Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * lwlock.c
4 : * Lightweight lock manager
5 : *
6 : * Lightweight locks are intended primarily to provide mutual exclusion of
7 : * access to shared-memory data structures. Therefore, they offer both
8 : * exclusive and shared lock modes (to support read/write and read-only
9 : * access to a shared object). There are few other frammishes. User-level
10 : * locking should be done with the full lock manager --- which depends on
11 : * LWLocks to protect its shared state.
12 : *
13 : * In addition to exclusive and shared modes, lightweight locks can be used to
14 : * wait until a variable changes value. The variable is initially not set
15 : * when the lock is acquired with LWLockAcquire, i.e. it remains set to the
16 : * value it was set to when the lock was released last, and can be updated
17 : * without releasing the lock by calling LWLockUpdateVar. LWLockWaitForVar
18 : * waits for the variable to be updated, or until the lock is free. When
19 : * releasing the lock with LWLockReleaseClearVar() the value can be set to an
20 : * appropriate value for a free lock. The meaning of the variable is up to
21 : * the caller, the lightweight lock code just assigns and compares it.
22 : *
23 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
24 : * Portions Copyright (c) 1994, Regents of the University of California
25 : *
26 : * IDENTIFICATION
27 : * src/backend/storage/lmgr/lwlock.c
28 : *
29 : * NOTES:
30 : *
31 : * This used to be a pretty straight forward reader-writer lock
32 : * implementation, in which the internal state was protected by a
33 : * spinlock. Unfortunately the overhead of taking the spinlock proved to be
34 : * too high for workloads/locks that were taken in shared mode very
35 : * frequently. Often we were spinning in the (obviously exclusive) spinlock,
36 : * while trying to acquire a shared lock that was actually free.
37 : *
38 : * Thus a new implementation was devised that provides wait-free shared lock
39 : * acquisition for locks that aren't exclusively locked.
40 : *
41 : * The basic idea is to have a single atomic variable 'lockcount' instead of
42 : * the formerly separate shared and exclusive counters and to use atomic
43 : * operations to acquire the lock. That's fairly easy to do for plain
44 : * rw-spinlocks, but a lot harder for something like LWLocks that want to wait
45 : * in the OS.
46 : *
47 : * For lock acquisition we use an atomic compare-and-exchange on the lockcount
48 : * variable. For exclusive lock we swap in a sentinel value
49 : * (LW_VAL_EXCLUSIVE), for shared locks we count the number of holders.
50 : *
51 : * To release the lock we use an atomic decrement to release the lock. If the
52 : * new value is zero (we get that atomically), we know we can/have to release
53 : * waiters.
54 : *
55 : * Obviously it is important that the sentinel value for exclusive locks
56 : * doesn't conflict with the maximum number of possible share lockers -
57 : * luckily MAX_BACKENDS makes that easily possible.
58 : *
59 : *
60 : * The attentive reader might have noticed that naively doing the above has a
61 : * glaring race condition: We try to lock using the atomic operations and
62 : * notice that we have to wait. Unfortunately by the time we have finished
63 : * queuing, the former locker very well might have already finished its
64 : * work. That's problematic because we're now stuck waiting inside the OS.
65 :
66 : * To mitigate those races we use a two phased attempt at locking:
67 : * Phase 1: Try to do it atomically, if we succeed, nice
68 : * Phase 2: Add ourselves to the waitqueue of the lock
69 : * Phase 3: Try to grab the lock again, if we succeed, remove ourselves from
70 : * the queue
71 : * Phase 4: Sleep till wake-up, goto Phase 1
72 : *
73 : * This protects us against the problem from above as nobody can release too
74 : * quick, before we're queued, since after Phase 2 we're already queued.
75 : * -------------------------------------------------------------------------
76 : */
77 : #include "postgres.h"
78 :
79 : #include "miscadmin.h"
80 : #include "pg_trace.h"
81 : #include "pgstat.h"
82 : #include "port/pg_bitutils.h"
83 : #include "storage/proc.h"
84 : #include "storage/proclist.h"
85 : #include "storage/procnumber.h"
86 : #include "storage/spin.h"
87 : #include "utils/memutils.h"
88 :
89 : #ifdef LWLOCK_STATS
90 : #include "utils/hsearch.h"
91 : #endif
92 :
93 :
94 : #define LW_FLAG_HAS_WAITERS ((uint32) 1 << 31)
95 : #define LW_FLAG_RELEASE_OK ((uint32) 1 << 30)
96 : #define LW_FLAG_LOCKED ((uint32) 1 << 29)
97 : #define LW_FLAG_BITS 3
98 : #define LW_FLAG_MASK (((1<<LW_FLAG_BITS)-1)<<(32-LW_FLAG_BITS))
99 :
100 : /* assumes MAX_BACKENDS is a (power of 2) - 1, checked below */
101 : #define LW_VAL_EXCLUSIVE (MAX_BACKENDS + 1)
102 : #define LW_VAL_SHARED 1
103 :
104 : /* already (power of 2)-1, i.e. suitable for a mask */
105 : #define LW_SHARED_MASK MAX_BACKENDS
106 : #define LW_LOCK_MASK (MAX_BACKENDS | LW_VAL_EXCLUSIVE)
107 :
108 :
109 : StaticAssertDecl(((MAX_BACKENDS + 1) & MAX_BACKENDS) == 0,
110 : "MAX_BACKENDS + 1 needs to be a power of 2");
111 :
112 : StaticAssertDecl((MAX_BACKENDS & LW_FLAG_MASK) == 0,
113 : "MAX_BACKENDS and LW_FLAG_MASK overlap");
114 :
115 : StaticAssertDecl((LW_VAL_EXCLUSIVE & LW_FLAG_MASK) == 0,
116 : "LW_VAL_EXCLUSIVE and LW_FLAG_MASK overlap");
117 :
118 : /*
119 : * There are three sorts of LWLock "tranches":
120 : *
121 : * 1. The individually-named locks defined in lwlocklist.h each have their
122 : * own tranche. We absorb the names of these tranches from there into
123 : * BuiltinTrancheNames here.
124 : *
125 : * 2. There are some predefined tranches for built-in groups of locks defined
126 : * in lwlocklist.h. We absorb the names of these tranches, too.
127 : *
128 : * 3. Extensions can create new tranches, via either RequestNamedLWLockTranche
129 : * or LWLockNewTrancheId. These names are stored in shared memory and can be
130 : * accessed via LWLockTrancheNames.
131 : *
132 : * All these names are user-visible as wait event names, so choose with care
133 : * ... and do not forget to update the documentation's list of wait events.
134 : */
135 : static const char *const BuiltinTrancheNames[] = {
136 : #define PG_LWLOCK(id, lockname) [id] = CppAsString(lockname),
137 : #define PG_LWLOCKTRANCHE(id, lockname) [LWTRANCHE_##id] = CppAsString(lockname),
138 : #include "storage/lwlocklist.h"
139 : #undef PG_LWLOCK
140 : #undef PG_LWLOCKTRANCHE
141 : };
142 :
143 : StaticAssertDecl(lengthof(BuiltinTrancheNames) ==
144 : LWTRANCHE_FIRST_USER_DEFINED,
145 : "missing entries in BuiltinTrancheNames[]");
146 :
147 : /*
148 : * This is indexed by tranche ID minus LWTRANCHE_FIRST_USER_DEFINED, and
149 : * points to the shared memory locations of the names of all
150 : * dynamically-created tranches. Backends inherit the pointer by fork from the
151 : * postmaster (except in the EXEC_BACKEND case, where we have special measures
152 : * to pass it down).
153 : */
154 : char **LWLockTrancheNames = NULL;
155 :
156 : /*
157 : * This points to the main array of LWLocks in shared memory. Backends inherit
158 : * the pointer by fork from the postmaster (except in the EXEC_BACKEND case,
159 : * where we have special measures to pass it down).
160 : */
161 : LWLockPadded *MainLWLockArray = NULL;
162 :
163 : /*
164 : * We use this structure to keep track of locked LWLocks for release
165 : * during error recovery. Normally, only a few will be held at once, but
166 : * occasionally the number can be much higher.
167 : */
168 : #define MAX_SIMUL_LWLOCKS 200
169 :
170 : /* struct representing the LWLocks we're holding */
171 : typedef struct LWLockHandle
172 : {
173 : LWLock *lock;
174 : LWLockMode mode;
175 : } LWLockHandle;
176 :
177 : static int num_held_lwlocks = 0;
178 : static LWLockHandle held_lwlocks[MAX_SIMUL_LWLOCKS];
179 :
180 : /* struct representing the LWLock tranche request for named tranche */
181 : typedef struct NamedLWLockTrancheRequest
182 : {
183 : char tranche_name[NAMEDATALEN];
184 : int num_lwlocks;
185 : } NamedLWLockTrancheRequest;
186 :
187 : /*
188 : * NamedLWLockTrancheRequests is the valid length of the request array. These
189 : * variables are non-static so that launch_backend.c can copy them to child
190 : * processes in EXEC_BACKEND builds.
