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 : static NamedLWLockTrancheRequest *NamedLWLockTrancheRequestArray = NULL;
188 :
189 : /*
190 : * NamedLWLockTrancheRequests is the valid length of the request array. This
191 : * variable is non-static so that postmaster.c can copy them to child processes
192 : * in EXEC_BACKEND builds.
193 : */
194 : int NamedLWLockTrancheRequests = 0;
195 :
196 : /* shared memory counter of registered tranches */
197 : int *LWLockCounter = NULL;
198 :
199 : /* backend-local counter of registered tranches */
200 : static int LocalLWLockCounter;
201 :
202 : #define MAX_NAMED_TRANCHES 256
203 :
204 : static void InitializeLWLocks(void);
205 : static inline void LWLockReportWaitStart(LWLock *lock);
206 : static inline void LWLockReportWaitEnd(void);
207 : static const char *GetLWTrancheName(uint16 trancheId);
208 :
209 : #define T_NAME(lock) \
210 : GetLWTrancheName((lock)->tranche)
211 :
212 : #ifdef LWLOCK_STATS
213 : typedef struct lwlock_stats_key
214 : {
215 : int tranche;
216 : void *instance;
217 : } lwlock_stats_key;
218 :
219 : typedef struct lwlock_stats
220 : {
221 : lwlock_stats_key key;
222 : int sh_acquire_count;
223 : int ex_acquire_count;
224 : int block_count;
225 : int dequeue_self_count;
226 : int spin_delay_count;
227 : } lwlock_stats;
228 :
229 : static HTAB *lwlock_stats_htab;
230 : static lwlock_stats lwlock_stats_dummy;
231 : #endif
232 :
233 : #ifdef LOCK_DEBUG
234 : bool Trace_lwlocks = false;
235 :
236 : inline static void
237 : PRINT_LWDEBUG(const char *where, LWLock *lock, LWLockMode mode)
238 : {
239 : /* hide statement & context here, otherwise the log is just too verbose */
240 : if (Trace_lwlocks)
241 : {
242 : uint32 state = pg_atomic_read_u32(&lock->state);
243 :
244 : ereport(LOG,
245 : (errhidestmt(true),
246 : errhidecontext(true),
247 : errmsg_internal("%d: %s(%s %p): excl %u shared %u haswaiters %u waiters %u rOK %d",
248 : MyProcPid,
249 : where, T_NAME(lock), lock,
250 : (state & LW_VAL_EXCLUSIVE) != 0,
251 : state & LW_SHARED_MASK,
252 : (state & LW_FLAG_HAS_WAITERS) != 0,
253 : pg_atomic_read_u32(&lock->nwaiters),
254 : (state & LW_FLAG_RELEASE_OK) != 0)));
255 : }
256 : }
257 :
258 : inline static void
259 : LOG_LWDEBUG(const char *where, LWLock *lock, const char *msg)
260 : {
261 : /* hide statement & context here, otherwise the log is just too verbose */
262 : if (Trace_lwlocks)
263 : {
264 : ereport(LOG,
265 : (errhidestmt(true),
266 : errhidecontext(true),
267 : errmsg_internal("%s(%s %p): %s", where,
268 : T_NAME(lock), lock, msg)));
269 : }
270 : }
271 :
272 : #else /* not LOCK_DEBUG */
273 : #define PRINT_LWDEBUG(a,b,c) ((void)0)
274 : #define LOG_LWDEBUG(a,b,c) ((void)0)
275 : #endif /* LOCK_DEBUG */
276 :
277 : #ifdef LWLOCK_STATS
278 :
279 : static void init_lwlock_stats(void);
280 : static void print_lwlock_stats(int code, Datum arg);
281 : static lwlock_stats * get_lwlock_stats_entry(LWLock *lock);
282 :
283 : static void
284 : init_lwlock_stats(void)
285 : {
286 : HASHCTL ctl;
287 : static MemoryContext lwlock_stats_cxt = NULL;
288 : static bool exit_registered = false;
289 :
290 : if (lwlock_stats_cxt != NULL)
291 : MemoryContextDelete(lwlock_stats_cxt);
292 :
293 : /*
294 : * The LWLock stats will be updated within a critical section, which
295 : * requires allocating new hash entries. Allocations within a critical
296 : * section are normally not allowed because running out of memory would
297 : * lead to a PANIC, but LWLOCK_STATS is debugging code that's not normally
298 : * turned on in production, so that's an acceptable risk. The hash entries
299 : * are small, so the risk of running out of memory is minimal in practice.
