xref: /freebsd/contrib/llvm-project/openmp/runtime/src/kmp_lock.h (revision 439352ac8257c8419cb4a662abb7f260f31f9932)
1 /*
2  * kmp_lock.h -- lock header file
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 #ifndef KMP_LOCK_H
14 #define KMP_LOCK_H
15 
16 #include <limits.h> // CHAR_BIT
17 #include <stddef.h> // offsetof
18 
19 #include "kmp_debug.h"
20 #include "kmp_os.h"
21 
22 #ifdef __cplusplus
23 #include <atomic>
24 
25 extern "C" {
26 #endif // __cplusplus
27 
28 // ----------------------------------------------------------------------------
29 // Have to copy these definitions from kmp.h because kmp.h cannot be included
30 // due to circular dependencies.  Will undef these at end of file.
31 
32 #define KMP_PAD(type, sz)                                                      \
33   (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
34 #define KMP_GTID_DNE (-2)
35 
36 // Forward declaration of ident and ident_t
37 
38 struct ident;
39 typedef struct ident ident_t;
40 
41 // End of copied code.
42 // ----------------------------------------------------------------------------
43 
44 // We need to know the size of the area we can assume that the compiler(s)
45 // allocated for objects of type omp_lock_t and omp_nest_lock_t.  The Intel
46 // compiler always allocates a pointer-sized area, as does visual studio.
47 //
48 // gcc however, only allocates 4 bytes for regular locks, even on 64-bit
49 // intel archs.  It allocates at least 8 bytes for nested lock (more on
50 // recent versions), but we are bounded by the pointer-sized chunks that
51 // the Intel compiler allocates.
52 
53 #if (KMP_OS_LINUX || KMP_OS_AIX) && defined(KMP_GOMP_COMPAT)
54 #define OMP_LOCK_T_SIZE sizeof(int)
55 #define OMP_NEST_LOCK_T_SIZE sizeof(void *)
56 #else
57 #define OMP_LOCK_T_SIZE sizeof(void *)
58 #define OMP_NEST_LOCK_T_SIZE sizeof(void *)
59 #endif
60 
61 // The Intel compiler allocates a 32-byte chunk for a critical section.
62 // Both gcc and visual studio only allocate enough space for a pointer.
63 // Sometimes we know that the space was allocated by the Intel compiler.
64 #define OMP_CRITICAL_SIZE sizeof(void *)
65 #define INTEL_CRITICAL_SIZE 32
66 
67 // lock flags
68 typedef kmp_uint32 kmp_lock_flags_t;
69 
70 #define kmp_lf_critical_section 1
71 
72 // When a lock table is used, the indices are of kmp_lock_index_t
73 typedef kmp_uint32 kmp_lock_index_t;
74 
75 // When memory allocated for locks are on the lock pool (free list),
76 // it is treated as structs of this type.
77 struct kmp_lock_pool {
78   union kmp_user_lock *next;
79   kmp_lock_index_t index;
80 };
81 
82 typedef struct kmp_lock_pool kmp_lock_pool_t;
83 
84 extern void __kmp_validate_locks(void);
85 
86 // ----------------------------------------------------------------------------
87 //  There are 5 lock implementations:
88 //       1. Test and set locks.
89 //       2. futex locks (Linux* OS on x86 and
90 //          Intel(R) Many Integrated Core Architecture)
91 //       3. Ticket (Lamport bakery) locks.
92 //       4. Queuing locks (with separate spin fields).
93 //       5. DRPA (Dynamically Reconfigurable Distributed Polling Area) locks
94 //
95 //   and 3 lock purposes:
96 //       1. Bootstrap locks -- Used for a few locks available at library
97 //       startup-shutdown time.
98 //          These do not require non-negative global thread ID's.
99 //       2. Internal RTL locks -- Used everywhere else in the RTL
100 //       3. User locks (includes critical sections)
101 // ----------------------------------------------------------------------------
102 
103 // ============================================================================
104 // Lock implementations.
105 //
106 // Test and set locks.
107 //
108 // Non-nested test and set locks differ from the other lock kinds (except
109 // futex) in that we use the memory allocated by the compiler for the lock,
110 // rather than a pointer to it.
111 //
112 // On lin32, lin_32e, and win_32, the space allocated may be as small as 4
113 // bytes, so we have to use a lock table for nested locks, and avoid accessing
114 // the depth_locked field for non-nested locks.
115 //
116 // Information normally available to the tools, such as lock location, lock
117 // usage (normal lock vs. critical section), etc. is not available with test and
118 // set locks.
119 // ----------------------------------------------------------------------------
120 
121 struct kmp_base_tas_lock {
122   // KMP_LOCK_FREE(tas) => unlocked; locked: (gtid+1) of owning thread
123 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) &&     \
124     __LP64__
125   // Flip the ordering of the high and low 32-bit member to be consistent
126   // with the memory layout of the address in 64-bit big-endian.
127   kmp_int32 depth_locked; // depth locked, for nested locks only
128   std::atomic<kmp_int32> poll;
129 #else
130   std::atomic<kmp_int32> poll;
131   kmp_int32 depth_locked; // depth locked, for nested locks only
132 #endif
133 };
134 
135 typedef struct kmp_base_tas_lock kmp_base_tas_lock_t;
136 
137 union kmp_tas_lock {
138   kmp_base_tas_lock_t lk;
139   kmp_lock_pool_t pool; // make certain struct is large enough
140   double lk_align; // use worst case alignment; no cache line padding
141 };
142 
143 typedef union kmp_tas_lock kmp_tas_lock_t;
144 
145 // Static initializer for test and set lock variables. Usage:
146 //    kmp_tas_lock_t xlock = KMP_TAS_LOCK_INITIALIZER( xlock );
147 #define KMP_TAS_LOCK_INITIALIZER(lock)                                         \
148   {                                                                            \
149     { KMP_LOCK_FREE(tas), 0 }                                                  \
150   }
151 
152 extern int __kmp_acquire_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
153 extern int __kmp_test_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
154 extern int __kmp_release_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
155 extern void __kmp_init_tas_lock(kmp_tas_lock_t *lck);
156 extern void __kmp_destroy_tas_lock(kmp_tas_lock_t *lck);
157 
158 extern int __kmp_acquire_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
159 extern int __kmp_test_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
160 extern int __kmp_release_nested_tas_lock(kmp_tas_lock_t *lck, kmp_int32 gtid);
161 extern void __kmp_init_nested_tas_lock(kmp_tas_lock_t *lck);
162 extern void __kmp_destroy_nested_tas_lock(kmp_tas_lock_t *lck);
163 
164 #define KMP_LOCK_RELEASED 1
165 #define KMP_LOCK_STILL_HELD 0
166 #define KMP_LOCK_ACQUIRED_FIRST 1
167 #define KMP_LOCK_ACQUIRED_NEXT 0
168 #ifndef KMP_USE_FUTEX
169 #define KMP_USE_FUTEX                                                          \
170   (KMP_OS_LINUX &&                                                             \
171    (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64))
172 #endif
173 #if KMP_USE_FUTEX
174 
175 // ----------------------------------------------------------------------------
176 // futex locks.  futex locks are only available on Linux* OS.
177 //
178 // Like non-nested test and set lock, non-nested futex locks use the memory
179 // allocated by the compiler for the lock, rather than a pointer to it.
180 //
181 // Information normally available to the tools, such as lock location, lock
182 // usage (normal lock vs. critical section), etc. is not available with test and
183 // set locks. With non-nested futex locks, the lock owner is not even available.