191 : */
192 : int NamedLWLockTrancheRequests = 0;
193 : NamedLWLockTrancheRequest *NamedLWLockTrancheRequestArray = NULL;
194 :
195 : /* postmaster's local copy of the request array */
196 : static NamedLWLockTrancheRequest *LocalNamedLWLockTrancheRequestArray = NULL;
197 :
198 : /* shared memory counter of registered tranches */
199 : int *LWLockCounter = NULL;
200 :
201 : /* backend-local counter of registered tranches */
202 : static int LocalLWLockCounter;
203 :
204 : #define MAX_NAMED_TRANCHES 256
205 :
206 : static void InitializeLWLocks(void);
207 : static inline void LWLockReportWaitStart(LWLock *lock);
208 : static inline void LWLockReportWaitEnd(void);
209 : static const char *GetLWTrancheName(uint16 trancheId);
210 :
211 : #define T_NAME(lock) \
212 : GetLWTrancheName((lock)->tranche)
213 :
214 : #ifdef LWLOCK_STATS
215 : typedef struct lwlock_stats_key
216 : {
217 : int tranche;
218 : void *instance;
219 : } lwlock_stats_key;
220 :
221 : typedef struct lwlock_stats
222 : {
223 : lwlock_stats_key key;
224 : int sh_acquire_count;
225 : int ex_acquire_count;
226 : int block_count;
227 : int dequeue_self_count;
228 : int spin_delay_count;
229 : } lwlock_stats;
230 :
231 : static HTAB *lwlock_stats_htab;
232 : static lwlock_stats lwlock_stats_dummy;
233 : #endif
234 :
235 : #ifdef LOCK_DEBUG
236 : bool Trace_lwlocks = false;
237 :
238 : inline static void
239 : PRINT_LWDEBUG(const char *where, LWLock *lock, LWLockMode mode)
240 : {
241 : /* hide statement & context here, otherwise the log is just too verbose */
242 : if (Trace_lwlocks)
243 : {
244 : uint32 state = pg_atomic_read_u32(&lock->state);
245 :
246 : ereport(LOG,
247 : (errhidestmt(true),
248 : errhidecontext(true),
249 : errmsg_internal("%d: %s(%s %p): excl %u shared %u haswaiters %u waiters %u rOK %d",
250 : MyProcPid,
251 : where, T_NAME(lock), lock,
252 : (state & LW_VAL_EXCLUSIVE) != 0,
253 : state & LW_SHARED_MASK,
254 : (state & LW_FLAG_HAS_WAITERS) != 0,
255 : pg_atomic_read_u32(&lock->nwaiters),
256 : (state & LW_FLAG_RELEASE_OK) != 0)));
257 : }
258 : }
259 :
260 : inline static void
261 : LOG_LWDEBUG(const char *where, LWLock *lock, const char *msg)
262 : {
263 : /* hide statement & context here, otherwise the log is just too verbose */
264 : if (Trace_lwlocks)
265 : {
266 : ereport(LOG,
267 : (errhidestmt(true),
268 : errhidecontext(true),
269 : errmsg_internal("%s(%s %p): %s", where,
270 : T_NAME(lock), lock, msg)));
271 : }
272 : }
273 :
274 : #else /* not LOCK_DEBUG */
275 : #define PRINT_LWDEBUG(a,b,c) ((void)0)
276 : #define LOG_LWDEBUG(a,b,c) ((void)0)
277 : #endif /* LOCK_DEBUG */
278 :
279 : #ifdef LWLOCK_STATS
280 :
281 : static void init_lwlock_stats(void);
282 : static void print_lwlock_stats(int code, Datum arg);
283 : static lwlock_stats * get_lwlock_stats_entry(LWLock *lock);
284 :
285 : static void
286 : init_lwlock_stats(void)
287 : {
288 : HASHCTL ctl;
289 : static MemoryContext lwlock_stats_cxt = NULL;
290 : static bool exit_registered = false;
291 :
292 : if (lwlock_stats_cxt != NULL)
293 : MemoryContextDelete(lwlock_stats_cxt);
294 :
295 : /*
296 : * The LWLock stats will be updated within a critical section, which
297 : * requires allocating new hash entries. Allocations within a critical
298 : * section are normally not allowed because running out of memory would
299 : * lead to a PANIC, but LWLOCK_STATS is debugging code that's not normally
300 : * turned on in production, so that's an acceptable risk. The hash entries
301 : * are small, so the risk of running out of memory is minimal in practice.
302 : */
303 : lwlock_stats_cxt = AllocSetContextCreate(TopMemoryContext,
304 : "LWLock stats",
305 : ALLOCSET_DEFAULT_SIZES);
306 : MemoryContextAllowInCriticalSection(lwlock_stats_cxt, true);
307 :
308 : ctl.keysize = sizeof(lwlock_stats_key);
309 : ctl.entrysize = sizeof(lwlock_stats);
310 : ctl.hcxt = lwlock_stats_cxt;
311 : lwlock_stats_htab = hash_create("lwlock stats", 16384, &ctl,
312 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
313 : if (!exit_registered)
314 : {
315 : on_shmem_exit(print_lwlock_stats, 0);
316 : exit_registered = true;
317 : }
318 : }
319 :
320 : static void
321 : print_lwlock_stats(int code, Datum arg)
322 : {
323 : HASH_SEQ_STATUS scan;
324 : lwlock_stats *lwstats;
325 :
326 : hash_seq_init(&scan, lwlock_stats_htab);
327 :
328 : /* Grab an LWLock to keep different backends from mixing reports */
329 : LWLockAcquire(&MainLWLockArray[0].lock, LW_EXCLUSIVE);
330 :
331 : while ((lwstats = (lwlock_stats *) hash_seq_search(&scan)) != NULL)
332 : {
333 : fprintf(stderr,
334 : "PID %d lwlock %s %p: shacq %u exacq %u blk %u spindelay %u dequeue self %u\n",
335 : MyProcPid, GetLWTrancheName(lwstats->key.tranche),
336 : lwstats->key.instance, lwstats->sh_acquire_count,
337 : lwstats->ex_acquire_count, lwstats->block_count,
338 : lwstats->spin_delay_count, lwstats->dequeue_self_count);
339 : }
340 :
341 : LWLockRelease(&MainLWLockArray[0].lock);
342 : }
343 :
344 : static lwlock_stats *
345 : get_lwlock_stats_entry(LWLock *lock)
346 : {
347 : lwlock_stats_key key;
348 : lwlock_stats *lwstats;
349 : bool found;
350 :
351 : /*
352 : * During shared memory initialization, the hash table doesn't exist yet.
353 : * Stats of that phase aren't very interesting, so just collect operations
354 : * on all locks in a single dummy entry.
355 : */
356 : if (lwlock_stats_htab == NULL)
357 : return &lwlock_stats_dummy;
358 :
359 : /* Fetch or create the entry. */
360 : MemSet(&key, 0, sizeof(key));
361 : key.tranche = lock->tranche;
362 : key.instance = lock;
363 : lwstats = hash_search(lwlock_stats_htab, &key, HASH_ENTER, &found);
364 : if (!found)
365 : {
366 : lwstats->sh_acquire_count = 0;
367 : lwstats->ex_acquire_count = 0;
368 : lwstats->block_count = 0;
369 : lwstats->dequeue_self_count = 0;
370 : lwstats->spin_delay_count = 0;
371 : }
372 : return lwstats;
373 : }
374 : #endif /* LWLOCK_STATS */
375 :
376 :
377 : /*
378 : * Compute number of LWLocks required by named tranches. These will be
379 : * allocated in the main array.
380 : */
381 : static int
382 6240 : NumLWLocksForNamedTranches(void)
383 : {
384 6240 : int numLocks = 0;
385 : int i;
386 :
387 6474 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
388 234 : numLocks += NamedLWLockTrancheRequestArray[i].num_lwlocks;
389 :
390 6240 : return numLocks;
391 : }
392 :
393 : /*
394 : * Compute shmem space needed for LWLocks and named tranches.
395 : */
396 : Size
397 6240 : LWLockShmemSize(void)
398 : {
399 : Size size;
400 6240 : int numLocks = NUM_FIXED_LWLOCKS;
401 :
402 : /*
403 : * If re-initializing shared memory, the request array will no longer be
404 : * accessible, so switch to the copy in postmaster's local memory. We'll
405 : * copy it back into shared memory later when CreateLWLocks() is called
406 : * again.
407 : */
408 6240 : if (LocalNamedLWLockTrancheRequestArray)
409 0 : NamedLWLockTrancheRequestArray = LocalNamedLWLockTrancheRequestArray;
410 :
411 : /* Calculate total number of locks needed in the main array. */
412 6240 : numLocks += NumLWLocksForNamedTranches();
413 :
414 : /* Space for dynamic allocation counter. */
415 6240 : size = MAXALIGN(sizeof(int));
416 :
417 : /* Space for named tranches. */
418 6240 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, sizeof(char *)));
419 6240 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, NAMEDATALEN));
420 :
421 : /*
422 : * Make space for named tranche requests. This is done for the benefit of
423 : * EXEC_BACKEND builds, which otherwise wouldn't be able to call
424 : * GetNamedLWLockTranche() outside postmaster.
425 : */
426 6240 : size = add_size(size, mul_size(NamedLWLockTrancheRequests,
427 : sizeof(NamedLWLockTrancheRequest)));
428 :
429 : /* Space for the LWLock array, plus room for cache line alignment. */
430 6240 : size = add_size(size, LWLOCK_PADDED_SIZE);
431 6240 : size = add_size(size, mul_size(numLocks, sizeof(LWLockPadded)));
432 :
433 6240 : return size;
434 : }
435 :
436 : /*
437 : * Allocate shmem space for the main LWLock array and all tranches and
438 : * initialize it.
