail_id);
1056 __kmp_printf_no_lock(
"\t\thead: %d ", lck->lk.head_id);
1058 if (lck->lk.head_id >= 1) {
1059 t = __kmp_threads[lck->lk.head_id - 1]->th.th_next_waiting;
1061 __kmp_printf_no_lock(
"-> %d ", t);
1062 t = __kmp_threads[t - 1]->th.th_next_waiting;
1065 __kmp_printf_no_lock(
"; tail: %d ", lck->lk.tail_id);
1066 __kmp_printf_no_lock(
"\n\n");
1071static kmp_int32 __kmp_get_queuing_lock_owner(kmp_queuing_lock_t *lck) {
1072 return TCR_4(lck->lk.owner_id) - 1;
1075static inline bool __kmp_is_queuing_lock_nestable(kmp_queuing_lock_t *lck) {
1076 return lck->lk.depth_locked != -1;
1080template <
bool takeTime>
1083__forceinline
static int
1084__kmp_acquire_queuing_lock_timed_template(kmp_queuing_lock_t *lck,
1086 kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1087 volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1088 volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1089 volatile kmp_uint32 *spin_here_p;
1092 ompt_state_t prev_state = ompt_state_undefined;
1096 (
"__kmp_acquire_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1098 KMP_FSYNC_PREPARE(lck);
1099 KMP_DEBUG_ASSERT(this_thr != NULL);
1100 spin_here_p = &this_thr->th.th_spin_here;
1102#ifdef DEBUG_QUEUING_LOCKS
1103 TRACE_LOCK(gtid + 1,
"acq ent");
1105 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1106 if (this_thr->th.th_next_waiting != 0)
1107 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1109 KMP_DEBUG_ASSERT(!*spin_here_p);
1110 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1119 *spin_here_p = TRUE;
1131#ifdef DEBUG_QUEUING_LOCKS
1133 TRACE_LOCK_HT(gtid + 1,
"acq read: ", head, tail);
1141 enqueued = KMP_COMPARE_AND_STORE_ACQ64((
volatile kmp_int64 *)tail_id_p,
1143 KMP_PACK_64(gtid + 1, gtid + 1));
1144#ifdef DEBUG_QUEUING_LOCKS
1146 TRACE_LOCK(gtid + 1,
"acq enq: (-1,0)->(tid,tid)");
1152 KMP_DEBUG_ASSERT(tail != gtid + 1);
1154#ifdef DEBUG_QUEUING_LOCKS
1155 TRACE_LOCK_HT(gtid + 1,
"acq read: ", head, tail);
1162 enqueued = KMP_COMPARE_AND_STORE_ACQ32(tail_id_p, tail, gtid + 1);
1164#ifdef DEBUG_QUEUING_LOCKS
1166 TRACE_LOCK(gtid + 1,
"acq enq: (h,t)->(h,tid)");
1173 kmp_int32 grabbed_lock;
1175#ifdef DEBUG_QUEUING_LOCKS
1177 TRACE_LOCK_HT(gtid + 1,
"acq read: ", head, tail);
1183 grabbed_lock = KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1);
1187 *spin_here_p = FALSE;
1191 (
"__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: no queuing\n",
1193#ifdef DEBUG_QUEUING_LOCKS
1194 TRACE_LOCK_HT(gtid + 1,
"acq exit: ", head, 0);
1198 if (ompt_enabled.enabled && prev_state != ompt_state_undefined) {
1200 this_thr->th.ompt_thread_info.state = prev_state;
1201 this_thr->th.ompt_thread_info.wait_id = 0;
1205 KMP_FSYNC_ACQUIRED(lck);
1206 return KMP_LOCK_ACQUIRED_FIRST;
1213 if (ompt_enabled.enabled && prev_state == ompt_state_undefined) {
1215 prev_state = this_thr->th.ompt_thread_info.state;
1216 this_thr->th.ompt_thread_info.wait_id = (uint64_t)lck;
1217 this_thr->th.ompt_thread_info.state = ompt_state_wait_lock;