184 // ----------------------------------------------------------------------------
185 
186 struct kmp_base_futex_lock {
187   volatile kmp_int32 poll; // KMP_LOCK_FREE(futex) => unlocked
188   // 2*(gtid+1) of owning thread, 0 if unlocked
189   // locked: (gtid+1) of owning thread
190   kmp_int32 depth_locked; // depth locked, for nested locks only
191 };
192 
193 typedef struct kmp_base_futex_lock kmp_base_futex_lock_t;
194 
195 union kmp_futex_lock {
196   kmp_base_futex_lock_t lk;
197   kmp_lock_pool_t pool; // make certain struct is large enough
198   double lk_align; // use worst case alignment
199   // no cache line padding
200 };
201 
202 typedef union kmp_futex_lock kmp_futex_lock_t;
203 
204 // Static initializer for futex lock variables. Usage:
205 //    kmp_futex_lock_t xlock = KMP_FUTEX_LOCK_INITIALIZER( xlock );
206 #define KMP_FUTEX_LOCK_INITIALIZER(lock)                                       \
207   {                                                                            \
208     { KMP_LOCK_FREE(futex), 0 }                                                \
209   }
210 
211 extern int __kmp_acquire_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
212 extern int __kmp_test_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
213 extern int __kmp_release_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
214 extern void __kmp_init_futex_lock(kmp_futex_lock_t *lck);
215 extern void __kmp_destroy_futex_lock(kmp_futex_lock_t *lck);
216 
217 extern int __kmp_acquire_nested_futex_lock(kmp_futex_lock_t *lck,
218                                            kmp_int32 gtid);
219 extern int __kmp_test_nested_futex_lock(kmp_futex_lock_t *lck, kmp_int32 gtid);
220 extern int __kmp_release_nested_futex_lock(kmp_futex_lock_t *lck,
221                                            kmp_int32 gtid);
222 extern void __kmp_init_nested_futex_lock(kmp_futex_lock_t *lck);
223 extern void __kmp_destroy_nested_futex_lock(kmp_futex_lock_t *lck);
224 
225 #endif // KMP_USE_FUTEX
226 
227 // ----------------------------------------------------------------------------
228 // Ticket locks.
229 
230 #ifdef __cplusplus
231 
232 #ifdef _MSC_VER
233 // MSVC won't allow use of std::atomic<> in a union since it has non-trivial
234 // copy constructor.
235 
236 struct kmp_base_ticket_lock {
237   // `initialized' must be the first entry in the lock data structure!
238   std::atomic_bool initialized;
239   volatile union kmp_ticket_lock *self; // points to the lock union
240   ident_t const *location; // Source code location of omp_init_lock().
241   std::atomic_uint
242       next_ticket; // ticket number to give to next thread which acquires
243   std::atomic_uint now_serving; // ticket number for thread which holds the lock
244   std::atomic_int owner_id; // (gtid+1) of owning thread, 0 if unlocked
245   std::atomic_int depth_locked; // depth locked, for nested locks only
246   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
247 };
248 #else
249 struct kmp_base_ticket_lock {
250   // `initialized' must be the first entry in the lock data structure!
251   std::atomic<bool> initialized;
252   volatile union kmp_ticket_lock *self; // points to the lock union
253   ident_t const *location; // Source code location of omp_init_lock().
254   std::atomic<unsigned>
255       next_ticket; // ticket number to give to next thread which acquires
256   std::atomic<unsigned>
257       now_serving; // ticket number for thread which holds the lock
258   std::atomic<int> owner_id; // (gtid+1) of owning thread, 0 if unlocked
259   std::atomic<int> depth_locked; // depth locked, for nested locks only
260   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
261 };
262 #endif
263 
264 #else // __cplusplus
265 
266 struct kmp_base_ticket_lock;
267 
268 #endif // !__cplusplus
269 
270 typedef struct kmp_base_ticket_lock kmp_base_ticket_lock_t;
271 
272 union KMP_ALIGN_CACHE kmp_ticket_lock {
273   kmp_base_ticket_lock_t
274       lk; // This field must be first to allow static initializing.
275   kmp_lock_pool_t pool;
276   double lk_align; // use worst case alignment
277   char lk_pad[KMP_PAD(kmp_base_ticket_lock_t, CACHE_LINE)];
278 };
279 
280 typedef union kmp_ticket_lock kmp_ticket_lock_t;
281 
282 // Static initializer for simple ticket lock variables. Usage:
283 //    kmp_ticket_lock_t xlock = KMP_TICKET_LOCK_INITIALIZER( xlock );
284 // Note the macro argument. It is important to make var properly initialized.
285 #define KMP_TICKET_LOCK_INITIALIZER(lock)                                      \
286   {                                                                            \
287     { true, &(lock), NULL, 0U, 0U, 0, -1 }                                     \
288   }
289 
290 extern int __kmp_acquire_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid);
291 extern int __kmp_test_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid);
292 extern int __kmp_test_ticket_lock_with_cheks(kmp_ticket_lock_t *lck,
293                                              kmp_int32 gtid);
294 extern int __kmp_release_ticket_lock(kmp_ticket_lock_t *lck, kmp_int32 gtid);
295 extern void __kmp_init_ticket_lock(kmp_ticket_lock_t *lck);
296 extern void __kmp_destroy_ticket_lock(kmp_ticket_lock_t *lck);
297 
298 extern int __kmp_acquire_nested_ticket_lock(kmp_ticket_lock_t *lck,
299                                             kmp_int32 gtid);
300 extern int __kmp_test_nested_ticket_lock(kmp_ticket_lock_t *lck,
301                                          kmp_int32 gtid);
302 extern int __kmp_release_nested_ticket_lock(kmp_ticket_lock_t *lck,
303                                             kmp_int32 gtid);
304 extern void __kmp_init_nested_ticket_lock(kmp_ticket_lock_t *lck);
305 extern void __kmp_destroy_nested_ticket_lock(kmp_ticket_lock_t *lck);
306 
307 // ----------------------------------------------------------------------------
308 // Queuing locks.
309 
310 #if KMP_USE_ADAPTIVE_LOCKS
311 
312 struct kmp_adaptive_lock_info;
313 
314 typedef struct kmp_adaptive_lock_info kmp_adaptive_lock_info_t;
315 
316 #if KMP_DEBUG_ADAPTIVE_LOCKS
317 
318 struct kmp_adaptive_lock_statistics {
319   /* So we can get stats from locks that haven't been destroyed. */
320   kmp_adaptive_lock_info_t *next;
321   kmp_adaptive_lock_info_t *prev;
322 
323   /* Other statistics */
324   kmp_uint32 successfulSpeculations;
325   kmp_uint32 hardFailedSpeculations;
326   kmp_uint32 softFailedSpeculations;
327   kmp_uint32 nonSpeculativeAcquires;
328   kmp_uint32 nonSpeculativeAcquireAttempts;
329   kmp_uint32 lemmingYields;
330 };
331 
332 typedef struct kmp_adaptive_lock_statistics kmp_adaptive_lock_statistics_t;
333 
334 extern void __kmp_print_speculative_stats();
335 extern void __kmp_init_speculative_stats();
336 
337 #endif // KMP_DEBUG_ADAPTIVE_LOCKS
338 
339 struct kmp_adaptive_lock_info {
340   /* Values used for adaptivity.
341      Although these are accessed from multiple threads we don't access them
342      atomically, because if we miss updates it probably doesn't matter much. (It
343      just affects our decision about whether to try speculation on the lock). */
344   kmp_uint32 volatile badness;
345   kmp_uint32 volatile acquire_attempts;
346   /* Parameters of the lock. */
347   kmp_uint32 max_badness;
348   kmp_uint32 max_soft_retries;
349 
350 #if KMP_DEBUG_ADAPTIVE_LOCKS
351   kmp_adaptive_lock_statistics_t volatile stats;
352 #endif
353 };
354 
355 #endif // KMP_USE_ADAPTIVE_LOCKS
356 
357 struct kmp_base_queuing_lock {
358 
359   //  `initialized' must be the first entry in the lock data structure!
360   volatile union kmp_queuing_lock
361       *initialized; // Points to the lock union if in initialized state.
362 
363   ident_t const *location; // Source code location of omp_init_lock().