439 : */
440 : void
441 2180 : CreateLWLocks(void)
442 : {
443 2180 : if (!IsUnderPostmaster)
444 : {
445 2180 : Size spaceLocks = LWLockShmemSize();
446 : char *ptr;
447 :
448 : /* Allocate space */
449 2180 : ptr = (char *) ShmemAlloc(spaceLocks);
450 :
451 : /* Initialize the dynamic-allocation counter for tranches */
452 2180 : LWLockCounter = (int *) ptr;
453 2180 : *LWLockCounter = LWTRANCHE_FIRST_USER_DEFINED;
454 2180 : ptr += MAXALIGN(sizeof(int));
455 :
456 : /* Initialize tranche names */
457 2180 : LWLockTrancheNames = (char **) ptr;
458 2180 : ptr += MAX_NAMED_TRANCHES * sizeof(char *);
459 560260 : for (int i = 0; i < MAX_NAMED_TRANCHES; i++)
460 : {
461 558080 : LWLockTrancheNames[i] = ptr;
462 558080 : ptr += NAMEDATALEN;
463 : }
464 :
465 : /*
466 : * Move named tranche requests to shared memory. This is done for the
467 : * benefit of EXEC_BACKEND builds, which otherwise wouldn't be able to
468 : * call GetNamedLWLockTranche() outside postmaster.
469 : */
470 2180 : if (NamedLWLockTrancheRequests > 0)
471 : {
472 : /*
473 : * Save the pointer to the request array in postmaster's local
474 : * memory. We'll need it if we ever need to re-initialize shared
475 : * memory after a crash.
476 : */
477 16 : LocalNamedLWLockTrancheRequestArray = NamedLWLockTrancheRequestArray;
478 :
479 16 : memcpy(ptr, NamedLWLockTrancheRequestArray,
480 : NamedLWLockTrancheRequests * sizeof(NamedLWLockTrancheRequest));
481 16 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *) ptr;
482 16 : ptr += NamedLWLockTrancheRequests * sizeof(NamedLWLockTrancheRequest);
483 : }
484 :
485 : /* Ensure desired alignment of LWLock array */
486 2180 : ptr += LWLOCK_PADDED_SIZE - ((uintptr_t) ptr) % LWLOCK_PADDED_SIZE;
487 2180 : MainLWLockArray = (LWLockPadded *) ptr;
488 :
489 : /* Initialize all LWLocks */
490 2180 : InitializeLWLocks();
491 : }
492 2180 : }
493 :
494 : /*
495 : * Initialize LWLocks that are fixed and those belonging to named tranches.
496 : */
497 : static void
498 2180 : InitializeLWLocks(void)
499 : {
500 : int id;
501 : int i;
502 : int j;
503 : LWLockPadded *lock;
504 :
505 : /* Initialize all individual LWLocks in main array */
506 119900 : for (id = 0, lock = MainLWLockArray; id < NUM_INDIVIDUAL_LWLOCKS; id++, lock++)
507 117720 : LWLockInitialize(&lock->lock, id);
508 :
509 : /* Initialize buffer mapping LWLocks in main array */
510 2180 : lock = MainLWLockArray + BUFFER_MAPPING_LWLOCK_OFFSET;
511 281220 : for (id = 0; id < NUM_BUFFER_PARTITIONS; id++, lock++)
512 279040 : LWLockInitialize(&lock->lock, LWTRANCHE_BUFFER_MAPPING);
513 :
514 : /* Initialize lmgrs' LWLocks in main array */
515 2180 : lock = MainLWLockArray + LOCK_MANAGER_LWLOCK_OFFSET;
516 37060 : for (id = 0; id < NUM_LOCK_PARTITIONS; id++, lock++)
517 34880 : LWLockInitialize(&lock->lock, LWTRANCHE_LOCK_MANAGER);
518 :
519 : /* Initialize predicate lmgrs' LWLocks in main array */
520 2180 : lock = MainLWLockArray + PREDICATELOCK_MANAGER_LWLOCK_OFFSET;
521 37060 : for (id = 0; id < NUM_PREDICATELOCK_PARTITIONS; id++, lock++)
522 34880 : LWLockInitialize(&lock->lock, LWTRANCHE_PREDICATE_LOCK_MANAGER);
523 :
524 : /*
525 : * Copy the info about any named tranches into shared memory (so that
526 : * other processes can see it), and initialize the requested LWLocks.
527 : */
528 2180 : if (NamedLWLockTrancheRequests > 0)
529 : {
530 16 : lock = &MainLWLockArray[NUM_FIXED_LWLOCKS];
531 :
532 94 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
533 : {
534 : NamedLWLockTrancheRequest *request;
535 : int tranche;
536 :
537 78 : request = &NamedLWLockTrancheRequestArray[i];
538 78 : tranche = LWLockNewTrancheId(request->tranche_name);
539 :
540 156 : for (j = 0; j < request->num_lwlocks; j++, lock++)
541 78 : LWLockInitialize(&lock->lock, tranche);
542 : }
543 : }
544 2180 : }
545 :
546 : /*
547 : * InitLWLockAccess - initialize backend-local state needed to hold LWLocks
548 : */
549 : void
550 44464 : InitLWLockAccess(void)
551 : {
552 : #ifdef LWLOCK_STATS
553 : init_lwlock_stats();
554 : #endif
555 44464 : }
556 :
557 : /*
558 : * GetNamedLWLockTranche - returns the base address of LWLock from the
559 : * specified tranche.
560 : *
561 : * Caller needs to retrieve the requested number of LWLocks starting from
562 : * the base lock address returned by this API. This can be used for
563 : * tranches that are requested by using RequestNamedLWLockTranche() API.
564 : */
565 : LWLockPadded *
566 18 : GetNamedLWLockTranche(const char *tranche_name)
567 : {
568 : int lock_pos;
569 : int i;
570 :
571 : /*
572 : * Obtain the position of base address of LWLock belonging to requested
573 : * tranche_name in MainLWLockArray. LWLocks for named tranches are placed
574 : * in MainLWLockArray after fixed locks.
575 : */
576 18 : lock_pos = NUM_FIXED_LWLOCKS;
577 82 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
578 : {
579 80 : if (strcmp(NamedLWLockTrancheRequestArray[i].tranche_name,
580 : tranche_name) == 0)
581 16 : return &MainLWLockArray[lock_pos];
582 :
583 64 : lock_pos += NamedLWLockTrancheRequestArray[i].num_lwlocks;
584 : }
585 :
586 2 : elog(ERROR, "requested tranche is not registered");
587 :
588 : /* just to keep compiler quiet */
589 : return NULL;
590 : }
591 :
592 : /*
593 : * Allocate a new tranche ID with the provided name.
594 : */
595 : int
596 556 : LWLockNewTrancheId(const char *name)
597 : {
598 : int result;
599 :
600 556 : if (!name)
601 2 : ereport(ERROR,
602 : (errcode(ERRCODE_INVALID_NAME),
603 : errmsg("tranche name cannot be NULL")));
604 :
605 554 : if (strlen(name) >= NAMEDATALEN)
606 2 : ereport(ERROR,
607 : (errcode(ERRCODE_NAME_TOO_LONG),
608 : errmsg("tranche name too long"),
609 : errdetail("LWLock tranche names must be no longer than %d bytes.",
610 : NAMEDATALEN - 1)));
611 :
612 : /*
613 : * We use the ShmemLock spinlock to protect LWLockCounter and
614 : * LWLockTrancheNames.
615 : */
616 552 : SpinLockAcquire(ShmemLock);
617 :
618 552 : if (*LWLockCounter - LWTRANCHE_FIRST_USER_DEFINED >= MAX_NAMED_TRANCHES)
619 : {
620 2 : SpinLockRelease(ShmemLock);
621 2 : ereport(ERROR,
622 : (errmsg("maximum number of tranches already registered"),
623 : errdetail("No more than %d tranches may be registered.",
624 : MAX_NAMED_TRANCHES)));
625 : }
626 :
627 550 : result = (*LWLockCounter)++;
628 550 : LocalLWLockCounter = *LWLockCounter;
629 550 : strlcpy(LWLockTrancheNames[result - LWTRANCHE_FIRST_USER_DEFINED], name, NAMEDATALEN);
630 :
631 550 : SpinLockRelease(ShmemLock);
632 :
633 550 : return result;
634 : }
635 :
636 : /*
637 : * RequestNamedLWLockTranche
638 : * Request that extra LWLocks be allocated during postmaster
639 : * startup.
640 : *
641 : * This may only be called via the shmem_request_hook of a library that is
642 : * loaded into the postmaster via shared_preload_libraries. Calls from
643 : * elsewhere will fail.
644 : *
645 : * The tranche name will be user-visible as a wait event name, so try to
646 : * use a name that fits the style for those.
647 : */
648 : void
649 78 : RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
650 : {
651 : NamedLWLockTrancheRequest *request;
652 : static int NamedLWLockTrancheRequestsAllocated;
653 :
654 78 : if (!process_shmem_requests_in_progress)
655 0 : elog(FATAL, "cannot request additional LWLocks outside shmem_request_hook");
656 :
657 78 : if (!tranche_name)
658 0 : ereport(ERROR,
659 : (errcode(ERRCODE_INVALID_NAME),
660 : errmsg("tranche name cannot be NULL")));
661 :
662 78 : if (strlen(tranche_name) >= NAMEDATALEN)
663 0 : ereport(ERROR,
664 : (errcode(ERRCODE_NAME_TOO_LONG),
665 : errmsg("tranche name too long"),
666 : errdetail("LWLock tranche names must be no longer than %d bytes.",
667 : NAMEDATALEN - 1)));
668 :
669 78 : if (NamedLWLockTrancheRequestArray == NULL)
670 : {
671 16 : NamedLWLockTrancheRequestsAllocated = 16;
672 16 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
673 16 : MemoryContextAlloc(TopMemoryContext,
674 : NamedLWLockTrancheRequestsAllocated
675 : * sizeof(NamedLWLockTrancheRequest));
676 : }
677 :
678 78 : if (NamedLWLockTrancheRequests >= NamedLWLockTrancheRequestsAllocated)
679 : {
680 2 : int i = pg_nextpower2_32(NamedLWLockTrancheRequests + 1);
681 :
682 2 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
683 2 : repalloc(NamedLWLockTrancheRequestArray,
684 : i * sizeof(NamedLWLockTrancheRequest));
685 2 : NamedLWLockTrancheRequestsAllocated = i;
686 : }
687 :
688 78 : request = &NamedLWLockTrancheRequestArray[NamedLWLockTrancheRequests];
689 78 : strlcpy(request->tranche_name, tranche_name, NAMEDATALEN);
690 78 : request->num_lwlocks = num_lwlocks;
691 78 : NamedLWLockTrancheRequests++;
692 78 : }
693 :
694 : /*
695 : * LWLockInitialize - initialize a new lwlock; it's initially unlocked
696 : */
697 : void
698 24752138 : LWLockInitialize(LWLock *lock, int tranche_id)
699 : {
700 : /* verify the tranche_id is valid */
701 24752138 : (void) GetLWTrancheName(tranche_id);
702 :
703 24752136 : pg_atomic_init_u32(&lock->state, LW_FLAG_RELEASE_OK);
704 : #ifdef LOCK_DEBUG
705 : pg_atomic_init_u32(&lock->nwaiters, 0);
706 : #endif
707 24752136 : lock->tranche = tranche_id;
708 24752136 : proclist_init(&lock->waiters);
709 24752136 : }
710 :
711 : /*
712 : * Report start of wait event for light-weight locks.