1223 kmp_info_t *tail_thr = __kmp_thread_from_gtid(tail - 1);
1224 KMP_ASSERT(tail_thr != NULL);
1225 tail_thr->th.th_next_waiting = gtid + 1;
1229 (
"__kmp_acquire_queuing_lock: lck:%p, T#%d waiting for lock\n",
1234 KMP_WAIT(spin_here_p, FALSE, KMP_EQ, lck);
1239#ifdef DEBUG_QUEUING_LOCKS
1240 TRACE_LOCK(gtid + 1,
"acq spin");
1242 if (this_thr->th.th_next_waiting != 0)
1243 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1245 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1246 KA_TRACE(1000, (
"__kmp_acquire_queuing_lock: lck:%p, T#%d exiting: after "
1247 "waiting on queue\n",
1250#ifdef DEBUG_QUEUING_LOCKS
1251 TRACE_LOCK(gtid + 1,
"acq exit 2");
1256 this_thr->th.ompt_thread_info.state = prev_state;
1257 this_thr->th.ompt_thread_info.wait_id = 0;
1261 return KMP_LOCK_ACQUIRED_FIRST;
1267 KMP_YIELD_OVERSUB();
1269#ifdef DEBUG_QUEUING_LOCKS
1270 TRACE_LOCK(gtid + 1,
"acq retry");
1273 KMP_ASSERT2(0,
"should not get here");
1274 return KMP_LOCK_ACQUIRED_FIRST;
1277int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1278 KMP_DEBUG_ASSERT(gtid >= 0);
1280 int retval = __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1284static int __kmp_acquire_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1286 char const *
const func =
"omp_set_lock";
1287 if (lck->lk.initialized != lck) {
1288 KMP_FATAL(LockIsUninitialized, func);
1290 if (__kmp_is_queuing_lock_nestable(lck)) {
1291 KMP_FATAL(LockNestableUsedAsSimple, func);
1293 if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1294 KMP_FATAL(LockIsAlreadyOwned, func);
1297 __kmp_acquire_queuing_lock(lck, gtid);
1299 lck->lk.owner_id = gtid + 1;
1300 return KMP_LOCK_ACQUIRED_FIRST;
1303int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1304 volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1307 kmp_info_t *this_thr;
1310 KA_TRACE(1000, (
"__kmp_test_queuing_lock: T#%d entering\n", gtid));
1311 KMP_DEBUG_ASSERT(gtid >= 0);
1313 this_thr = __kmp_thread_from_gtid(gtid);
1314 KMP_DEBUG_ASSERT(this_thr != NULL);
1315 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1322 if (KMP_COMPARE_AND_STORE_ACQ32(head_id_p, 0, -1)) {
1324 (
"__kmp_test_queuing_lock: T#%d exiting: holding lock\n", gtid));
1325 KMP_FSYNC_ACQUIRED(lck);
1331 (
"__kmp_test_queuing_lock: T#%d exiting: without lock\n", gtid));
1335static int __kmp_test_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1337 char const *
const func =
"omp_test_lock";
1338 if (lck->lk.initialized != lck) {
1339 KMP_FATAL(LockIsUninitialized, func);
1341 if (__kmp_is_queuing_lock_nestable(lck)) {
1342 KMP_FATAL(LockNestableUsedAsSimple, func);
1345 int retval = __kmp_test_queuing_lock(lck, gtid);
1348 lck->lk.owner_id = gtid + 1;
1353int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1354 volatile kmp_int32 *head_id_p = &lck->lk.head_id;
1355 volatile kmp_int32 *tail_id_p = &lck->lk.tail_id;
1358 (
"__kmp_release_queuing_lock: lck:%p, T#%d entering\n", lck, gtid));
1359 KMP_DEBUG_ASSERT(gtid >= 0);
1360#if KMP_DEBUG || DEBUG_QUEUING_LOCKS