364 
365   KMP_ALIGN(8) // tail_id  must be 8-byte aligned!
366 
367   volatile kmp_int32
368       tail_id; // (gtid+1) of thread at tail of wait queue, 0 if empty
369   // Must be no padding here since head/tail used in 8-byte CAS
370   volatile kmp_int32
371       head_id; // (gtid+1) of thread at head of wait queue, 0 if empty
372   // Decl order assumes little endian
373   // bakery-style lock
374   volatile kmp_uint32
375       next_ticket; // ticket number to give to next thread which acquires
376   volatile kmp_uint32
377       now_serving; // ticket number for thread which holds the lock
378   volatile kmp_int32 owner_id; // (gtid+1) of owning thread, 0 if unlocked
379   kmp_int32 depth_locked; // depth locked, for nested locks only
380 
381   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
382 };
383 
384 typedef struct kmp_base_queuing_lock kmp_base_queuing_lock_t;
385 
386 KMP_BUILD_ASSERT(offsetof(kmp_base_queuing_lock_t, tail_id) % 8 == 0);
387 
388 union KMP_ALIGN_CACHE kmp_queuing_lock {
389   kmp_base_queuing_lock_t
390       lk; // This field must be first to allow static initializing.
391   kmp_lock_pool_t pool;
392   double lk_align; // use worst case alignment
393   char lk_pad[KMP_PAD(kmp_base_queuing_lock_t, CACHE_LINE)];
394 };
395 
396 typedef union kmp_queuing_lock kmp_queuing_lock_t;
397 
398 extern int __kmp_acquire_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid);
399 extern int __kmp_test_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid);
400 extern int __kmp_release_queuing_lock(kmp_queuing_lock_t *lck, kmp_int32 gtid);
401 extern void __kmp_init_queuing_lock(kmp_queuing_lock_t *lck);
402 extern void __kmp_destroy_queuing_lock(kmp_queuing_lock_t *lck);
403 
404 extern int __kmp_acquire_nested_queuing_lock(kmp_queuing_lock_t *lck,
405                                              kmp_int32 gtid);
406 extern int __kmp_test_nested_queuing_lock(kmp_queuing_lock_t *lck,
407                                           kmp_int32 gtid);
408 extern int __kmp_release_nested_queuing_lock(kmp_queuing_lock_t *lck,
409                                              kmp_int32 gtid);
410 extern void __kmp_init_nested_queuing_lock(kmp_queuing_lock_t *lck);
411 extern void __kmp_destroy_nested_queuing_lock(kmp_queuing_lock_t *lck);
412 
413 #if KMP_USE_ADAPTIVE_LOCKS
414 
415 // ----------------------------------------------------------------------------
416 // Adaptive locks.
417 struct kmp_base_adaptive_lock {
418   kmp_base_queuing_lock qlk;
419   KMP_ALIGN(CACHE_LINE)
420   kmp_adaptive_lock_info_t
421       adaptive; // Information for the speculative adaptive lock
422 };
423 
424 typedef struct kmp_base_adaptive_lock kmp_base_adaptive_lock_t;
425 
426 union KMP_ALIGN_CACHE kmp_adaptive_lock {
427   kmp_base_adaptive_lock_t lk;
428   kmp_lock_pool_t pool;
429   double lk_align;
430   char lk_pad[KMP_PAD(kmp_base_adaptive_lock_t, CACHE_LINE)];
431 };
432 typedef union kmp_adaptive_lock kmp_adaptive_lock_t;
433 
434 #define GET_QLK_PTR(l) ((kmp_queuing_lock_t *)&(l)->lk.qlk)
435 
436 #endif // KMP_USE_ADAPTIVE_LOCKS
437 
438 // ----------------------------------------------------------------------------
439 // DRDPA ticket locks.
440 struct kmp_base_drdpa_lock {
441   // All of the fields on the first cache line are only written when
442   // initializing or reconfiguring the lock.  These are relatively rare
443   // operations, so data from the first cache line will usually stay resident in
444   // the cache of each thread trying to acquire the lock.
445   //
446   // initialized must be the first entry in the lock data structure!
447   KMP_ALIGN_CACHE
448 
449   volatile union kmp_drdpa_lock
450       *initialized; // points to the lock union if in initialized state
451   ident_t const *location; // Source code location of omp_init_lock().
452   std::atomic<std::atomic<kmp_uint64> *> polls;
453   std::atomic<kmp_uint64> mask; // is 2**num_polls-1 for mod op
454   kmp_uint64 cleanup_ticket; // thread with cleanup ticket
455   std::atomic<kmp_uint64> *old_polls; // will deallocate old_polls
456   kmp_uint32 num_polls; // must be power of 2
457 
458   // next_ticket it needs to exist in a separate cache line, as it is
459   // invalidated every time a thread takes a new ticket.
460   KMP_ALIGN_CACHE
461 
462   std::atomic<kmp_uint64> next_ticket;
463 
464   // now_serving is used to store our ticket value while we hold the lock. It
465   // has a slightly different meaning in the DRDPA ticket locks (where it is
466   // written by the acquiring thread) than it does in the simple ticket locks
467   // (where it is written by the releasing thread).
468   //
469   // Since now_serving is only read and written in the critical section,
470   // it is non-volatile, but it needs to exist on a separate cache line,
471   // as it is invalidated at every lock acquire.
472   //
473   // Likewise, the vars used for nested locks (owner_id and depth_locked) are
474   // only written by the thread owning the lock, so they are put in this cache
475   // line.  owner_id is read by other threads, so it must be declared volatile.
476   KMP_ALIGN_CACHE
477   kmp_uint64 now_serving; // doesn't have to be volatile
478   volatile kmp_uint32 owner_id; // (gtid+1) of owning thread, 0 if unlocked
479   kmp_int32 depth_locked; // depth locked
480   kmp_lock_flags_t flags; // lock specifics, e.g. critical section lock
481 };
482 
483 typedef struct kmp_base_drdpa_lock kmp_base_drdpa_lock_t;
484 
485 union KMP_ALIGN_CACHE kmp_drdpa_lock {
486   kmp_base_drdpa_lock_t
487       lk; // This field must be first to allow static initializing. */
488   kmp_lock_pool_t pool;
489   double lk_align; // use worst case alignment
490   char lk_pad[KMP_PAD(kmp_base_drdpa_lock_t, CACHE_LINE)];
491 };
492 
493 typedef union kmp_drdpa_lock kmp_drdpa_lock_t;
494 
495 extern int __kmp_acquire_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
496 extern int __kmp_test_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
497 extern int __kmp_release_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
498 extern void __kmp_init_drdpa_lock(kmp_drdpa_lock_t *lck);
499 extern void __kmp_destroy_drdpa_lock(kmp_drdpa_lock_t *lck);
500 
501 extern int __kmp_acquire_nested_drdpa_lock(kmp_drdpa_lock_t *lck,
502                                            kmp_int32 gtid);
503 extern int __kmp_test_nested_drdpa_lock(kmp_drdpa_lock_t *lck, kmp_int32 gtid);
504 extern int __kmp_release_nested_drdpa_lock(kmp_drdpa_lock_t *lck,
505                                            kmp_int32 gtid);
506 extern void __kmp_init_nested_drdpa_lock(kmp_drdpa_lock_t *lck);
507 extern void __kmp_destroy_nested_drdpa_lock(kmp_drdpa_lock_t *lck);
508 
509 // ============================================================================
510 // Lock purposes.
511 // ============================================================================
512 
513 // Bootstrap locks.
514 //
515 // Bootstrap locks -- very few locks used at library initialization time.
516 // Bootstrap locks are currently implemented as ticket locks.
517 // They could also be implemented as test and set lock, but cannot be
518 // implemented with other lock kinds as they require gtids which are not
519 // available at initialization time.