713 : *
714 : * This function will be used by all the light-weight lock calls which
715 : * needs to wait to acquire the lock. This function distinguishes wait
716 : * event based on tranche and lock id.
717 : */
718 : static inline void
719 3145476 : LWLockReportWaitStart(LWLock *lock)
720 : {
721 3145476 : pgstat_report_wait_start(PG_WAIT_LWLOCK | lock->tranche);
722 3145476 : }
723 :
724 : /*
725 : * Report end of wait event for light-weight locks.
726 : */
727 : static inline void
728 3145476 : LWLockReportWaitEnd(void)
729 : {
730 3145476 : pgstat_report_wait_end();
731 3145476 : }
732 :
733 : /*
734 : * Return the name of an LWLock tranche.
735 : */
736 : static const char *
737 24752686 : GetLWTrancheName(uint16 trancheId)
738 : {
739 : /* Built-in tranche or individual LWLock? */
740 24752686 : if (trancheId < LWTRANCHE_FIRST_USER_DEFINED)
741 24751372 : return BuiltinTrancheNames[trancheId];
742 :
743 : /*
744 : * We only ever add new entries to LWLockTrancheNames, so most lookups can
745 : * avoid taking the spinlock as long as the backend-local counter
746 : * (LocalLWLockCounter) is greater than the requested tranche ID. Else,
747 : * we need to first update the backend-local counter with ShmemLock held
748 : * before attempting the lookup again. In practice, the latter case is
749 : * probably rare.
750 : */
751 1314 : if (trancheId >= LocalLWLockCounter)
752 : {
753 2 : SpinLockAcquire(ShmemLock);
754 2 : LocalLWLockCounter = *LWLockCounter;
755 2 : SpinLockRelease(ShmemLock);
756 :
757 2 : if (trancheId >= LocalLWLockCounter)
758 2 : elog(ERROR, "tranche %d is not registered", trancheId);
759 : }
760 :
761 : /*
762 : * It's an extension tranche, so look in LWLockTrancheNames.
763 : */
764 1312 : trancheId -= LWTRANCHE_FIRST_USER_DEFINED;
765 :
766 1312 : return LWLockTrancheNames[trancheId];
767 : }
768 :
769 : /*
770 : * Return an identifier for an LWLock based on the wait class and event.
771 : */
772 : const char *
773 548 : GetLWLockIdentifier(uint32 classId, uint16 eventId)
774 : {
775 : Assert(classId == PG_WAIT_LWLOCK);
776 : /* The event IDs are just tranche numbers. */
777 548 : return GetLWTrancheName(eventId);
778 : }
779 :
780 : /*
781 : * Internal function that tries to atomically acquire the lwlock in the passed
782 : * in mode.
783 : *
784 : * This function will not block waiting for a lock to become free - that's the
785 : * caller's job.
786 : *
787 : * Returns true if the lock isn't free and we need to wait.
788 : */
789 : static bool
790 756123172 : LWLockAttemptLock(LWLock *lock, LWLockMode mode)
791 : {
792 : uint32 old_state;
793 :
794 : Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
795 :
796 : /*
797 : * Read once outside the loop, later iterations will get the newer value
798 : * via compare & exchange.
799 : */
800 756123172 : old_state = pg_atomic_read_u32(&lock->state);
801 :
802 : /* loop until we've determined whether we could acquire the lock or not */
803 : while (true)
804 1076236 : {
805 : uint32 desired_state;
806 : bool lock_free;
807 :
808 757199408 : desired_state = old_state;
809 :
810 757199408 : if (mode == LW_EXCLUSIVE)
811 : {
812 464658404 : lock_free = (old_state & LW_LOCK_MASK) == 0;
813 464658404 : if (lock_free)
814 462412484 : desired_state += LW_VAL_EXCLUSIVE;
815 : }
816 : else
817 : {
818 292541004 : lock_free = (old_state & LW_VAL_EXCLUSIVE) == 0;
819 292541004 : if (lock_free)
820 288312192 : desired_state += LW_VAL_SHARED;
821 : }
822 :
823 : /*
824 : * Attempt to swap in the state we are expecting. If we didn't see
825 : * lock to be free, that's just the old value. If we saw it as free,
826 : * we'll attempt to mark it acquired. The reason that we always swap
827 : * in the value is that this doubles as a memory barrier. We could try
828 : * to be smarter and only swap in values if we saw the lock as free,
829 : * but benchmark haven't shown it as beneficial so far.
830 : *
831 : * Retry if the value changed since we last looked at it.
832 : */
833 757199408 : if (pg_atomic_compare_exchange_u32(&lock->state,
834 : &old_state, desired_state))
835 : {
836 756123172 : if (lock_free)
837 : {
838 : /* Great! Got the lock. */
839 : #ifdef LOCK_DEBUG
840 : if (mode == LW_EXCLUSIVE)
841 : lock->owner = MyProc;
842 : #endif
843 749952560 : return false;
844 : }
845 : else
846 6170612 : return true; /* somebody else has the lock */
847 : }
848 : }
849 : pg_unreachable();
850 : }
851 :
852 : /*
853 : * Lock the LWLock's wait list against concurrent activity.
854 : *
855 : * NB: even though the wait list is locked, non-conflicting lock operations
856 : * may still happen concurrently.
857 : *
858 : * Time spent holding mutex should be short!
859 : */
860 : static void
861 12967796 : LWLockWaitListLock(LWLock *lock)
862 : {
863 : uint32 old_state;
864 : #ifdef LWLOCK_STATS
865 : lwlock_stats *lwstats;
866 : uint32 delays = 0;
867 :
868 : lwstats = get_lwlock_stats_entry(lock);
869 : #endif
870 :
871 : while (true)
872 : {
873 : /* always try once to acquire lock directly */
874 13243308 : old_state = pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_LOCKED);
875 13243308 : if (!(old_state & LW_FLAG_LOCKED))
876 12967796 : break; /* got lock */
877 :
878 : /* and then spin without atomic operations until lock is released */
879 : {
880 : SpinDelayStatus delayStatus;
881 :
882 275512 : init_local_spin_delay(&delayStatus);
883 :
884 810044 : while (old_state & LW_FLAG_LOCKED)
885 : {
886 534532 : perform_spin_delay(&delayStatus);
887 534532 : old_state = pg_atomic_read_u32(&lock->state);
888 : }
889 : #ifdef LWLOCK_STATS
890 : delays += delayStatus.delays;
891 : #endif
892 275512 : finish_spin_delay(&delayStatus);
893 : }
894 :
895 : /*
896 : * Retry. The lock might obviously already be re-acquired by the time
897 : * we're attempting to get it again.
898 : */
899 : }
900 :
901 : #ifdef LWLOCK_STATS
902 : lwstats->spin_delay_count += delays;
903 : #endif
904 12967796 : }
905 :
906 : /*
907 : * Unlock the LWLock's wait list.
908 : *
909 : * Note that it can be more efficient to manipulate flags and release the
910 : * locks in a single atomic operation.
911 : */
912 : static void
913 8969364 : LWLockWaitListUnlock(LWLock *lock)
914 : {
915 : uint32 old_state PG_USED_FOR_ASSERTS_ONLY;
916 :
917 8969364 : old_state = pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_LOCKED);
918 :
919 : Assert(old_state & LW_FLAG_LOCKED);
920 8969364 : }
921 :
922 : /*
923 : * Wakeup all the lockers that currently have a chance to acquire the lock.
924 : */
925 : static void
926 3998432 : LWLockWakeup(LWLock *lock)
927 : {
928 : bool new_release_ok;
929 3998432 : bool wokeup_somebody = false;
930 : proclist_head wakeup;
931 : proclist_mutable_iter iter;
932 :
933 3998432 : proclist_init(&wakeup);
934 :
935 3998432 : new_release_ok = true;
936 :
937 : /* lock wait list while collecting backends to wake up */
938 3998432 : LWLockWaitListLock(lock);
939 :
940 6269492 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
941 : {
942 3245250 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
943 :
944 3245250 : if (wokeup_somebody && waiter->lwWaitMode == LW_EXCLUSIVE)
945 22628 : continue;
946 :
947 3222622 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
948 3222622 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
949 :
950 3222622 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
951 : {
952 : /*
953 : * Prevent additional wakeups until retryer gets to run. Backends
954 : * that are just waiting for the lock to become free don't retry
955 : * automatically.