300 : */
301 : lwlock_stats_cxt = AllocSetContextCreate(TopMemoryContext,
302 : "LWLock stats",
303 : ALLOCSET_DEFAULT_SIZES);
304 : MemoryContextAllowInCriticalSection(lwlock_stats_cxt, true);
305 :
306 : ctl.keysize = sizeof(lwlock_stats_key);
307 : ctl.entrysize = sizeof(lwlock_stats);
308 : ctl.hcxt = lwlock_stats_cxt;
309 : lwlock_stats_htab = hash_create("lwlock stats", 16384, &ctl,
310 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
311 : if (!exit_registered)
312 : {
313 : on_shmem_exit(print_lwlock_stats, 0);
314 : exit_registered = true;
315 : }
316 : }
317 :
318 : static void
319 : print_lwlock_stats(int code, Datum arg)
320 : {
321 : HASH_SEQ_STATUS scan;
322 : lwlock_stats *lwstats;
323 :
324 : hash_seq_init(&scan, lwlock_stats_htab);
325 :
326 : /* Grab an LWLock to keep different backends from mixing reports */
327 : LWLockAcquire(&MainLWLockArray[0].lock, LW_EXCLUSIVE);
328 :
329 : while ((lwstats = (lwlock_stats *) hash_seq_search(&scan)) != NULL)
330 : {
331 : fprintf(stderr,
332 : "PID %d lwlock %s %p: shacq %u exacq %u blk %u spindelay %u dequeue self %u\n",
333 : MyProcPid, GetLWTrancheName(lwstats->key.tranche),
334 : lwstats->key.instance, lwstats->sh_acquire_count,
335 : lwstats->ex_acquire_count, lwstats->block_count,
336 : lwstats->spin_delay_count, lwstats->dequeue_self_count);
337 : }
338 :
339 : LWLockRelease(&MainLWLockArray[0].lock);
340 : }
341 :
342 : static lwlock_stats *
343 : get_lwlock_stats_entry(LWLock *lock)
344 : {
345 : lwlock_stats_key key;
346 : lwlock_stats *lwstats;
347 : bool found;
348 :
349 : /*
350 : * During shared memory initialization, the hash table doesn't exist yet.
351 : * Stats of that phase aren't very interesting, so just collect operations
352 : * on all locks in a single dummy entry.
353 : */
354 : if (lwlock_stats_htab == NULL)
355 : return &lwlock_stats_dummy;
356 :
357 : /* Fetch or create the entry. */
358 : MemSet(&key, 0, sizeof(key));
359 : key.tranche = lock->tranche;
360 : key.instance = lock;
361 : lwstats = hash_search(lwlock_stats_htab, &key, HASH_ENTER, &found);
362 : if (!found)
363 : {
364 : lwstats->sh_acquire_count = 0;
365 : lwstats->ex_acquire_count = 0;
366 : lwstats->block_count = 0;
367 : lwstats->dequeue_self_count = 0;
368 : lwstats->spin_delay_count = 0;
369 : }
370 : return lwstats;
371 : }
372 : #endif /* LWLOCK_STATS */
373 :
374 :
375 : /*
376 : * Compute number of LWLocks required by named tranches. These will be
377 : * allocated in the main array.