520 
521 typedef kmp_ticket_lock_t kmp_bootstrap_lock_t;
522 
523 #define KMP_BOOTSTRAP_LOCK_INITIALIZER(lock) KMP_TICKET_LOCK_INITIALIZER((lock))
524 #define KMP_BOOTSTRAP_LOCK_INIT(lock)                                          \
525   kmp_bootstrap_lock_t lock = KMP_TICKET_LOCK_INITIALIZER(lock)
526 
527 static inline int __kmp_acquire_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
528   return __kmp_acquire_ticket_lock(lck, KMP_GTID_DNE);
529 }
530 
531 static inline int __kmp_test_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
532   return __kmp_test_ticket_lock(lck, KMP_GTID_DNE);
533 }
534 
535 static inline void __kmp_release_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
536   __kmp_release_ticket_lock(lck, KMP_GTID_DNE);
537 }
538 
539 static inline void __kmp_init_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
540   __kmp_init_ticket_lock(lck);
541 }
542 
543 static inline void __kmp_destroy_bootstrap_lock(kmp_bootstrap_lock_t *lck) {
544   __kmp_destroy_ticket_lock(lck);
545 }
546 
547 // Internal RTL locks.
548 //
549 // Internal RTL locks are also implemented as ticket locks, for now.
550 //
551 // FIXME - We should go through and figure out which lock kind works best for
552 // each internal lock, and use the type declaration and function calls for
553 // that explicit lock kind (and get rid of this section).
554 
555 typedef kmp_ticket_lock_t kmp_lock_t;
556 
557 #define KMP_LOCK_INIT(lock) kmp_lock_t lock = KMP_TICKET_LOCK_INITIALIZER(lock)
558 
559 static inline int __kmp_acquire_lock(kmp_lock_t *lck, kmp_int32 gtid) {
560   return __kmp_acquire_ticket_lock(lck, gtid);
561 }
562 
563 static inline int __kmp_test_lock(kmp_lock_t *lck, kmp_int32 gtid) {
564   return __kmp_test_ticket_lock(lck, gtid);
565 }
566 
567 static inline void __kmp_release_lock(kmp_lock_t *lck, kmp_int32 gtid) {
568   __kmp_release_ticket_lock(lck, gtid);
569 }
570 
571 static inline void __kmp_init_lock(kmp_lock_t *lck) {
572   __kmp_init_ticket_lock(lck);
573 }
574 
575 static inline void __kmp_destroy_lock(kmp_lock_t *lck) {
576   __kmp_destroy_ticket_lock(lck);
577 }
578 
579 // User locks.
580 //
581 // Do not allocate objects of type union kmp_user_lock!!! This will waste space
582 // unless __kmp_user_lock_kind == lk_drdpa. Instead, check the value of
583 // __kmp_user_lock_kind and allocate objects of the type of the appropriate
584 // union member, and cast their addresses to kmp_user_lock_p.
585 
586 enum kmp_lock_kind {
587   lk_default = 0,
588   lk_tas,
589 #if KMP_USE_FUTEX
590   lk_futex,
591 #endif
592 #if KMP_USE_DYNAMIC_LOCK && KMP_USE_TSX
593   lk_hle,
594   lk_rtm_queuing,
595   lk_rtm_spin,
596 #endif
597   lk_ticket,
598   lk_queuing,
599   lk_drdpa,
600 #if KMP_USE_ADAPTIVE_LOCKS
601   lk_adaptive
602 #endif // KMP_USE_ADAPTIVE_LOCKS
603 };
604 
605 typedef enum kmp_lock_kind kmp_lock_kind_t;
606 
607 extern kmp_lock_kind_t __kmp_user_lock_kind;
608 
609 union kmp_user_lock {
610   kmp_tas_lock_t tas;
611 #if KMP_USE_FUTEX
612   kmp_futex_lock_t futex;
613 #endif
614   kmp_ticket_lock_t ticket;
615   kmp_queuing_lock_t queuing;
616   kmp_drdpa_lock_t drdpa;
617 #if KMP_USE_ADAPTIVE_LOCKS
618   kmp_adaptive_lock_t adaptive;
619 #endif // KMP_USE_ADAPTIVE_LOCKS
620   kmp_lock_pool_t pool;
621 };
622 
623 typedef union kmp_user_lock *kmp_user_lock_p;
624 
625 #if !KMP_USE_DYNAMIC_LOCK
626 
627 extern size_t __kmp_base_user_lock_size;
628 extern size_t __kmp_user_lock_size;
629 
630 extern kmp_int32 (*__kmp_get_user_lock_owner_)(kmp_user_lock_p lck);
631 
632 static inline kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p lck) {
633   KMP_DEBUG_ASSERT(__kmp_get_user_lock_owner_ != NULL);
634   return (*__kmp_get_user_lock_owner_)(lck);
635 }
636 
637 extern int (*__kmp_acquire_user_lock_with_checks_)(kmp_user_lock_p lck,
638                                                    kmp_int32 gtid);
639 
640 #if KMP_OS_LINUX &&                                                            \
641     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
642 
643 #define __kmp_acquire_user_lock_with_checks(lck, gtid)                         \
644   if (__kmp_user_lock_kind == lk_tas) {                                        \
645     if (__kmp_env_consistency_check) {                                         \
646       char const *const func = "omp_set_lock";                                 \
647       if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&                       \
648           lck->tas.lk.depth_locked != -1) {                                    \
649         KMP_FATAL(LockNestableUsedAsSimple, func);                             \
650       }                                                                        \
651       if ((gtid >= 0) && (lck->tas.lk.poll - 1 == gtid)) {                     \
652         KMP_FATAL(LockIsAlreadyOwned, func);                                   \
653       }                                                                        \
654     }                                                                          \
655     if (lck->tas.lk.poll != 0 ||                                               \
656         !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)) {     \
657       kmp_uint32 spins;                                                        \
658       kmp_uint64 time;                                                         \
659       KMP_FSYNC_PREPARE(lck);                                                  \
660       KMP_INIT_YIELD(spins);                                                   \
661       KMP_INIT_BACKOFF(time);                                                  \
662       do {                                                                     \
663         KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time);                              \
664       } while (                                                                \
665           lck->tas.lk.poll != 0 ||                                             \
666           !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));    \
667     }                                                                          \
668     KMP_FSYNC_ACQUIRED(lck);                                                   \
669   } else {                                                                     \
670     KMP_DEBUG_ASSERT(__kmp_acquire_user_lock_with_checks_ != NULL);            \
671     (*__kmp_acquire_user_lock_with_checks_)(lck, gtid);                        \
672   }
673 
674 #else
675 static inline int __kmp_acquire_user_lock_with_checks(kmp_user_lock_p lck,
676                                                       kmp_int32 gtid) {
677   KMP_DEBUG_ASSERT(__kmp_acquire_user_lock_with_checks_ != NULL);
678   return (*__kmp_acquire_user_lock_with_checks_)(lck, gtid);
679 }
680 #endif
681 
682 extern int (*__kmp_test_user_lock_with_checks_)(kmp_user_lock_p lck,
683                                                 kmp_int32 gtid);
684 
685 #if KMP_OS_LINUX &&                                                            \
686     (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64)
687 
688 #include "kmp_i18n.h" /* AC: KMP_FATAL definition */
689 extern int __kmp_env_consistency_check; /* AC: copy from kmp.h here */
690 static inline int __kmp_test_user_lock_with_checks(kmp_user_lock_p lck,
691                                                    kmp_int32 gtid) {
692   if (__kmp_user_lock_kind == lk_tas) {
693     if (__kmp_env_consistency_check) {
694       char const *const func = "omp_test_lock";
695       if ((sizeof(kmp_tas_lock_t) <= OMP_LOCK_T_SIZE) &&
696           lck->tas.lk.depth_locked != -1) {
697         KMP_FATAL(LockNestableUsedAsSimple, func);
698       }
699     }
700     return ((lck->tas.lk.poll == 0) &&
701             __kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));
702   } else {
703     KMP_DEBUG_ASSERT(__kmp_test_user_lock_with_checks_ != NULL);
704     return (*__kmp_test_user_lock_with_checks_)(lck, gtid);
705   }
706 }
707 #else
708 static inline int __kmp_test_user_lock_with_checks(kmp_user_lock_p lck,
709                                                    kmp_int32 gtid) {
710   KMP_DEBUG_ASSERT(__kmp_test_user_lock_with_checks_ != NULL);
711   return (*__kmp_test_user_lock_with_checks_)(lck, gtid);
712 }
713 #endif
714 
715 extern int (*__kmp_release_user_lock_with_checks_)(kmp_user_lock_p lck,
716                                                    kmp_int32 gtid);
717 
718 static inline void __kmp_release_user_lock_with_checks(kmp_user_lock_p lck,
719                                                        kmp_int32 gtid) {
720   KMP_DEBUG_ASSERT(__kmp_release_user_lock_with_checks_ != NULL);
721   (*__kmp_release_user_lock_with_checks_)(lck, gtid);
722 }
723 
724 extern void (*__kmp_init_user_lock_with_checks_)(kmp_user_lock_p lck);
725 
726 static inline void __kmp_init_user_lock_with_checks(kmp_user_lock_p lck) {
727   KMP_DEBUG_ASSERT(__kmp_init_user_lock_with_checks_ != NULL);
728   (*__kmp_init_user_lock_with_checks_)(lck);
729 }
730 
731 // We need a non-checking version of destroy lock for when the RTL is
732 // doing the cleanup as it can't always tell if the lock is nested or not.