1361 kmp_info_t *this_thr = __kmp_thread_from_gtid(gtid);
1363 KMP_DEBUG_ASSERT(this_thr != NULL);
1364#ifdef DEBUG_QUEUING_LOCKS
1365 TRACE_LOCK(gtid + 1,
"rel ent");
1367 if (this_thr->th.th_spin_here)
1368 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1369 if (this_thr->th.th_next_waiting != 0)
1370 __kmp_dump_queuing_lock(this_thr, gtid, lck, *head_id_p, *tail_id_p);
1372 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here);
1373 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0);
1375 KMP_FSYNC_RELEASING(lck);
1384#ifdef DEBUG_QUEUING_LOCKS
1386 TRACE_LOCK_HT(gtid + 1,
"rel read: ", head, tail);
1388 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1390 KMP_DEBUG_ASSERT(head !=
1395 if (KMP_COMPARE_AND_STORE_REL32(head_id_p, -1, 0)) {
1398 (
"__kmp_release_queuing_lock: lck:%p, T#%d exiting: queue empty\n",
1400#ifdef DEBUG_QUEUING_LOCKS
1401 TRACE_LOCK_HT(gtid + 1,
"rel exit: ", 0, 0);
1407 return KMP_LOCK_RELEASED;
1414#ifdef DEBUG_QUEUING_LOCKS
1416 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1418 KMP_DEBUG_ASSERT(head > 0);
1421 dequeued = KMP_COMPARE_AND_STORE_REL64(
1422 RCAST(
volatile kmp_int64 *, tail_id_p), KMP_PACK_64(head, head),
1423 KMP_PACK_64(-1, 0));
1424#ifdef DEBUG_QUEUING_LOCKS
1425 TRACE_LOCK(gtid + 1,
"rel deq: (h,h)->(-1,0)");
1429 volatile kmp_int32 *waiting_id_p;
1430 kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1431 KMP_DEBUG_ASSERT(head_thr != NULL);
1432 waiting_id_p = &head_thr->th.th_next_waiting;
1435#ifdef DEBUG_QUEUING_LOCKS
1436 if (head <= 0 || tail <= 0)
1437 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1439 KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1446 KMP_WAIT((
volatile kmp_uint32 *)waiting_id_p, 0, KMP_NEQ, NULL);
1447#ifdef DEBUG_QUEUING_LOCKS
1448 TRACE_LOCK(gtid + 1,
"rel deq: (h,t)->(h',t)");
1455 kmp_info_t *head_thr = __kmp_thread_from_gtid(head - 1);
1456 KMP_DEBUG_ASSERT(head_thr != NULL);
1459#ifdef DEBUG_QUEUING_LOCKS
1460 if (head <= 0 || tail <= 0)
1461 __kmp_dump_queuing_lock(this_thr, gtid, lck, head, tail);
1463 KMP_DEBUG_ASSERT(head > 0 && tail > 0);
1467 head_thr->th.th_next_waiting = 0;
1468#ifdef DEBUG_QUEUING_LOCKS
1469 TRACE_LOCK_T(gtid + 1,
"rel nw=0 for t=", head);
1474 head_thr->th.th_spin_here = FALSE;
1476 KA_TRACE(1000, (
"__kmp_release_queuing_lock: lck:%p, T#%d exiting: after "
1479#ifdef DEBUG_QUEUING_LOCKS
1480 TRACE_LOCK(gtid + 1,
"rel exit 2");
1482 return KMP_LOCK_RELEASED;
1487#ifdef DEBUG_QUEUING_LOCKS
1488 TRACE_LOCK(gtid + 1,
"rel retry");
1492 KMP_ASSERT2(0,
"should not get here");
1493 return KMP_LOCK_RELEASED;
1496static int __kmp_release_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1498 char const *
const func =
"omp_unset_lock";
1500 if (lck->lk.initialized != lck) {
1501 KMP_FATAL(LockIsUninitialized, func);
1503 if (__kmp_is_queuing_lock_nestable(lck)) {
1504 KMP_FATAL(LockNestableUsedAsSimple, func);
1506 if (__kmp_get_queuing_lock_owner(lck) == -1) {
1507 KMP_FATAL(LockUnsettingFree, func);