956 : */
957 3001480 : new_release_ok = false;
958 :
959 : /*
960 : * Don't wakeup (further) exclusive locks.
961 : */
962 3001480 : wokeup_somebody = true;
963 : }
964 :
965 : /*
966 : * Signal that the process isn't on the wait list anymore. This allows
967 : * LWLockDequeueSelf() to remove itself of the waitlist with a
968 : * proclist_delete(), rather than having to check if it has been
969 : * removed from the list.
970 : */
971 : Assert(waiter->lwWaiting == LW_WS_WAITING);
972 3222622 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
973 :
974 : /*
975 : * Once we've woken up an exclusive lock, there's no point in waking
976 : * up anybody else.
977 : */
978 3222622 : if (waiter->lwWaitMode == LW_EXCLUSIVE)
979 974190 : break;
980 : }
981 :
982 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS);
983 :
984 : /* unset required flags, and release lock, in one fell swoop */
985 : {
986 : uint32 old_state;
987 : uint32 desired_state;
988 :
989 3998432 : old_state = pg_atomic_read_u32(&lock->state);
990 : while (true)
991 : {
992 4149650 : desired_state = old_state;
993 :
994 : /* compute desired flags */
995 :
996 4149650 : if (new_release_ok)
997 1332714 : desired_state |= LW_FLAG_RELEASE_OK;
998 : else
999 2816936 : desired_state &= ~LW_FLAG_RELEASE_OK;
1000 :
1001 4149650 : if (proclist_is_empty(&wakeup))
1002 1231154 : desired_state &= ~LW_FLAG_HAS_WAITERS;
1003 :
1004 4149650 : desired_state &= ~LW_FLAG_LOCKED; /* release lock */
1005 :
1006 4149650 : if (pg_atomic_compare_exchange_u32(&lock->state, &old_state,
1007 : desired_state))
1008 3998432 : break;
1009 : }
1010 : }
1011 :
1012 : /* Awaken any waiters I removed from the queue. */
1013 7221054 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1014 : {
1015 3222622 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1016 :
1017 : LOG_LWDEBUG("LWLockRelease", lock, "release waiter");
1018 3222622 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1019 :
1020 : /*
1021 : * Guarantee that lwWaiting being unset only becomes visible once the
1022 : * unlink from the link has completed. Otherwise the target backend
1023 : * could be woken up for other reason and enqueue for a new lock - if
1024 : * that happens before the list unlink happens, the list would end up
1025 : * being corrupted.
1026 : *
1027 : * The barrier pairs with the LWLockWaitListLock() when enqueuing for
1028 : * another lock.
1029 : */
1030 3222622 : pg_write_barrier();
1031 3222622 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1032 3222622 : PGSemaphoreUnlock(waiter->sem);
1033 : }
1034 3998432 : }
1035 :
1036 : /*
1037 : * Add ourselves to the end of the queue.
1038 : *
1039 : * NB: Mode can be LW_WAIT_UNTIL_FREE here!
1040 : */
1041 : static void
1042 3444168 : LWLockQueueSelf(LWLock *lock, LWLockMode mode)
1043 : {
1044 : /*
1045 : * If we don't have a PGPROC structure, there's no way to wait. This
1046 : * should never occur, since MyProc should only be null during shared
1047 : * memory initialization.
1048 : */
1049 3444168 : if (MyProc == NULL)
1050 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1051 :
1052 3444168 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
1053 0 : elog(PANIC, "queueing for lock while waiting on another one");
1054 :
1055 3444168 : LWLockWaitListLock(lock);
1056 :
1057 : /* setting the flag is protected by the spinlock */
1058 3444168 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_HAS_WAITERS);
1059 :
1060 3444168 : MyProc->lwWaiting = LW_WS_WAITING;
1061 3444168 : MyProc->lwWaitMode = mode;
1062 :
1063 : /* LW_WAIT_UNTIL_FREE waiters are always at the front of the queue */
1064 3444168 : if (mode == LW_WAIT_UNTIL_FREE)
1065 227638 : proclist_push_head(&lock->waiters, MyProcNumber, lwWaitLink);
1066 : else
1067 3216530 : proclist_push_tail(&lock->waiters, MyProcNumber, lwWaitLink);
1068 :
1069 : /* Can release the mutex now */
1070 3444168 : LWLockWaitListUnlock(lock);
1071 :
1072 : #ifdef LOCK_DEBUG
1073 : pg_atomic_fetch_add_u32(&lock->nwaiters, 1);
1074 : #endif
1075 3444168 : }
1076 :
1077 : /*
1078 : * Remove ourselves from the waitlist.
1079 : *
1080 : * This is used if we queued ourselves because we thought we needed to sleep
1081 : * but, after further checking, we discovered that we don't actually need to
1082 : * do so.
1083 : */
1084 : static void
1085 298692 : LWLockDequeueSelf(LWLock *lock)
1086 : {
1087 : bool on_waitlist;
1088 :
1089 : #ifdef LWLOCK_STATS
1090 : lwlock_stats *lwstats;
1091 :
1092 : lwstats = get_lwlock_stats_entry(lock);
1093 :
1094 : lwstats->dequeue_self_count++;
1095 : #endif
1096 :
1097 298692 : LWLockWaitListLock(lock);
1098 :
1099 : /*
1100 : * Remove ourselves from the waitlist, unless we've already been removed.
1101 : * The removal happens with the wait list lock held, so there's no race in
1102 : * this check.
1103 : */
1104 298692 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
1105 298692 : if (on_waitlist)
1106 222168 : proclist_delete(&lock->waiters, MyProcNumber, lwWaitLink);
1107 :
1108 298692 : if (proclist_is_empty(&lock->waiters) &&
1109 284186 : (pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS) != 0)
1110 : {
1111 283742 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_HAS_WAITERS);
1112 : }
1113 :
1114 : /* XXX: combine with fetch_and above? */
1115 298692 : LWLockWaitListUnlock(lock);
1116 :
1117 : /* clear waiting state again, nice for debugging */
1118 298692 : if (on_waitlist)
1119 222168 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
1120 : else
1121 : {
1122 76524 : int extraWaits = 0;
1123 :
1124 : /*
1125 : * Somebody else dequeued us and has or will wake us up. Deal with the
1126 : * superfluous absorption of a wakeup.
1127 : */
1128 :
1129 : /*
1130 : * Reset RELEASE_OK flag if somebody woke us before we removed
1131 : * ourselves - they'll have set it to false.
1132 : */
1133 76524 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_RELEASE_OK);
1134 :
1135 : /*
1136 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
1137 : * get reset at some inconvenient point later. Most of the time this
1138 : * will immediately return.
1139 : */
1140 : for (;;)
1141 : {
1142 76524 : PGSemaphoreLock(MyProc->sem);
1143 76524 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
1144 76524 : break;
1145 0 : extraWaits++;
1146 : }
1147 :
1148 : /*
1149 : * Fix the process wait semaphore's count for any absorbed wakeups.
1150 : */
1151 76524 : while (extraWaits-- > 0)
1152 0 : PGSemaphoreUnlock(MyProc->sem);
1153 : }
1154 :
1155 : #ifdef LOCK_DEBUG
1156 : {
1157 : /* not waiting anymore */
1158 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1159 :
1160 : Assert(nwaiters < MAX_BACKENDS);
1161 : }
1162 : #endif
1163 298692 : }
1164 :
1165 : /*
1166 : * LWLockAcquire - acquire a lightweight lock in the specified mode
1167 : *
1168 : * If the lock is not available, sleep until it is. Returns true if the lock
1169 : * was available immediately, false if we had to sleep.
1170 : *
1171 : * Side effect: cancel/die interrupts are held off until lock release.
1172 : */
1173 : bool
1174 744944246 : LWLockAcquire(LWLock *lock, LWLockMode mode)
1175 : {
1176 744944246 : PGPROC *proc = MyProc;
1177 744944246 : bool result = true;
1178 744944246 : int extraWaits = 0;
1179 : #ifdef LWLOCK_STATS
1180 : lwlock_stats *lwstats;
1181 :
1182 : lwstats = get_lwlock_stats_entry(lock);
1183 : #endif
1184 :
1185 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1186 :
1187 : PRINT_LWDEBUG("LWLockAcquire", lock, mode);
1188 :
1189 : #ifdef LWLOCK_STATS
1190 : /* Count lock acquisition attempts */
1191 : if (mode == LW_EXCLUSIVE)
1192 : lwstats->ex_acquire_count++;
1193 : else
1194 : lwstats->sh_acquire_count++;
1195 : #endif /* LWLOCK_STATS */
1196 :
1197 : /*
1198 : * We can't wait if we haven't got a PGPROC. This should only occur
1199 : * during bootstrap or shared memory initialization. Put an Assert here
1200 : * to catch unsafe coding practices.
1201 : */
1202 : Assert(!(proc == NULL && IsUnderPostmaster));
1203 :
1204 : /* Ensure we will have room to remember the lock */
1205 744944246 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1206 0 : elog(ERROR, "too many LWLocks taken");
1207 :
1208 : /*
1209 : * Lock out cancel/die interrupts until we exit the code section protected
1210 : * by the LWLock. This ensures that interrupts will not interfere with
1211 : * manipulations of data structures in shared memory.
1212 : */
1213 744944246 : HOLD_INTERRUPTS();
1214 :
1215 : /*
1216 : * Loop here to try to acquire lock after each time we are signaled by
1217 : * LWLockRelease.