378 : */
379 : static int
380 6222 : NumLWLocksForNamedTranches(void)
381 : {
382 6222 : int numLocks = 0;
383 : int i;
384 :
385 6264 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
386 42 : numLocks += NamedLWLockTrancheRequestArray[i].num_lwlocks;
387 :
388 6222 : return numLocks;
389 : }
390 :
391 : /*
392 : * Compute shmem space needed for LWLocks and named tranches.
393 : */
394 : Size
395 6222 : LWLockShmemSize(void)
396 : {
397 : Size size;
398 6222 : int numLocks = NUM_FIXED_LWLOCKS;
399 :
400 : /* Calculate total number of locks needed in the main array. */
401 6222 : numLocks += NumLWLocksForNamedTranches();
402 :
403 : /* Space for dynamic allocation counter. */
404 6222 : size = MAXALIGN(sizeof(int));
405 :
406 : /* Space for named tranches. */
407 6222 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, sizeof(char *)));
408 6222 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, NAMEDATALEN));
409 :
410 : /* Space for the LWLock array, plus room for cache line alignment. */
411 6222 : size = add_size(size, LWLOCK_PADDED_SIZE);
412 6222 : size = add_size(size, mul_size(numLocks, sizeof(LWLockPadded)));
413 :
414 6222 : return size;
415 : }
416 :
417 : /*
418 : * Allocate shmem space for the main LWLock array and all tranches and
419 : * initialize it.
420 : */
421 : void
422 2174 : CreateLWLocks(void)
423 : {
424 2174 : if (!IsUnderPostmaster)
425 : {
426 2174 : Size spaceLocks = LWLockShmemSize();
427 : char *ptr;
428 :
429 : /* Allocate space */
430 2174 : ptr = (char *) ShmemAlloc(spaceLocks);
431 :
432 : /* Initialize the dynamic-allocation counter for tranches */
433 2174 : LWLockCounter = (int *) ptr;
434 2174 : *LWLockCounter = LWTRANCHE_FIRST_USER_DEFINED;
435 2174 : ptr += MAXALIGN(sizeof(int));
436 :
437 : /* Initialize tranche names */
438 2174 : LWLockTrancheNames = (char **) ptr;
439 2174 : ptr += MAX_NAMED_TRANCHES * sizeof(char *);
440 558718 : for (int i = 0; i < MAX_NAMED_TRANCHES; i++)
441 : {
442 556544 : LWLockTrancheNames[i] = ptr;
443 556544 : ptr += NAMEDATALEN;
444 : }
445 :
446 : /* Ensure desired alignment of LWLock array */
447 2174 : ptr += LWLOCK_PADDED_SIZE - ((uintptr_t) ptr) % LWLOCK_PADDED_SIZE;
448 2174 : MainLWLockArray = (LWLockPadded *) ptr;
449 :
450 : /* Initialize all LWLocks */
451 2174 : InitializeLWLocks();
452 : }
453 2174 : }
454 :
455 : /*
456 : * Initialize LWLocks that are fixed and those belonging to named tranches.