1509 if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1510 KMP_FATAL(LockUnsettingSetByAnother, func);
1512 lck->lk.owner_id = 0;
1513 return __kmp_release_queuing_lock(lck, gtid);
1516void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck) {
1517 lck->lk.location = NULL;
1518 lck->lk.head_id = 0;
1519 lck->lk.tail_id = 0;
1520 lck->lk.next_ticket = 0;
1521 lck->lk.now_serving = 0;
1522 lck->lk.owner_id = 0;
1523 lck->lk.depth_locked = -1;
1524 lck->lk.initialized = lck;
1526 KA_TRACE(1000, (
"__kmp_init_queuing_lock: lock %p initialized\n", lck));
1529void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck) {
1530 lck->lk.initialized = NULL;
1531 lck->lk.location = NULL;
1532 lck->lk.head_id = 0;
1533 lck->lk.tail_id = 0;
1534 lck->lk.next_ticket = 0;
1535 lck->lk.now_serving = 0;
1536 lck->lk.owner_id = 0;
1537 lck->lk.depth_locked = -1;
1540static void __kmp_destroy_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1541 char const *
const func =
"omp_destroy_lock";
1542 if (lck->lk.initialized != lck) {
1543 KMP_FATAL(LockIsUninitialized, func);
1545 if (__kmp_is_queuing_lock_nestable(lck)) {
1546 KMP_FATAL(LockNestableUsedAsSimple, func);
1548 if (__kmp_get_queuing_lock_owner(lck) != -1) {
1549 KMP_FATAL(LockStillOwned, func);
1551 __kmp_destroy_queuing_lock(lck);
1556int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1557 KMP_DEBUG_ASSERT(gtid >= 0);
1559 if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1560 lck->lk.depth_locked += 1;
1561 return KMP_LOCK_ACQUIRED_NEXT;
1563 __kmp_acquire_queuing_lock_timed_template<false>(lck, gtid);
1565 lck->lk.depth_locked = 1;
1567 lck->lk.owner_id = gtid + 1;
1568 return KMP_LOCK_ACQUIRED_FIRST;
1573__kmp_acquire_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1575 char const *
const func =
"omp_set_nest_lock";
1576 if (lck->lk.initialized != lck) {
1577 KMP_FATAL(LockIsUninitialized, func);
1579 if (!__kmp_is_queuing_lock_nestable(lck)) {
1580 KMP_FATAL(LockSimpleUsedAsNestable, func);
1582 return __kmp_acquire_nested_queuing_lock(lck, gtid);
1585int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1588 KMP_DEBUG_ASSERT(gtid >= 0);
1590 if (__kmp_get_queuing_lock_owner(lck) == gtid) {
1591 retval = ++lck->lk.depth_locked;
1592 }
else if (!__kmp_test_queuing_lock(lck, gtid)) {
1596 retval = lck->lk.depth_locked = 1;
1598 lck->lk.owner_id = gtid + 1;
1603static int __kmp_test_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1605 char const *
const func =
"omp_test_nest_lock";
1606 if (lck->lk.initialized != lck) {
1607 KMP_FATAL(LockIsUninitialized, func);
1609 if (!__kmp_is_queuing_lock_nestable(lck)) {
1610 KMP_FATAL(LockSimpleUsedAsNestable, func);
1612 return __kmp_test_nested_queuing_lock(lck, gtid);
1615int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid) {
1616 KMP_DEBUG_ASSERT(gtid >= 0);
1619 if (--(lck->lk.depth_locked) == 0) {
1621 lck->lk.owner_id = 0;
1622 __kmp_release_queuing_lock(lck, gtid);
1623 return KMP_LOCK_RELEASED;
1625 return KMP_LOCK_STILL_HELD;