1218 : *
1219 : * NOTE: it might seem better to have LWLockRelease actually grant us the
1220 : * lock, rather than retrying and possibly having to go back to sleep. But
1221 : * in practice that is no good because it means a process swap for every
1222 : * lock acquisition when two or more processes are contending for the same
1223 : * lock. Since LWLocks are normally used to protect not-very-long
1224 : * sections of computation, a process needs to be able to acquire and
1225 : * release the same lock many times during a single CPU time slice, even
1226 : * in the presence of contention. The efficiency of being able to do that
1227 : * outweighs the inefficiency of sometimes wasting a process dispatch
1228 : * cycle because the lock is not free when a released waiter finally gets
1229 : * to run. See pgsql-hackers archives for 29-Dec-01.
1230 : */
1231 : for (;;)
1232 2922838 : {
1233 : bool mustwait;
1234 :
1235 : /*
1236 : * Try to grab the lock the first time, we're not in the waitqueue
1237 : * yet/anymore.
1238 : */
1239 747867084 : mustwait = LWLockAttemptLock(lock, mode);
1240 :
1241 747867084 : if (!mustwait)
1242 : {
1243 : LOG_LWDEBUG("LWLockAcquire", lock, "immediately acquired lock");
1244 744650554 : break; /* got the lock */
1245 : }
1246 :
1247 : /*
1248 : * Ok, at this point we couldn't grab the lock on the first try. We
1249 : * cannot simply queue ourselves to the end of the list and wait to be
1250 : * woken up because by now the lock could long have been released.
1251 : * Instead add us to the queue and try to grab the lock again. If we
1252 : * succeed we need to revert the queuing and be happy, otherwise we
1253 : * recheck the lock. If we still couldn't grab it, we know that the
1254 : * other locker will see our queue entries when releasing since they
1255 : * existed before we checked for the lock.
1256 : */
1257 :
1258 : /* add to the queue */
1259 3216530 : LWLockQueueSelf(lock, mode);
1260 :
1261 : /* we're now guaranteed to be woken up if necessary */
1262 3216530 : mustwait = LWLockAttemptLock(lock, mode);
1263 :
1264 : /* ok, grabbed the lock the second time round, need to undo queueing */
1265 3216530 : if (!mustwait)
1266 : {
1267 : LOG_LWDEBUG("LWLockAcquire", lock, "acquired, undoing queue");
1268 :
1269 293692 : LWLockDequeueSelf(lock);
1270 293692 : break;
1271 : }
1272 :
1273 : /*
1274 : * Wait until awakened.
1275 : *
1276 : * It is possible that we get awakened for a reason other than being
1277 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1278 : * we've gotten the LWLock, re-increment the sema by the number of
1279 : * additional signals received.
1280 : */
1281 : LOG_LWDEBUG("LWLockAcquire", lock, "waiting");
1282 :
1283 : #ifdef LWLOCK_STATS
1284 : lwstats->block_count++;
1285 : #endif
1286 :
1287 2922838 : LWLockReportWaitStart(lock);
1288 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1289 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1290 :
1291 : for (;;)
1292 : {
1293 2922838 : PGSemaphoreLock(proc->sem);
1294 2922838 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1295 2922838 : break;
1296 0 : extraWaits++;
1297 : }
1298 :
1299 : /* Retrying, allow LWLockRelease to release waiters again. */
1300 2922838 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_RELEASE_OK);
1301 :
1302 : #ifdef LOCK_DEBUG
1303 : {
1304 : /* not waiting anymore */
1305 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1306 :
1307 : Assert(nwaiters < MAX_BACKENDS);
1308 : }
1309 : #endif
1310 :
1311 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1312 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1313 2922838 : LWLockReportWaitEnd();
1314 :
1315 : LOG_LWDEBUG("LWLockAcquire", lock, "awakened");
1316 :
1317 : /* Now loop back and try to acquire lock again. */
1318 2922838 : result = false;
1319 : }
1320 :
1321 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_ENABLED())
1322 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE(T_NAME(lock), mode);
1323 :
1324 : /* Add lock to list of locks held by this backend */
1325 744944246 : held_lwlocks[num_held_lwlocks].lock = lock;
1326 744944246 : held_lwlocks[num_held_lwlocks++].mode = mode;
1327 :
1328 : /*
1329 : * Fix the process wait semaphore's count for any absorbed wakeups.
1330 : */
1331 744944246 : while (extraWaits-- > 0)
1332 0 : PGSemaphoreUnlock(proc->sem);
1333 :
1334 744944246 : return result;
1335 : }
1336 :
1337 : /*
1338 : * LWLockConditionalAcquire - acquire a lightweight lock in the specified mode
1339 : *
1340 : * If the lock is not available, return false with no side-effects.
1341 : *
1342 : * If successful, cancel/die interrupts are held off until lock release.
1343 : */
1344 : bool
1345 4759390 : LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
1346 : {
1347 : bool mustwait;
1348 :
1349 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1350 :
1351 : PRINT_LWDEBUG("LWLockConditionalAcquire", lock, mode);
1352 :
1353 : /* Ensure we will have room to remember the lock */
1354 4759390 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1355 0 : elog(ERROR, "too many LWLocks taken");
1356 :
1357 : /*
1358 : * Lock out cancel/die interrupts until we exit the code section protected
1359 : * by the LWLock. This ensures that interrupts will not interfere with
1360 : * manipulations of data structures in shared memory.
1361 : */
1362 4759390 : HOLD_INTERRUPTS();
1363 :
1364 : /* Check for the lock */
1365 4759390 : mustwait = LWLockAttemptLock(lock, mode);
1366 :
1367 4759390 : if (mustwait)
1368 : {
1369 : /* Failed to get lock, so release interrupt holdoff */
1370 8182 : RESUME_INTERRUPTS();
1371 :
1372 : LOG_LWDEBUG("LWLockConditionalAcquire", lock, "failed");
1373 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL_ENABLED())
1374 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL(T_NAME(lock), mode);
1375 : }
1376 : else
1377 : {
1378 : /* Add lock to list of locks held by this backend */
1379 4751208 : held_lwlocks[num_held_lwlocks].lock = lock;
1380 4751208 : held_lwlocks[num_held_lwlocks++].mode = mode;
1381 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_ENABLED())
1382 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE(T_NAME(lock), mode);
1383 : }
1384 4759390 : return !mustwait;
1385 : }
1386 :
1387 : /*
1388 : * LWLockAcquireOrWait - Acquire lock, or wait until it's free
1389 : *
1390 : * The semantics of this function are a bit funky. If the lock is currently
1391 : * free, it is acquired in the given mode, and the function returns true. If
1392 : * the lock isn't immediately free, the function waits until it is released
1393 : * and returns false, but does not acquire the lock.
1394 : *
1395 : * This is currently used for WALWriteLock: when a backend flushes the WAL,
1396 : * holding WALWriteLock, it can flush the commit records of many other
1397 : * backends as a side-effect. Those other backends need to wait until the
1398 : * flush finishes, but don't need to acquire the lock anymore. They can just
1399 : * wake up, observe that their records have already been flushed, and return.
1400 : */
1401 : bool
1402 268476 : LWLockAcquireOrWait(LWLock *lock, LWLockMode mode)
1403 : {
1404 268476 : PGPROC *proc = MyProc;
1405 : bool mustwait;
1406 268476 : int extraWaits = 0;
1407 : #ifdef LWLOCK_STATS
1408 : lwlock_stats *lwstats;
1409 :
1410 : lwstats = get_lwlock_stats_entry(lock);
1411 : #endif
1412 :
1413 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1414 :
1415 : PRINT_LWDEBUG("LWLockAcquireOrWait", lock, mode);
1416 :
1417 : /* Ensure we will have room to remember the lock */
1418 268476 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1419 0 : elog(ERROR, "too many LWLocks taken");
1420 :
1421 : /*
1422 : * Lock out cancel/die interrupts until we exit the code section protected
1423 : * by the LWLock. This ensures that interrupts will not interfere with
1424 : * manipulations of data structures in shared memory.
1425 : */
1426 268476 : HOLD_INTERRUPTS();
1427 :
1428 : /*
1429 : * NB: We're using nearly the same twice-in-a-row lock acquisition
1430 : * protocol as LWLockAcquire(). Check its comments for details.
1431 : */
1432 268476 : mustwait = LWLockAttemptLock(lock, mode);
1433 :
1434 268476 : if (mustwait)
1435 : {
1436 11692 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1437 :
1438 11692 : mustwait = LWLockAttemptLock(lock, mode);
1439 :
1440 11692 : if (mustwait)
1441 : {
1442 : /*
1443 : * Wait until awakened. Like in LWLockAcquire, be prepared for
1444 : * bogus wakeups.
1445 : */
1446 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "waiting");
1447 :
1448 : #ifdef LWLOCK_STATS
1449 : lwstats->block_count++;
1450 : #endif
1451 :
1452 11370 : LWLockReportWaitStart(lock);
1453 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1454 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1455 :
1456 : for (;;)
1457 : {
1458 11370 : PGSemaphoreLock(proc->sem);
1459 11370 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1460 11370 : break;
1461 0 : extraWaits++;
1462 : }
1463 :
1464 : #ifdef LOCK_DEBUG
1465 : {
1466 : /* not waiting anymore */
1467 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1468 :
1469 : Assert(nwaiters < MAX_BACKENDS);
1470 : }
1471 : #endif
1472 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1473 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1474 11370 : LWLockReportWaitEnd();
1475 :
1476 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "awakened");
1477 : }
1478 : else
1479 : {
1480 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "acquired, undoing queue");
1481 :
1482 : /*
1483 : * Got lock in the second attempt, undo queueing. We need to treat
1484 : * this as having successfully acquired the lock, otherwise we'd
1485 : * not necessarily wake up people we've prevented from acquiring
1486 : * the lock.