457 : */
458 : static void
459 2174 : InitializeLWLocks(void)
460 : {
461 : int id;
462 : int i;
463 : int j;
464 : LWLockPadded *lock;
465 :
466 : /* Initialize all individual LWLocks in main array */
467 119570 : for (id = 0, lock = MainLWLockArray; id < NUM_INDIVIDUAL_LWLOCKS; id++, lock++)
468 117396 : LWLockInitialize(&lock->lock, id);
469 :
470 : /* Initialize buffer mapping LWLocks in main array */
471 2174 : lock = MainLWLockArray + BUFFER_MAPPING_LWLOCK_OFFSET;
472 280446 : for (id = 0; id < NUM_BUFFER_PARTITIONS; id++, lock++)
473 278272 : LWLockInitialize(&lock->lock, LWTRANCHE_BUFFER_MAPPING);
474 :
475 : /* Initialize lmgrs' LWLocks in main array */
476 2174 : lock = MainLWLockArray + LOCK_MANAGER_LWLOCK_OFFSET;
477 36958 : for (id = 0; id < NUM_LOCK_PARTITIONS; id++, lock++)
478 34784 : LWLockInitialize(&lock->lock, LWTRANCHE_LOCK_MANAGER);
479 :
480 : /* Initialize predicate lmgrs' LWLocks in main array */
481 2174 : lock = MainLWLockArray + PREDICATELOCK_MANAGER_LWLOCK_OFFSET;
482 36958 : for (id = 0; id < NUM_PREDICATELOCK_PARTITIONS; id++, lock++)
483 34784 : LWLockInitialize(&lock->lock, LWTRANCHE_PREDICATE_LOCK_MANAGER);
484 :
485 : /*
486 : * Copy the info about any named tranches into shared memory (so that
487 : * other processes can see it), and initialize the requested LWLocks.
488 : */
489 2174 : if (NamedLWLockTrancheRequests > 0)
490 : {
491 14 : lock = &MainLWLockArray[NUM_FIXED_LWLOCKS];
492 :
493 28 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
494 : {
495 : NamedLWLockTrancheRequest *request;
496 : int tranche;
497 :
498 14 : request = &NamedLWLockTrancheRequestArray[i];
499 14 : tranche = LWLockNewTrancheId(request->tranche_name);
500 :
501 28 : for (j = 0; j < request->num_lwlocks; j++, lock++)
502 14 : LWLockInitialize(&lock->lock, tranche);
503 : }
504 : }
505 2174 : }
506 :
507 : /*
508 : * InitLWLockAccess - initialize backend-local state needed to hold LWLocks
509 : */
510 : void
511 44518 : InitLWLockAccess(void)
512 : {
513 : #ifdef LWLOCK_STATS
514 : init_lwlock_stats();
515 : #endif
516 44518 : }
517 :
518 : /*
519 : * GetNamedLWLockTranche - returns the base address of LWLock from the
520 : * specified tranche.
521 : *
522 : * Caller needs to retrieve the requested number of LWLocks starting from
523 : * the base lock address returned by this API. This can be used for
524 : * tranches that are requested by using RequestNamedLWLockTranche() API.
525 : */
526 : LWLockPadded *
527 14 : GetNamedLWLockTranche(const char *tranche_name)
528 : {
529 : int lock_pos;
530 : int i;
531 :
532 : /*
533 : * Obtain the position of base address of LWLock belonging to requested
534 : * tranche_name in MainLWLockArray. LWLocks for named tranches are placed
535 : * in MainLWLockArray after fixed locks.
536 : */
537 14 : lock_pos = NUM_FIXED_LWLOCKS;
538 14 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
539 : {
540 14 : if (strcmp(NamedLWLockTrancheRequestArray[i].tranche_name,
541 : tranche_name) == 0)
542 14 : return &MainLWLockArray[lock_pos];
543 :
544 0 : lock_pos += NamedLWLockTrancheRequestArray[i].num_lwlocks;
545 : }
546 :
547 0 : elog(ERROR, "requested tranche is not registered");
548 :
549 : /* just to keep compiler quiet */
550 : return NULL;
551 : }
552 :
553 : /*
554 : * Allocate a new tranche ID with the provided name.
555 : */
556 : int
557 38 : LWLockNewTrancheId(const char *name)
558 : {
559 : int result;
560 :
561 38 : if (!name)
562 0 : ereport(ERROR,
563 : (errcode(ERRCODE_INVALID_NAME),
564 : errmsg("tranche name cannot be NULL")));
565 :
566 38 : if (strlen(name) >= NAMEDATALEN)
567 0 : ereport(ERROR,
568 : (errcode(ERRCODE_NAME_TOO_LONG),
569 : errmsg("tranche name too long"),
570 : errdetail("LWLock tranche names must be no longer than %d bytes.",
571 : NAMEDATALEN - 1)));
572 :
573 : /*
574 : * We use the ShmemLock spinlock to protect LWLockCounter and
575 : * LWLockTrancheNames.