733 extern void (*__kmp_destroy_user_lock_)(kmp_user_lock_p lck);
734 
735 static inline void __kmp_destroy_user_lock(kmp_user_lock_p lck) {
736   KMP_DEBUG_ASSERT(__kmp_destroy_user_lock_ != NULL);
737   (*__kmp_destroy_user_lock_)(lck);
738 }
739 
740 extern void (*__kmp_destroy_user_lock_with_checks_)(kmp_user_lock_p lck);
741 
742 static inline void __kmp_destroy_user_lock_with_checks(kmp_user_lock_p lck) {
743   KMP_DEBUG_ASSERT(__kmp_destroy_user_lock_with_checks_ != NULL);
744   (*__kmp_destroy_user_lock_with_checks_)(lck);
745 }
746 
747 extern int (*__kmp_acquire_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
748                                                           kmp_int32 gtid);
749 
750 #if KMP_OS_LINUX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
751 
752 #define __kmp_acquire_nested_user_lock_with_checks(lck, gtid, depth)           \
753   if (__kmp_user_lock_kind == lk_tas) {                                        \
754     if (__kmp_env_consistency_check) {                                         \
755       char const *const func = "omp_set_nest_lock";                            \
756       if ((sizeof(kmp_tas_lock_t) <= OMP_NEST_LOCK_T_SIZE) &&                  \
757           lck->tas.lk.depth_locked == -1) {                                    \
758         KMP_FATAL(LockSimpleUsedAsNestable, func);                             \
759       }                                                                        \
760     }                                                                          \
761     if (lck->tas.lk.poll - 1 == gtid) {                                        \
762       lck->tas.lk.depth_locked += 1;                                           \
763       *depth = KMP_LOCK_ACQUIRED_NEXT;                                         \
764     } else {                                                                   \
765       if ((lck->tas.lk.poll != 0) ||                                           \
766           !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1)) {   \
767         kmp_uint32 spins;                                                      \
768         kmp_uint64 time;                                                       \
769         KMP_FSYNC_PREPARE(lck);                                                \
770         KMP_INIT_YIELD(spins);                                                 \
771         KMP_INIT_BACKOFF(time);                                                \
772         do {                                                                   \
773           KMP_YIELD_OVERSUB_ELSE_SPIN(spins, time);                            \
774         } while (                                                              \
775             (lck->tas.lk.poll != 0) ||                                         \
776             !__kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));  \
777       }                                                                        \
778       lck->tas.lk.depth_locked = 1;                                            \
779       *depth = KMP_LOCK_ACQUIRED_FIRST;                                        \
780     }                                                                          \
781     KMP_FSYNC_ACQUIRED(lck);                                                   \
782   } else {                                                                     \
783     KMP_DEBUG_ASSERT(__kmp_acquire_nested_user_lock_with_checks_ != NULL);     \
784     *depth = (*__kmp_acquire_nested_user_lock_with_checks_)(lck, gtid);        \
785   }
786 
787 #else
788 static inline void
789 __kmp_acquire_nested_user_lock_with_checks(kmp_user_lock_p lck, kmp_int32 gtid,
790                                            int *depth) {
791   KMP_DEBUG_ASSERT(__kmp_acquire_nested_user_lock_with_checks_ != NULL);
792   *depth = (*__kmp_acquire_nested_user_lock_with_checks_)(lck, gtid);
793 }
794 #endif
795 
796 extern int (*__kmp_test_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
797                                                        kmp_int32 gtid);
798 
799 #if KMP_OS_LINUX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
800 static inline int __kmp_test_nested_user_lock_with_checks(kmp_user_lock_p lck,
801                                                           kmp_int32 gtid) {
802   if (__kmp_user_lock_kind == lk_tas) {
803     int retval;
804     if (__kmp_env_consistency_check) {
805       char const *const func = "omp_test_nest_lock";
806       if ((sizeof(kmp_tas_lock_t) <= OMP_NEST_LOCK_T_SIZE) &&
807           lck->tas.lk.depth_locked == -1) {
808         KMP_FATAL(LockSimpleUsedAsNestable, func);
809       }
810     }
811     KMP_DEBUG_ASSERT(gtid >= 0);
812     if (lck->tas.lk.poll - 1 ==
813         gtid) { /* __kmp_get_tas_lock_owner( lck ) == gtid */
814       return ++lck->tas.lk.depth_locked; /* same owner, depth increased */
815     }
816     retval = ((lck->tas.lk.poll == 0) &&
817               __kmp_atomic_compare_store_acq(&lck->tas.lk.poll, 0, gtid + 1));
818     if (retval) {
819       KMP_MB();
820       lck->tas.lk.depth_locked = 1;
821     }
822     return retval;
823   } else {
824     KMP_DEBUG_ASSERT(__kmp_test_nested_user_lock_with_checks_ != NULL);
825     return (*__kmp_test_nested_user_lock_with_checks_)(lck, gtid);
826   }
827 }
828 #else
829 static inline int __kmp_test_nested_user_lock_with_checks(kmp_user_lock_p lck,
830                                                           kmp_int32 gtid) {
831   KMP_DEBUG_ASSERT(__kmp_test_nested_user_lock_with_checks_ != NULL);
832   return (*__kmp_test_nested_user_lock_with_checks_)(lck, gtid);
833 }
834 #endif
835 
836 extern int (*__kmp_release_nested_user_lock_with_checks_)(kmp_user_lock_p lck,
837                                                           kmp_int32 gtid);
838 
839 static inline int
840 __kmp_release_nested_user_lock_with_checks(kmp_user_lock_p lck,
841                                            kmp_int32 gtid) {
842   KMP_DEBUG_ASSERT(__kmp_release_nested_user_lock_with_checks_ != NULL);
843   return (*__kmp_release_nested_user_lock_with_checks_)(lck, gtid);
844 }
845 
846 extern void (*__kmp_init_nested_user_lock_with_checks_)(kmp_user_lock_p lck);
847 
848 static inline void
849 __kmp_init_nested_user_lock_with_checks(kmp_user_lock_p lck) {
850   KMP_DEBUG_ASSERT(__kmp_init_nested_user_lock_with_checks_ != NULL);
851   (*__kmp_init_nested_user_lock_with_checks_)(lck);
852 }
853 
854 extern void (*__kmp_destroy_nested_user_lock_with_checks_)(kmp_user_lock_p lck);
855 
856 static inline void
857 __kmp_destroy_nested_user_lock_with_checks(kmp_user_lock_p lck) {
858   KMP_DEBUG_ASSERT(__kmp_destroy_nested_user_lock_with_checks_ != NULL);
859   (*__kmp_destroy_nested_user_lock_with_checks_)(lck);
860 }
861 
862 // user lock functions which do not necessarily exist for all lock kinds.
863 //
864 // The "set" functions usually have wrapper routines that check for a NULL set
865 // function pointer and call it if non-NULL.