1487 : */
1488 322 : LWLockDequeueSelf(lock);
1489 : }
1490 : }
1491 :
1492 : /*
1493 : * Fix the process wait semaphore's count for any absorbed wakeups.
1494 : */
1495 268476 : while (extraWaits-- > 0)
1496 0 : PGSemaphoreUnlock(proc->sem);
1497 :
1498 268476 : if (mustwait)
1499 : {
1500 : /* Failed to get lock, so release interrupt holdoff */
1501 11370 : RESUME_INTERRUPTS();
1502 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "failed");
1503 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL_ENABLED())
1504 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL(T_NAME(lock), mode);
1505 : }
1506 : else
1507 : {
1508 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "succeeded");
1509 : /* Add lock to list of locks held by this backend */
1510 257106 : held_lwlocks[num_held_lwlocks].lock = lock;
1511 257106 : held_lwlocks[num_held_lwlocks++].mode = mode;
1512 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_ENABLED())
1513 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT(T_NAME(lock), mode);
1514 : }
1515 :
1516 268476 : return !mustwait;
1517 : }
1518 :
1519 : /*
1520 : * Does the lwlock in its current state need to wait for the variable value to
1521 : * change?
1522 : *
1523 : * If we don't need to wait, and it's because the value of the variable has
1524 : * changed, store the current value in newval.
1525 : *
1526 : * *result is set to true if the lock was free, and false otherwise.
1527 : */
1528 : static bool
1529 7906228 : LWLockConflictsWithVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1530 : uint64 *newval, bool *result)
1531 : {
1532 : bool mustwait;
1533 : uint64 value;
1534 :
1535 : /*
1536 : * Test first to see if it the slot is free right now.
1537 : *
1538 : * XXX: the unique caller of this routine, WaitXLogInsertionsToFinish()
1539 : * via LWLockWaitForVar(), uses an implied barrier with a spinlock before
1540 : * this, so we don't need a memory barrier here as far as the current
1541 : * usage is concerned. But that might not be safe in general.
1542 : */
1543 7906228 : mustwait = (pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE) != 0;
1544 :
1545 7906228 : if (!mustwait)
1546 : {
1547 5201146 : *result = true;
1548 5201146 : return false;
1549 : }
1550 :
1551 2705082 : *result = false;
1552 :
1553 : /*
1554 : * Reading this value atomically is safe even on platforms where uint64
1555 : * cannot be read without observing a torn value.
1556 : */
1557 2705082 : value = pg_atomic_read_u64(valptr);
1558 :
1559 2705082 : if (value != oldval)
1560 : {
1561 2277868 : mustwait = false;
1562 2277868 : *newval = value;
1563 : }
1564 : else
1565 : {
1566 427214 : mustwait = true;
1567 : }
1568 :
1569 2705082 : return mustwait;
1570 : }
1571 :
1572 : /*
1573 : * LWLockWaitForVar - Wait until lock is free, or a variable is updated.
1574 : *
1575 : * If the lock is held and *valptr equals oldval, waits until the lock is
1576 : * either freed, or the lock holder updates *valptr by calling
1577 : * LWLockUpdateVar. If the lock is free on exit (immediately or after
1578 : * waiting), returns true. If the lock is still held, but *valptr no longer
1579 : * matches oldval, returns false and sets *newval to the current value in
1580 : * *valptr.
1581 : *
1582 : * Note: this function ignores shared lock holders; if the lock is held
1583 : * in shared mode, returns 'true'.
1584 : *
1585 : * Be aware that LWLockConflictsWithVar() does not include a memory barrier,
1586 : * hence the caller of this function may want to rely on an explicit barrier or
1587 : * an implied barrier via spinlock or LWLock to avoid memory ordering issues.
1588 : */
1589 : bool
1590 7479014 : LWLockWaitForVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1591 : uint64 *newval)
1592 : {
1593 7479014 : PGPROC *proc = MyProc;
1594 7479014 : int extraWaits = 0;
1595 7479014 : bool result = false;
1596 : #ifdef LWLOCK_STATS
1597 : lwlock_stats *lwstats;
1598 :
1599 : lwstats = get_lwlock_stats_entry(lock);
1600 : #endif
1601 :
1602 : PRINT_LWDEBUG("LWLockWaitForVar", lock, LW_WAIT_UNTIL_FREE);
1603 :
1604 : /*
1605 : * Lock out cancel/die interrupts while we sleep on the lock. There is no
1606 : * cleanup mechanism to remove us from the wait queue if we got
1607 : * interrupted.
1608 : */
1609 7479014 : HOLD_INTERRUPTS();
1610 :
1611 : /*
1612 : * Loop here to check the lock's status after each time we are signaled.
1613 : */
1614 : for (;;)
1615 211268 : {
1616 : bool mustwait;
1617 :
1618 7690282 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1619 : &result);
1620 :
1621 7690282 : if (!mustwait)
1622 7474336 : break; /* the lock was free or value didn't match */
1623 :
1624 : /*
1625 : * Add myself to wait queue. Note that this is racy, somebody else
1626 : * could wakeup before we're finished queuing. NB: We're using nearly
1627 : * the same twice-in-a-row lock acquisition protocol as
1628 : * LWLockAcquire(). Check its comments for details. The only
1629 : * difference is that we also have to check the variable's values when
1630 : * checking the state of the lock.
1631 : */
1632 215946 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1633 :
1634 : /*
1635 : * Set RELEASE_OK flag, to make sure we get woken up as soon as the
1636 : * lock is released.
1637 : */
1638 215946 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_RELEASE_OK);
1639 :
1640 : /*
1641 : * We're now guaranteed to be woken up if necessary. Recheck the lock
1642 : * and variables state.
1643 : */
1644 215946 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1645 : &result);
1646 :
1647 : /* Ok, no conflict after we queued ourselves. Undo queueing. */
1648 215946 : if (!mustwait)
1649 : {
1650 : LOG_LWDEBUG("LWLockWaitForVar", lock, "free, undoing queue");
1651 :
1652 4678 : LWLockDequeueSelf(lock);
1653 4678 : break;
1654 : }
1655 :
1656 : /*
1657 : * Wait until awakened.
1658 : *
1659 : * It is possible that we get awakened for a reason other than being
1660 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1661 : * we've gotten the LWLock, re-increment the sema by the number of
1662 : * additional signals received.
1663 : */
1664 : LOG_LWDEBUG("LWLockWaitForVar", lock, "waiting");
1665 :
1666 : #ifdef LWLOCK_STATS
1667 : lwstats->block_count++;
1668 : #endif
1669 :
1670 211268 : LWLockReportWaitStart(lock);
1671 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1672 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), LW_EXCLUSIVE);
1673 :
1674 : for (;;)
1675 : {
1676 211268 : PGSemaphoreLock(proc->sem);
1677 211268 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1678 211268 : break;
1679 0 : extraWaits++;
1680 : }
1681 :
1682 : #ifdef LOCK_DEBUG
1683 : {
1684 : /* not waiting anymore */
1685 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1686 :
1687 : Assert(nwaiters < MAX_BACKENDS);
1688 : }
1689 : #endif
1690 :
1691 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1692 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), LW_EXCLUSIVE);
1693 211268 : LWLockReportWaitEnd();
1694 :
1695 : LOG_LWDEBUG("LWLockWaitForVar", lock, "awakened");
1696 :
1697 : /* Now loop back and check the status of the lock again. */
1698 : }
1699 :
1700 : /*
1701 : * Fix the process wait semaphore's count for any absorbed wakeups.
1702 : */
1703 7479014 : while (extraWaits-- > 0)
1704 0 : PGSemaphoreUnlock(proc->sem);
1705 :
1706 : /*
1707 : * Now okay to allow cancel/die interrupts.
1708 : */
1709 7479014 : RESUME_INTERRUPTS();
1710 :
1711 7479014 : return result;
1712 : }
1713 :
1714 :
1715 : /*
1716 : * LWLockUpdateVar - Update a variable and wake up waiters atomically
1717 : *
1718 : * Sets *valptr to 'val', and wakes up all processes waiting for us with
1719 : * LWLockWaitForVar(). It first sets the value atomically and then wakes up
1720 : * waiting processes so that any process calling LWLockWaitForVar() on the same
1721 : * lock is guaranteed to see the new value, and act accordingly.
1722 : *
1723 : * The caller must be holding the lock in exclusive mode.
1724 : */
1725 : void
1726 5226504 : LWLockUpdateVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1727 : {
1728 : proclist_head wakeup;
1729 : proclist_mutable_iter iter;
1730 :
1731 : PRINT_LWDEBUG("LWLockUpdateVar", lock, LW_EXCLUSIVE);
1732 :
1733 : /*
1734 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1735 : * that the variable is updated before waking up waiters.
1736 : */
1737 5226504 : pg_atomic_exchange_u64(valptr, val);
1738 :
1739 5226504 : proclist_init(&wakeup);
1740 :
1741 5226504 : LWLockWaitListLock(lock);
1742 :
1743 : Assert(pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE);
1744 :
1745 : /*
1746 : * See if there are any LW_WAIT_UNTIL_FREE waiters that need to be woken
1747 : * up. They are always in the front of the queue.
1748 : */
1749 5229192 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
1750 : {
1751 135738 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1752 :
1753 135738 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
1754 133050 : break;
1755 :
1756 2688 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
1757 2688 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
1758 :
1759 : /* see LWLockWakeup() */
1760 : Assert(waiter->lwWaiting == LW_WS_WAITING);
1761 2688 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
1762 : }
1763 :
1764 : /* We are done updating shared state of the lock itself. */
1765 5226504 : LWLockWaitListUnlock(lock);
1766 :
1767 : /*
1768 : * Awaken any waiters I removed from the queue.