576 : */
577 38 : SpinLockAcquire(ShmemLock);
578 :
579 38 : if (*LWLockCounter - LWTRANCHE_FIRST_USER_DEFINED >= MAX_NAMED_TRANCHES)
580 : {
581 0 : SpinLockRelease(ShmemLock);
582 0 : ereport(ERROR,
583 : (errmsg("maximum number of tranches already registered"),
584 : errdetail("No more than %d tranches may be registered.",
585 : MAX_NAMED_TRANCHES)));
586 : }
587 :
588 38 : result = (*LWLockCounter)++;
589 38 : LocalLWLockCounter = *LWLockCounter;
590 38 : strlcpy(LWLockTrancheNames[result - LWTRANCHE_FIRST_USER_DEFINED], name, NAMEDATALEN);
591 :
592 38 : SpinLockRelease(ShmemLock);
593 :
594 38 : return result;
595 : }
596 :
597 : /*
598 : * RequestNamedLWLockTranche
599 : * Request that extra LWLocks be allocated during postmaster
600 : * startup.
601 : *
602 : * This may only be called via the shmem_request_hook of a library that is
603 : * loaded into the postmaster via shared_preload_libraries. Calls from
604 : * elsewhere will fail.
605 : *
606 : * The tranche name will be user-visible as a wait event name, so try to
607 : * use a name that fits the style for those.
608 : */
609 : void
610 14 : RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
611 : {
612 : NamedLWLockTrancheRequest *request;
613 : static int NamedLWLockTrancheRequestsAllocated;
614 :
615 14 : if (!process_shmem_requests_in_progress)
616 0 : elog(FATAL, "cannot request additional LWLocks outside shmem_request_hook");
617 :
618 14 : if (!tranche_name)
619 0 : ereport(ERROR,
620 : (errcode(ERRCODE_INVALID_NAME),
621 : errmsg("tranche name cannot be NULL")));
622 :
623 14 : if (strlen(tranche_name) >= NAMEDATALEN)
624 0 : ereport(ERROR,
625 : (errcode(ERRCODE_NAME_TOO_LONG),
626 : errmsg("tranche name too long"),
627 : errdetail("LWLock tranche names must be no longer than %d bytes.",
628 : NAMEDATALEN - 1)));
629 :
630 14 : if (NamedLWLockTrancheRequestArray == NULL)
631 : {
632 14 : NamedLWLockTrancheRequestsAllocated = 16;
633 14 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
634 14 : MemoryContextAlloc(TopMemoryContext,
635 : NamedLWLockTrancheRequestsAllocated
636 : * sizeof(NamedLWLockTrancheRequest));
637 : }
638 :
639 14 : if (NamedLWLockTrancheRequests >= NamedLWLockTrancheRequestsAllocated)
640 : {
641 0 : int i = pg_nextpower2_32(NamedLWLockTrancheRequests + 1);
642 :
643 0 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
644 0 : repalloc(NamedLWLockTrancheRequestArray,
645 : i * sizeof(NamedLWLockTrancheRequest));
646 0 : NamedLWLockTrancheRequestsAllocated = i;
647 : }
648 :
649 14 : request = &NamedLWLockTrancheRequestArray[NamedLWLockTrancheRequests];
650 14 : strlcpy(request->tranche_name, tranche_name, NAMEDATALEN);
651 14 : request->num_lwlocks = num_lwlocks;
652 14 : NamedLWLockTrancheRequests++;
653 14 : }
654 :
655 : /*
656 : * LWLockInitialize - initialize a new lwlock; it's initially unlocked
657 : */
658 : void
659 24636708 : LWLockInitialize(LWLock *lock, int tranche_id)
660 : {
661 : /* verify the tranche_id is valid */
662 24636708 : (void) GetLWTrancheName(tranche_id);
663 :
664 24636708 : pg_atomic_init_u32(&lock->state, LW_FLAG_RELEASE_OK);
665 : #ifdef LOCK_DEBUG
666 : pg_atomic_init_u32(&lock->nwaiters, 0);
667 : #endif
668 24636708 : lock->tranche = tranche_id;
669 24636708 : proclist_init(&lock->waiters);
670 24636708 : }
671 :
672 : /*
673 : * Report start of wait event for light-weight locks.