866 //
867 // In some cases, it makes sense to have a "get" wrapper function check for a
868 // NULL get function pointer and return NULL / invalid value / error code if
869 // the function pointer is NULL.
870 //
871 // In other cases, the calling code really should differentiate between an
872 // unimplemented function and one that is implemented but returning NULL /
873 // invalid value.  If this is the case, no get function wrapper exists.
874 
875 extern int (*__kmp_is_user_lock_initialized_)(kmp_user_lock_p lck);
876 
877 // no set function; fields set during local allocation
878 
879 extern const ident_t *(*__kmp_get_user_lock_location_)(kmp_user_lock_p lck);
880 
881 static inline const ident_t *__kmp_get_user_lock_location(kmp_user_lock_p lck) {
882   if (__kmp_get_user_lock_location_ != NULL) {
883     return (*__kmp_get_user_lock_location_)(lck);
884   } else {
885     return NULL;
886   }
887 }
888 
889 extern void (*__kmp_set_user_lock_location_)(kmp_user_lock_p lck,
890                                              const ident_t *loc);
891 
892 static inline void __kmp_set_user_lock_location(kmp_user_lock_p lck,
893                                                 const ident_t *loc) {
894   if (__kmp_set_user_lock_location_ != NULL) {
895     (*__kmp_set_user_lock_location_)(lck, loc);
896   }
897 }
898 
899 extern kmp_lock_flags_t (*__kmp_get_user_lock_flags_)(kmp_user_lock_p lck);
900 
901 extern void (*__kmp_set_user_lock_flags_)(kmp_user_lock_p lck,
902                                           kmp_lock_flags_t flags);
903 
904 static inline void __kmp_set_user_lock_flags(kmp_user_lock_p lck,
905                                              kmp_lock_flags_t flags) {
906   if (__kmp_set_user_lock_flags_ != NULL) {
907     (*__kmp_set_user_lock_flags_)(lck, flags);
908   }
909 }
910 
911 // The function which sets up all of the vtbl pointers for kmp_user_lock_t.
912 extern void __kmp_set_user_lock_vptrs(kmp_lock_kind_t user_lock_kind);
913 
914 // Macros for binding user lock functions.
915 #define KMP_BIND_USER_LOCK_TEMPLATE(nest, kind, suffix)                        \
916   {                                                                            \
917     __kmp_acquire##nest##user_lock_with_checks_ = (int (*)(                    \
918         kmp_user_lock_p, kmp_int32))__kmp_acquire##nest##kind##_##suffix;      \
919     __kmp_release##nest##user_lock_with_checks_ = (int (*)(                    \
920         kmp_user_lock_p, kmp_int32))__kmp_release##nest##kind##_##suffix;      \
921     __kmp_test##nest##user_lock_with_checks_ = (int (*)(                       \
922         kmp_user_lock_p, kmp_int32))__kmp_test##nest##kind##_##suffix;         \
923     __kmp_init##nest##user_lock_with_checks_ =                                 \
924         (void (*)(kmp_user_lock_p))__kmp_init##nest##kind##_##suffix;          \
925     __kmp_destroy##nest##user_lock_with_checks_ =                              \
926         (void (*)(kmp_user_lock_p))__kmp_destroy##nest##kind##_##suffix;       \
927   }
928 
929 #define KMP_BIND_USER_LOCK(kind) KMP_BIND_USER_LOCK_TEMPLATE(_, kind, lock)
930 #define KMP_BIND_USER_LOCK_WITH_CHECKS(kind)                                   \
931   KMP_BIND_USER_LOCK_TEMPLATE(_, kind, lock_with_checks)
932 #define KMP_BIND_NESTED_USER_LOCK(kind)                                        \
933   KMP_BIND_USER_LOCK_TEMPLATE(_nested_, kind, lock)
934 #define KMP_BIND_NESTED_USER_LOCK_WITH_CHECKS(kind)                            \
935   KMP_BIND_USER_LOCK_TEMPLATE(_nested_, kind, lock_with_checks)
936 
937 // User lock table & lock allocation
938 /* On 64-bit Linux* OS (and OS X*) GNU compiler allocates only 4 bytems memory
939    for lock variable, which is not enough to store a pointer, so we have to use
940    lock indexes instead of pointers and maintain lock table to map indexes to
941    pointers.
942 
943 
944    Note: The first element of the table is not a pointer to lock! It is a
945    pointer to previously allocated table (or NULL if it is the first table).
946 
947    Usage:
948 
949    if ( OMP_LOCK_T_SIZE < sizeof( <lock> ) ) { // or OMP_NEST_LOCK_T_SIZE
950      Lock table is fully utilized. User locks are indexes, so table is used on
951      user lock operation.
952      Note: it may be the case (lin_32) that we don't need to use a lock
953      table for regular locks, but do need the table for nested locks.
954    }
955    else {
956      Lock table initialized but not actually used.
957    }
958 */
959 
960 struct kmp_lock_table {
961   kmp_lock_index_t used; // Number of used elements
962   kmp_lock_index_t allocated; // Number of allocated elements
963   kmp_user_lock_p *table; // Lock table.
964 };
965 
966 typedef struct kmp_lock_table kmp_lock_table_t;
967 
968 extern kmp_lock_table_t __kmp_user_lock_table;
969 extern kmp_user_lock_p __kmp_lock_pool;
970 
971 struct kmp_block_of_locks {
972   struct kmp_block_of_locks *next_block;
973   void *locks;
974 };
975 
976 typedef struct kmp_block_of_locks kmp_block_of_locks_t;
977 
978 extern kmp_block_of_locks_t *__kmp_lock_blocks;
979 extern int __kmp_num_locks_in_block;
980 
981 extern kmp_user_lock_p __kmp_user_lock_allocate(void **user_lock,
982                                                 kmp_int32 gtid,
983                                                 kmp_lock_flags_t flags);
984 extern void __kmp_user_lock_free(void **user_lock, kmp_int32 gtid,
985                                  kmp_user_lock_p lck);
986 extern kmp_user_lock_p __kmp_lookup_user_lock(void **user_lock,
987                                               char const *func);
988 extern void __kmp_cleanup_user_locks();
989 
990 #define KMP_CHECK_USER_LOCK_INIT()                                             \
991   {                                                                            \
992     if (!TCR_4(__kmp_init_user_locks)) {                                       \
993       __kmp_acquire_bootstrap_lock(&__kmp_initz_lock);                         \
994       if (!TCR_4(__kmp_init_user_locks)) {                                     \
995         TCW_4(__kmp_init_user_locks, TRUE);                                    \
996       }                                                                        \
997       __kmp_release_bootstrap_lock(&__kmp_initz_lock);                         \
998     }                                                                          \
999   }
1000 
1001 #endif // KMP_USE_DYNAMIC_LOCK
1002 
1003 #undef KMP_PAD
1004 #undef KMP_GTID_DNE
1005 
1006 #if KMP_USE_DYNAMIC_LOCK
1007 // KMP_USE_DYNAMIC_LOCK enables dynamic dispatch of lock functions without
1008 // breaking the current compatibility. Essential functionality of this new code
1009 // is dynamic dispatch, but it also implements (or enables implementation of)
1010 // hinted user lock and critical section which will be part of OMP 4.5 soon.
1011 //
1012 // Lock type can be decided at creation time (i.e., lock initialization), and
1013 // subsequent lock function call on the created lock object requires type
1014 // extraction and call through jump table using the extracted type. This type
1015 // information is stored in two different ways depending on the size of the lock
1016 // object, and we differentiate lock types by this size requirement - direct and
1017 // indirect locks.
1018 //
1019 // Direct locks:
1020 // A direct lock object fits into the space created by the compiler for an
1021 // omp_lock_t object, and TAS/Futex lock falls into this category. We use low
1022 // one byte of the lock object as the storage for the lock type, and appropriate
1023 // bit operation is required to access the data meaningful to the lock
1024 // algorithms. Also, to differentiate direct lock from indirect lock, 1 is
1025 // written to LSB of the lock object. The newly introduced "hle" lock is also a
1026 // direct lock.