1769 : */
1770 5229192 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1771 : {
1772 2688 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1773 :
1774 2688 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1775 : /* check comment in LWLockWakeup() about this barrier */
1776 2688 : pg_write_barrier();
1777 2688 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1778 2688 : PGSemaphoreUnlock(waiter->sem);
1779 : }
1780 5226504 : }
1781 :
1782 :
1783 : /*
1784 : * Stop treating lock as held by current backend.
1785 : *
1786 : * This is the code that can be shared between actually releasing a lock
1787 : * (LWLockRelease()) and just not tracking ownership of the lock anymore
1788 : * without releasing the lock (LWLockDisown()).
1789 : *
1790 : * Returns the mode in which the lock was held by the current backend.
1791 : *
1792 : * NB: This does not call RESUME_INTERRUPTS(), but leaves that responsibility
1793 : * of the caller.
1794 : *
1795 : * NB: This will leave lock->owner pointing to the current backend (if
1796 : * LOCK_DEBUG is set). This is somewhat intentional, as it makes it easier to
1797 : * debug cases of missing wakeups during lock release.
1798 : */
1799 : static inline LWLockMode
1800 749952560 : LWLockDisownInternal(LWLock *lock)
1801 : {
1802 : LWLockMode mode;
1803 : int i;
1804 :
1805 : /*
1806 : * Remove lock from list of locks held. Usually, but not always, it will
1807 : * be the latest-acquired lock; so search array backwards.
1808 : */
1809 832962250 : for (i = num_held_lwlocks; --i >= 0;)
1810 832962250 : if (lock == held_lwlocks[i].lock)
1811 749952560 : break;
1812 :
1813 749952560 : if (i < 0)
1814 0 : elog(ERROR, "lock %s is not held", T_NAME(lock));
1815 :
1816 749952560 : mode = held_lwlocks[i].mode;
1817 :
1818 749952560 : num_held_lwlocks--;
1819 832962250 : for (; i < num_held_lwlocks; i++)
1820 83009690 : held_lwlocks[i] = held_lwlocks[i + 1];
1821 :
1822 749952560 : return mode;
1823 : }
1824 :
1825 : /*
1826 : * Helper function to release lock, shared between LWLockRelease() and
1827 : * LWLockReleaseDisowned().
1828 : */
1829 : static void
1830 749952560 : LWLockReleaseInternal(LWLock *lock, LWLockMode mode)
1831 : {
1832 : uint32 oldstate;
1833 : bool check_waiters;
1834 :
1835 : /*
1836 : * Release my hold on lock, after that it can immediately be acquired by
1837 : * others, even if we still have to wakeup other waiters.
1838 : */
1839 749952560 : if (mode == LW_EXCLUSIVE)
1840 462046608 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_EXCLUSIVE);
1841 : else
1842 287905952 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_SHARED);
1843 :
1844 : /* nobody else can have that kind of lock */
1845 : Assert(!(oldstate & LW_VAL_EXCLUSIVE));
1846 :
1847 : if (TRACE_POSTGRESQL_LWLOCK_RELEASE_ENABLED())
1848 : TRACE_POSTGRESQL_LWLOCK_RELEASE(T_NAME(lock));
1849 :
1850 : /*
1851 : * We're still waiting for backends to get scheduled, don't wake them up
1852 : * again.
1853 : */
1854 749952560 : if ((oldstate & (LW_FLAG_HAS_WAITERS | LW_FLAG_RELEASE_OK)) ==
1855 4021892 : (LW_FLAG_HAS_WAITERS | LW_FLAG_RELEASE_OK) &&
1856 4021892 : (oldstate & LW_LOCK_MASK) == 0)
1857 3998432 : check_waiters = true;
1858 : else
1859 745954128 : check_waiters = false;
1860 :
1861 : /*
1862 : * As waking up waiters requires the spinlock to be acquired, only do so
1863 : * if necessary.
1864 : */
1865 749952560 : if (check_waiters)
1866 : {
1867 : /* XXX: remove before commit? */
1868 : LOG_LWDEBUG("LWLockRelease", lock, "releasing waiters");
1869 3998432 : LWLockWakeup(lock);
1870 : }
1871 749952560 : }
1872 :
1873 :
1874 : /*
1875 : * Stop treating lock as held by current backend.
1876 : *
1877 : * After calling this function it's the callers responsibility to ensure that
1878 : * the lock gets released (via LWLockReleaseDisowned()), even in case of an
1879 : * error. This only is desirable if the lock is going to be released in a
1880 : * different process than the process that acquired it.
1881 : */
1882 : void
1883 0 : LWLockDisown(LWLock *lock)
1884 : {
1885 0 : LWLockDisownInternal(lock);
1886 :
1887 0 : RESUME_INTERRUPTS();
1888 0 : }
1889 :
1890 : /*
1891 : * LWLockRelease - release a previously acquired lock
1892 : */
1893 : void
1894 749952560 : LWLockRelease(LWLock *lock)
1895 : {
1896 : LWLockMode mode;
1897 :
1898 749952560 : mode = LWLockDisownInternal(lock);
1899 :
1900 : PRINT_LWDEBUG("LWLockRelease", lock, mode);
1901 :
1902 749952560 : LWLockReleaseInternal(lock, mode);
1903 :
1904 : /*
1905 : * Now okay to allow cancel/die interrupts.
1906 : */
1907 749952560 : RESUME_INTERRUPTS();
1908 749952560 : }
1909 :
1910 : /*
1911 : * Release lock previously disowned with LWLockDisown().
1912 : */
1913 : void
1914 0 : LWLockReleaseDisowned(LWLock *lock, LWLockMode mode)
1915 : {
1916 0 : LWLockReleaseInternal(lock, mode);
1917 0 : }
1918 :
1919 : /*
1920 : * LWLockReleaseClearVar - release a previously acquired lock, reset variable
1921 : */
1922 : void
1923 29295242 : LWLockReleaseClearVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1924 : {
1925 : /*
1926 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1927 : * that the variable is updated before releasing the lock.
1928 : */
1929 29295242 : pg_atomic_exchange_u64(valptr, val);
1930 :
1931 29295242 : LWLockRelease(lock);
1932 29295242 : }
1933 :
1934 :
1935 : /*
1936 : * LWLockReleaseAll - release all currently-held locks
1937 : *
1938 : * Used to clean up after ereport(ERROR). An important difference between this
1939 : * function and retail LWLockRelease calls is that InterruptHoldoffCount is
1940 : * unchanged by this operation. This is necessary since InterruptHoldoffCount
1941 : * has been set to an appropriate level earlier in error recovery. We could
1942 : * decrement it below zero if we allow it to drop for each released lock!
1943 : */
1944 : void
1945 112756 : LWLockReleaseAll(void)
1946 : {
1947 113154 : while (num_held_lwlocks > 0)
1948 : {
1949 398 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
1950 :
1951 398 : LWLockRelease(held_lwlocks[num_held_lwlocks - 1].lock);
1952 : }
1953 112756 : }
1954 :
1955 :
1956 : /*
1957 : * ForEachLWLockHeldByMe - run a callback for each held lock
1958 : *
1959 : * This is meant as debug support only.
1960 : */
1961 : void
1962 0 : ForEachLWLockHeldByMe(void (*callback) (LWLock *, LWLockMode, void *),
1963 : void *context)
1964 : {
1965 : int i;
1966 :
1967 0 : for (i = 0; i < num_held_lwlocks; i++)
1968 0 : callback(held_lwlocks[i].lock, held_lwlocks[i].mode, context);
1969 0 : }
1970 :
1971 : /*
1972 : * LWLockHeldByMe - test whether my process holds a lock in any mode
1973 : *
1974 : * This is meant as debug support only.
1975 : */
1976 : bool
1977 0 : LWLockHeldByMe(LWLock *lock)
1978 : {
1979 : int i;
1980 :
1981 0 : for (i = 0; i < num_held_lwlocks; i++)
1982 : {
1983 0 : if (held_lwlocks[i].lock == lock)
1984 0 : return true;
1985 : }
1986 0 : return false;
1987 : }
1988 :
1989 : /*
1990 : * LWLockAnyHeldByMe - test whether my process holds any of an array of locks
1991 : *
1992 : * This is meant as debug support only.
1993 : */
1994 : bool
1995 0 : LWLockAnyHeldByMe(LWLock *lock, int nlocks, size_t stride)
1996 : {
1997 : char *held_lock_addr;
1998 : char *begin;
1999 : char *end;
2000 : int i;
2001 :
2002 0 : begin = (char *) lock;
2003 0 : end = begin + nlocks * stride;
2004 0 : for (i = 0; i < num_held_lwlocks; i++)
2005 : {
2006 0 : held_lock_addr = (char *) held_lwlocks[i].lock;
2007 0 : if (held_lock_addr >= begin &&
2008 0 : held_lock_addr < end &&
2009 0 : (held_lock_addr - begin) % stride == 0)
2010 0 : return true;
2011 : }
2012 0 : return false;
2013 : }
2014 :
2015 : /*
2016 : * LWLockHeldByMeInMode - test whether my process holds a lock in given mode
2017 : *
2018 : * This is meant as debug support only.
2019 : */
2020 : bool
2021 0 : LWLockHeldByMeInMode(LWLock *lock, LWLockMode mode)
2022 : {
2023 : int i;
2024 :
2025 0 : for (i = 0; i < num_held_lwlocks; i++)
2026 : {
2027 0 : if (held_lwlocks[i].lock == lock && held_lwlocks[i].mode == mode)
2028 0 : return true;
2029 : }
2030 0 : return false;
2031 : }
|