1629__kmp_release_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck,
1631 char const *
const func =
"omp_unset_nest_lock";
1633 if (lck->lk.initialized != lck) {
1634 KMP_FATAL(LockIsUninitialized, func);
1636 if (!__kmp_is_queuing_lock_nestable(lck)) {
1637 KMP_FATAL(LockSimpleUsedAsNestable, func);
1639 if (__kmp_get_queuing_lock_owner(lck) == -1) {
1640 KMP_FATAL(LockUnsettingFree, func);
1642 if (__kmp_get_queuing_lock_owner(lck) != gtid) {
1643 KMP_FATAL(LockUnsettingSetByAnother, func);
1645 return __kmp_release_nested_queuing_lock(lck, gtid);
1648void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1649 __kmp_init_queuing_lock(lck);
1650 lck->lk.depth_locked = 0;
1653void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck) {
1654 __kmp_destroy_queuing_lock(lck);
1655 lck->lk.depth_locked = 0;
1659__kmp_destroy_nested_queuing_lock_with_checks(kmp_queuing_lock_t *lck) {
1660 char const *
const func =
"omp_destroy_nest_lock";
1661 if (lck->lk.initialized != lck) {
1662 KMP_FATAL(LockIsUninitialized, func);
1664 if (!__kmp_is_queuing_lock_nestable(lck)) {
1665 KMP_FATAL(LockSimpleUsedAsNestable, func);
1667 if (__kmp_get_queuing_lock_owner(lck) != -1) {
1668 KMP_FATAL(LockStillOwned, func);
1670 __kmp_destroy_nested_queuing_lock(lck);
1675static const ident_t *__kmp_get_queuing_lock_location(kmp_queuing_lock_t *lck) {
1676 return lck->lk.location;
1679static void __kmp_set_queuing_lock_location(kmp_queuing_lock_t *lck,
1681 lck->lk.location = loc;
1684static kmp_lock_flags_t __kmp_get_queuing_lock_flags(kmp_queuing_lock_t *lck) {
1685 return lck->lk.flags;
1688static void __kmp_set_queuing_lock_flags(kmp_queuing_lock_t *lck,
1689 kmp_lock_flags_t flags) {
1690 lck->lk.flags = flags;
1693#if KMP_USE_ADAPTIVE_LOCKS
1697#if KMP_HAVE_RTM_INTRINSICS
1698#include <immintrin.h>
1699#define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1704#define _XBEGIN_STARTED (~0u)
1705#define _XABORT_EXPLICIT (1 << 0)
1706#define _XABORT_RETRY (1 << 1)
1707#define _XABORT_CONFLICT (1 << 2)
1708#define _XABORT_CAPACITY (1 << 3)
1709#define _XABORT_DEBUG (1 << 4)
1710#define _XABORT_NESTED (1 << 5)
1711#define _XABORT_CODE(x) ((unsigned char)(((x) >> 24) & 0xFF))
1714#define SOFT_ABORT_MASK (_XABORT_RETRY | _XABORT_CONFLICT | _XABORT_EXPLICIT)
1716#define STRINGIZE_INTERNAL(arg) #arg
1717#define STRINGIZE(arg) STRINGIZE_INTERNAL(arg)
1723static __inline
int _xbegin() {
1760 __asm__
volatile(
"1: .byte 0xC7; .byte 0xF8;\n"
1763 "1: movl %%eax,%0\n"
1765 :
"+r"(res)::
"memory",
"%eax");
1771static __inline
void _xend() {
1779 __asm__
volatile(
".byte 0x0f; .byte 0x01; .byte 0xd5" :::
"memory");
1788#define _xabort(ARG) _asm _emit 0xc6 _asm _emit 0xf8 _asm _emit ARG
1790#define _xabort(ARG) \
1791 __asm__ volatile(".byte 0xC6; .byte 0xF8; .byte " STRINGIZE(ARG):::"memory");
1797#if KMP_DEBUG_ADAPTIVE_LOCKS
1802static kmp_adaptive_lock_statistics_t destroyedStats;
1805static kmp_adaptive_lock_info_t liveLocks;
1808static kmp_bootstrap_lock_t chain_lock =
1809 KMP_BOOTSTRAP_LOCK_INITIALIZER(chain_lock);