1027 //
1028 // Indirect locks:
1029 // An indirect lock object requires more space than the compiler-generated
1030 // space, and it should be allocated from heap. Depending on the size of the
1031 // compiler-generated space for the lock (i.e., size of omp_lock_t), this
1032 // omp_lock_t object stores either the address of the heap-allocated indirect
1033 // lock (void * fits in the object) or an index to the indirect lock table entry
1034 // that holds the address. Ticket/Queuing/DRDPA/Adaptive lock falls into this
1035 // category, and the newly introduced "rtm" lock is also an indirect lock which
1036 // was implemented on top of the Queuing lock. When the omp_lock_t object holds
1037 // an index (not lock address), 0 is written to LSB to differentiate the lock
1038 // from a direct lock, and the remaining part is the actual index to the
1039 // indirect lock table.
1040 
1041 #include <stdint.h> // for uintptr_t
1042 
1043 // Shortcuts
1044 #define KMP_USE_INLINED_TAS                                                    \
1045   (KMP_OS_LINUX && (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM)) && 1
1046 #define KMP_USE_INLINED_FUTEX KMP_USE_FUTEX && 0
1047 
1048 // List of lock definitions; all nested locks are indirect locks.
1049 // hle lock is xchg lock prefixed with XACQUIRE/XRELEASE.
1050 // All nested locks are indirect lock types.
1051 #if KMP_USE_TSX
1052 #if KMP_USE_FUTEX
1053 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a) m(futex, a) m(hle, a) m(rtm_spin, a)
1054 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1055   m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing, a)      \
1056       m(nested_tas, a) m(nested_futex, a) m(nested_ticket, a)                  \
1057           m(nested_queuing, a) m(nested_drdpa, a)
1058 #else
1059 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a) m(hle, a) m(rtm_spin, a)
1060 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1061   m(ticket, a) m(queuing, a) m(adaptive, a) m(drdpa, a) m(rtm_queuing, a)      \
1062       m(nested_tas, a) m(nested_ticket, a) m(nested_queuing, a)                \
1063           m(nested_drdpa, a)
1064 #endif // KMP_USE_FUTEX
1065 #define KMP_LAST_D_LOCK lockseq_rtm_spin
1066 #else
1067 #if KMP_USE_FUTEX
1068 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a) m(futex, a)
1069 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1070   m(ticket, a) m(queuing, a) m(drdpa, a) m(nested_tas, a) m(nested_futex, a)   \
1071       m(nested_ticket, a) m(nested_queuing, a) m(nested_drdpa, a)
1072 #define KMP_LAST_D_LOCK lockseq_futex
1073 #else
1074 #define KMP_FOREACH_D_LOCK(m, a) m(tas, a)
1075 #define KMP_FOREACH_I_LOCK(m, a)                                               \
1076   m(ticket, a) m(queuing, a) m(drdpa, a) m(nested_tas, a) m(nested_ticket, a)  \
1077       m(nested_queuing, a) m(nested_drdpa, a)
1078 #define KMP_LAST_D_LOCK lockseq_tas
1079 #endif // KMP_USE_FUTEX
1080 #endif // KMP_USE_TSX
1081 
1082 // Information used in dynamic dispatch
1083 #define KMP_LOCK_SHIFT                                                         \
1084   8 // number of low bits to be used as tag for direct locks
1085 #define KMP_FIRST_D_LOCK lockseq_tas
1086 #define KMP_FIRST_I_LOCK lockseq_ticket
1087 #define KMP_LAST_I_LOCK lockseq_nested_drdpa
1088 #define KMP_NUM_I_LOCKS                                                        \
1089   (locktag_nested_drdpa + 1) // number of indirect lock types
1090 
1091 // Base type for dynamic locks.
1092 typedef kmp_uint32 kmp_dyna_lock_t;
1093 
1094 // Lock sequence that enumerates all lock kinds. Always make this enumeration
1095 // consistent with kmp_lockseq_t in the include directory.
1096 typedef enum {
1097   lockseq_indirect = 0,
1098 #define expand_seq(l, a) lockseq_##l,
1099   KMP_FOREACH_D_LOCK(expand_seq, 0) KMP_FOREACH_I_LOCK(expand_seq, 0)
1100 #undef expand_seq
1101 } kmp_dyna_lockseq_t;
1102 
1103 // Enumerates indirect lock tags.
1104 typedef enum {
1105 #define expand_tag(l, a) locktag_##l,
1106   KMP_FOREACH_I_LOCK(expand_tag, 0)
1107 #undef expand_tag
1108 } kmp_indirect_locktag_t;
1109 
1110 // Utility macros that extract information from lock sequences.
1111 #define KMP_IS_D_LOCK(seq)                                                     \
1112   ((seq) >= KMP_FIRST_D_LOCK && (seq) <= KMP_LAST_D_LOCK)
1113 #define KMP_IS_I_LOCK(seq)                                                     \
1114   ((seq) >= KMP_FIRST_I_LOCK && (seq) <= KMP_LAST_I_LOCK)
1115 #define KMP_GET_I_TAG(seq) (kmp_indirect_locktag_t)((seq)-KMP_FIRST_I_LOCK)
1116 #define KMP_GET_D_TAG(seq) ((seq) << 1 | 1)
1117 
1118 // Enumerates direct lock tags starting from indirect tag.
1119 typedef enum {
1120 #define expand_tag(l, a) locktag_##l = KMP_GET_D_TAG(lockseq_##l),
1121   KMP_FOREACH_D_LOCK(expand_tag, 0)
1122 #undef expand_tag
1123 } kmp_direct_locktag_t;
1124 
1125 // Indirect lock type
1126 typedef struct {
1127   kmp_user_lock_p lock;
1128   kmp_indirect_locktag_t type;
1129 } kmp_indirect_lock_t;
1130 
1131 // Function tables for direct locks. Set/unset/test differentiate functions
1132 // with/without consistency checking.
1133 extern void (*__kmp_direct_init[])(kmp_dyna_lock_t *, kmp_dyna_lockseq_t);
1134 extern void (**__kmp_direct_destroy)(kmp_dyna_lock_t *);
1135 extern int (**__kmp_direct_set)(kmp_dyna_lock_t *, kmp_int32);
1136 extern int (**__kmp_direct_unset)(kmp_dyna_lock_t *, kmp_int32);
1137 extern int (**__kmp_direct_test)(kmp_dyna_lock_t *, kmp_int32);
1138 
1139 // Function tables for indirect locks. Set/unset/test differentiate functions
1140 // with/without consistency checking.
1141 extern void (*__kmp_indirect_init[])(kmp_user_lock_p);
1142 extern void (**__kmp_indirect_destroy)(kmp_user_lock_p);
1143 extern int (**__kmp_indirect_set)(kmp_user_lock_p, kmp_int32);
1144 extern int (**__kmp_indirect_unset)(kmp_user_lock_p, kmp_int32);
1145 extern int (**__kmp_indirect_test)(kmp_user_lock_p, kmp_int32);
1146 
1147 // Extracts direct lock tag from a user lock pointer
1148 #define KMP_EXTRACT_D_TAG(l)                                                   \
1149   ((kmp_dyna_lock_t)((kmp_base_tas_lock_t *)(l))->poll &                       \
1150    ((1 << KMP_LOCK_SHIFT) - 1) &                                               \
1151    -((kmp_dyna_lock_t)((kmp_tas_lock_t *)(l))->lk.poll & 1))
1152 
1153 // Extracts indirect lock index from a user lock pointer
1154 #define KMP_EXTRACT_I_INDEX(l)                                                 \
1155   ((kmp_lock_index_t)((kmp_base_tas_lock_t *)(l))->poll >> 1)
1156 
1157 // Returns function pointer to the direct lock function with l (kmp_dyna_lock_t
1158 // *) and op (operation type).
1159 #define KMP_D_LOCK_FUNC(l, op) __kmp_direct_##op[KMP_EXTRACT_D_TAG(l)]
1160 
1161 // Returns function pointer to the indirect lock function with l
1162 // (kmp_indirect_lock_t *) and op (operation type).
1163 #define KMP_I_LOCK_FUNC(l, op)                                                 \
1164   __kmp_indirect_##op[((kmp_indirect_lock_t *)(l))->type]
1165 
1166 // Initializes a direct lock with the given lock pointer and lock sequence.
1167 #define KMP_INIT_D_LOCK(l, seq)                                                \
1168   __kmp_direct_init[KMP_GET_D_TAG(seq)]((kmp_dyna_lock_t *)l, seq)
1169 
1170 // Initializes an indirect lock with the given lock pointer and lock sequence.
1171 #define KMP_INIT_I_LOCK(l, seq)                                                \
1172   __kmp_direct_init[0]((kmp_dyna_lock_t *)(l), seq)
1173 
1174 // Returns "free" lock value for the given lock type.
1175 #define KMP_LOCK_FREE(type) (locktag_##type)
1176 
1177 // Returns "busy" lock value for the given lock teyp.
1178 #define KMP_LOCK_BUSY(v, type) ((v) << KMP_LOCK_SHIFT | locktag_##type)
1179 
1180 // Returns lock value after removing (shifting) lock tag.
1181 #define KMP_LOCK_STRIP(v) ((v) >> KMP_LOCK_SHIFT)
1182 
1183 // Initializes global states and data structures for managing dynamic user
1184 // locks.
1185 extern void __kmp_init_dynamic_user_locks();
1186 
1187 // Allocates and returns an indirect lock with the given indirect lock tag.
1188 extern kmp_indirect_lock_t *
1189 __kmp_allocate_indirect_lock(void **, kmp_int32, kmp_indirect_locktag_t);
1190 
1191 // Cleans up global states and data structures for managing dynamic user locks.
1192 extern void __kmp_cleanup_indirect_user_locks();
1193 
1194 // Default user lock sequence when not using hinted locks.
1195 extern kmp_dyna_lockseq_t __kmp_user_lock_seq;
1196 
1197 // Jump table for "set lock location", available only for indirect locks.
1198 extern void (*__kmp_indirect_set_location[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
1199                                                             const ident_t *);
1200 #define KMP_SET_I_LOCK_LOCATION(lck, loc)                                      \
1201   {                                                                            \
1202     if (__kmp_indirect_set_location[(lck)->type] != NULL)                      \
1203       __kmp_indirect_set_location[(lck)->type]((lck)->lock, loc);              \
1204   }
1205 
1206 // Jump table for "set lock flags", available only for indirect locks.
1207 extern void (*__kmp_indirect_set_flags[KMP_NUM_I_LOCKS])(kmp_user_lock_p,
1208                                                          kmp_lock_flags_t);
1209 #define KMP_SET_I_LOCK_FLAGS(lck, flag)                                        \
1210   {                                                                            \
1211     if (__kmp_indirect_set_flags[(lck)->type] != NULL)                         \
1212       __kmp_indirect_set_flags[(lck)->type]((lck)->lock, flag);                \
1213   }
1214 
1215 // Jump table for "get lock location", available only for indirect locks.
1216 extern const ident_t *(*__kmp_indirect_get_location[KMP_NUM_I_LOCKS])(
1217     kmp_user_lock_p);
1218 #define KMP_GET_I_LOCK_LOCATION(lck)                                           \
1219   (__kmp_indirect_get_location[(lck)->type] != NULL                            \
1220        ? __kmp_indirect_get_location[(lck)->type]((lck)->lock)                 \
1221        : NULL)
1222 
1223 // Jump table for "get lock flags", available only for indirect locks.
1224 extern kmp_lock_flags_t (*__kmp_indirect_get_flags[KMP_NUM_I_LOCKS])(
1225     kmp_user_lock_p);
1226 #define KMP_GET_I_LOCK_FLAGS(lck)                                              \
1227   (__kmp_indirect_get_flags[(lck)->type] != NULL                               \
1228        ? __kmp_indirect_get_flags[(lck)->type]((lck)->lock)                    \
1229        : NULL)
1230 
1231 // number of kmp_indirect_lock_t objects to be allocated together
1232 #define KMP_I_LOCK_CHUNK 1024
1233 // Keep at a power of 2 since it is used in multiplication & division
1234 KMP_BUILD_ASSERT(KMP_I_LOCK_CHUNK % 2 == 0);
1235 // number of row entries in the initial lock table
1236 #define KMP_I_LOCK_TABLE_INIT_NROW_PTRS 8
1237 
1238 // Lock table for indirect locks.
1239 typedef struct kmp_indirect_lock_table {
1240   kmp_indirect_lock_t **table; // blocks of indirect locks allocated
1241   kmp_uint32 nrow_ptrs; // number *table pointer entries in table
1242   kmp_lock_index_t next; // index to the next lock to be allocated
1243   struct kmp_indirect_lock_table *next_table;
1244 } kmp_indirect_lock_table_t;
1245 
1246 extern kmp_indirect_lock_table_t __kmp_i_lock_table;
1247 
1248 // Returns the indirect lock associated with the given index.
1249 // Returns nullptr if no lock at given index
1250 static inline kmp_indirect_lock_t *__kmp_get_i_lock(kmp_lock_index_t idx) {
1251   kmp_indirect_lock_table_t *lock_table = &__kmp_i_lock_table;
1252   while (lock_table) {
1253     kmp_lock_index_t max_locks = lock_table->nrow_ptrs * KMP_I_LOCK_CHUNK;
1254     if (idx < max_locks) {
1255       kmp_lock_index_t row = idx / KMP_I_LOCK_CHUNK;
1256       kmp_lock_index_t col = idx % KMP_I_LOCK_CHUNK;
1257       if (!lock_table->table[row] || idx >= lock_table->next)
1258         break;
1259       return &lock_table->table[row][col];
1260     }
1261     idx -= max_locks;
1262     lock_table = lock_table->next_table;
1263   }
1264   return nullptr;
1265 }
1266 
1267 // Number of locks in a lock block, which is fixed to "1" now.
1268 // TODO: No lock block implementation now. If we do support, we need to manage
1269 // lock block data structure for each indirect lock type.
1270 extern int __kmp_num_locks_in_block;
1271 
1272 // Fast lock table lookup without consistency checking
1273 #define KMP_LOOKUP_I_LOCK(l)                                                   \
1274   ((OMP_LOCK_T_SIZE < sizeof(void *))                                          \
1275        ? __kmp_get_i_lock(KMP_EXTRACT_I_INDEX(l))                              \
1276        : *((kmp_indirect_lock_t **)(l)))
1277 
1278 // Used once in kmp_error.cpp
1279 extern kmp_int32 __kmp_get_user_lock_owner(kmp_user_lock_p, kmp_uint32);
1280 
1281 #else // KMP_USE_DYNAMIC_LOCK
1282 
1283 #define KMP_LOCK_BUSY(v, type) (v)
1284 #define KMP_LOCK_FREE(type) 0
1285 #define KMP_LOCK_STRIP(v) (v)
1286 
1287 #endif // KMP_USE_DYNAMIC_LOCK
1288 
1289 // data structure for using backoff within spin locks.
1290 typedef struct {
1291   kmp_uint32 step; // current step
1292   kmp_uint32 max_backoff; // upper bound of outer delay loop
1293   kmp_uint32 min_tick; // size of inner delay loop in ticks (machine-dependent)
1294 } kmp_backoff_t;
1295 
1296 // Runtime's default backoff parameters
1297 extern kmp_backoff_t __kmp_spin_backoff_params;
1298 
1299 // Backoff function
1300 extern void __kmp_spin_backoff(kmp_backoff_t *);
1301 
1302 #ifdef __cplusplus
1303 } // extern "C"
1304 #endif // __cplusplus
1305 
1306 #endif /* KMP_LOCK_H */
1307