1812void __kmp_init_speculative_stats() {
1813 kmp_adaptive_lock_info_t *lck = &liveLocks;
1815 memset(CCAST(kmp_adaptive_lock_statistics_t *, &(lck->stats)), 0,
1816 sizeof(lck->stats));
1817 lck->stats.next = lck;
1818 lck->stats.prev = lck;
1820 KMP_ASSERT(lck->stats.next->stats.prev == lck);
1821 KMP_ASSERT(lck->stats.prev->stats.next == lck);
1823 __kmp_init_bootstrap_lock(&chain_lock);
1827static void __kmp_remember_lock(kmp_adaptive_lock_info_t *lck) {
1828 __kmp_acquire_bootstrap_lock(&chain_lock);
1830 lck->stats.next = liveLocks.stats.next;
1831 lck->stats.prev = &liveLocks;
1833 liveLocks.stats.next = lck;
1834 lck->stats.next->stats.prev = lck;
1836 KMP_ASSERT(lck->stats.next->stats.prev == lck);
1837 KMP_ASSERT(lck->stats.prev->stats.next == lck);
1839 __kmp_release_bootstrap_lock(&chain_lock);
1842static void __kmp_forget_lock(kmp_adaptive_lock_info_t *lck) {
1843 KMP_ASSERT(lck->stats.next->stats.prev == lck);
1844 KMP_ASSERT(lck->stats.prev->stats.next == lck);
1846 kmp_adaptive_lock_info_t *n = lck->stats.next;
1847 kmp_adaptive_lock_info_t *p = lck->stats.prev;
1853static void __kmp_zero_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1854 memset(CCAST(kmp_adaptive_lock_statistics_t *, &lck->stats), 0,
1855 sizeof(lck->stats));
1856 __kmp_remember_lock(lck);
1859static void __kmp_add_stats(kmp_adaptive_lock_statistics_t *t,
1860 kmp_adaptive_lock_info_t *lck) {
1861 kmp_adaptive_lock_statistics_t
volatile *s = &lck->stats;
1863 t->nonSpeculativeAcquireAttempts += lck->acquire_attempts;
1864 t->successfulSpeculations += s->successfulSpeculations;
1865 t->hardFailedSpeculations += s->hardFailedSpeculations;
1866 t->softFailedSpeculations += s->softFailedSpeculations;
1867 t->nonSpeculativeAcquires += s->nonSpeculativeAcquires;
1868 t->lemmingYields += s->lemmingYields;
1871static void __kmp_accumulate_speculative_stats(kmp_adaptive_lock_info_t *lck) {
1872 __kmp_acquire_bootstrap_lock(&chain_lock);
1874 __kmp_add_stats(&destroyedStats, lck);
1875 __kmp_forget_lock(lck);
1877 __kmp_release_bootstrap_lock(&chain_lock);
1880static float percent(kmp_uint32 count, kmp_uint32 total) {
1881 return (total == 0) ? 0.0 : (100.0 * count) / total;
1884void __kmp_print_speculative_stats() {
1885 kmp_adaptive_lock_statistics_t total = destroyedStats;
1886 kmp_adaptive_lock_info_t *lck;
1888 for (lck = liveLocks.stats.next; lck != &liveLocks; lck = lck->stats.next) {
1889 __kmp_add_stats(&total, lck);
1891 kmp_adaptive_lock_statistics_t *t = &total;
1892 kmp_uint32 totalSections =
1893 t->nonSpeculativeAcquires + t->successfulSpeculations;
1894 kmp_uint32 totalSpeculations = t->successfulSpeculations +
1895 t->hardFailedSpeculations +
1896 t->softFailedSpeculations;
1897 if (totalSections <= 0)
1901 if (strcmp(__kmp_speculative_statsfile,
"-") == 0) {
1904 size_t buffLen = KMP_STRLEN(__kmp_speculative_statsfile) + 20;
1905 char buffer[buffLen];
1906 KMP_SNPRINTF(&buffer[0], buffLen, __kmp_speculative_statsfile,
1907 (kmp_int32)getpid());
1908 statsFile.
open(buffer,
"w");
1911 fprintf(statsFile,
"Speculative lock statistics (all approximate!)\n");
1913 " Lock parameters: \n"
1914 " max_soft_retries : %10d\n"
1915 " max_badness : %10d\n",
1916 __kmp_adaptive_backoff_params.max_soft_retries,
1917 __kmp_adaptive_backoff_params.max_badness);
1918 fprintf(statsFile,
" Non-speculative acquire attempts : %10d\n",
1919 t->nonSpeculativeAcquireAttempts);
1920 fprintf(statsFile,
" Total critical sections : %10d\n",
1922 fprintf(statsFile,
" Successful speculations : %10d (%5.1f%%)\n",
1923 t->successfulSpeculations,
1924 percent(t->successfulSpeculations, totalSections));
1925 fprintf(statsFile,
" Non-speculative acquires : %10d (%5.1f%%)\n",
1926 t->nonSpeculativeAcquires,
1927 percent(t->nonSpeculativeAcquires, totalSections));
1928 fprintf(statsFile,
" Lemming yields : %10d\n\n",
1931 fprintf(statsFile,
" Speculative acquire attempts : %10d\n",
1933 fprintf(statsFile,
" Successes : %10d (%5.1f%%)\n",
1934 t->successfulSpeculations,
1935 percent(t->successfulSpeculations, totalSpeculations));
1936 fprintf(statsFile,
" Soft failures : %10d (%5.1f%%)\n",
1937 t->softFailedSpeculations,
1938 percent(t->softFailedSpeculations, totalSpeculations));
1939 fprintf(statsFile,
" Hard failures : %10d (%5.1f%%)\n",
1940 t->hardFailedSpeculations,
1941 percent(t->hardFailedSpeculations, totalSpeculations));
1944#define KMP_INC_STAT(lck, stat) (lck->lk.adaptive.stats.stat++)
1946#define KMP_INC_STAT(lck, stat)
1950static inline bool __kmp_is_unlocked_queuing_lock(kmp_queuing_lock_t *lck) {
1953 bool res = lck->lk.head_id == 0;
1960 __sync_synchronize();
1968__kmp_update_badness_after_success(kmp_adaptive_lock_t *lck) {
1970 lck->lk.adaptive.badness = 0;
1971 KMP_INC_STAT(lck, successfulSpeculations);
1975static __inline
void __kmp_step_badness(kmp_adaptive_lock_t *lck) {
1976 kmp_uint32 newBadness = (lck->lk.adaptive.badness << 1) | 1;
1977 if (newBadness > lck->lk.adaptive.max_badness) {
1980 lck->lk.adaptive.badness = newBadness;
1985KMP_ATTRIBUTE_TARGET_RTM
1986static __inline
int __kmp_should_speculate(kmp_adaptive_lock_t *lck,
1988 kmp_uint32 badness = lck->lk.adaptive.badness;
1989 kmp_uint32 attempts = lck->lk.adaptive.acquire_attempts;
1990 int res = (attempts & badness) == 0;
1996KMP_ATTRIBUTE_TARGET_RTM
1997static int __kmp_test_adaptive_lock_only(kmp_adaptive_lock_t *lck,
1999 int retries = lck->lk.adaptive.max_soft_retries;