674 : *
675 : * This function will be used by all the light-weight lock calls which
676 : * needs to wait to acquire the lock. This function distinguishes wait
677 : * event based on tranche and lock id.
678 : */
679 : static inline void
680 3172830 : LWLockReportWaitStart(LWLock *lock)
681 : {
682 3172830 : pgstat_report_wait_start(PG_WAIT_LWLOCK | lock->tranche);
683 3172830 : }
684 :
685 : /*
686 : * Report end of wait event for light-weight locks.
687 : */
688 : static inline void
689 3172830 : LWLockReportWaitEnd(void)
690 : {
691 3172830 : pgstat_report_wait_end();
692 3172830 : }
693 :
694 : /*
695 : * Return the name of an LWLock tranche.
696 : */
697 : static const char *
698 24636802 : GetLWTrancheName(uint16 trancheId)
699 : {
700 : /* Built-in tranche or individual LWLock? */
701 24636802 : if (trancheId < LWTRANCHE_FIRST_USER_DEFINED)
702 24636066 : return BuiltinTrancheNames[trancheId];
703 :
704 : /*
705 : * We only ever add new entries to LWLockTrancheNames, so most lookups can
706 : * avoid taking the spinlock as long as the backend-local counter
707 : * (LocalLWLockCounter) is greater than the requested tranche ID. Else,
708 : * we need to first update the backend-local counter with ShmemLock held
709 : * before attempting the lookup again. In practice, the latter case is
710 : * probably rare.
711 : */
712 736 : if (trancheId >= LocalLWLockCounter)
713 : {
714 0 : SpinLockAcquire(ShmemLock);
715 0 : LocalLWLockCounter = *LWLockCounter;
716 0 : SpinLockRelease(ShmemLock);
717 :
718 0 : if (trancheId >= LocalLWLockCounter)
719 0 : elog(ERROR, "tranche %d is not registered", trancheId);
720 : }
721 :
722 : /*
723 : * It's an extension tranche, so look in LWLockTrancheNames.
724 : */
725 736 : trancheId -= LWTRANCHE_FIRST_USER_DEFINED;
726 :
727 736 : return LWLockTrancheNames[trancheId];
728 : }
729 :
730 : /*
731 : * Return an identifier for an LWLock based on the wait class and event.
732 : */
733 : const char *
734 94 : GetLWLockIdentifier(uint32 classId, uint16 eventId)
735 : {
736 : Assert(classId == PG_WAIT_LWLOCK);
737 : /* The event IDs are just tranche numbers. */
738 94 : return GetLWTrancheName(eventId);
739 : }
740 :
741 : /*
742 : * Internal function that tries to atomically acquire the lwlock in the passed
743 : * in mode.
744 : *
745 : * This function will not block waiting for a lock to become free - that's the
746 : * caller's job.
747 : *
748 : * Returns true if the lock isn't free and we need to wait.
749 : */
750 : static bool
751 757368250 : LWLockAttemptLock(LWLock *lock, LWLockMode mode)
752 : {
753 : uint32 old_state;
754 :
755 : Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
756 :
757 : /*
758 : * Read once outside the loop, later iterations will get the newer value
759 : * via compare & exchange.
760 : */
761 757368250 : old_state = pg_atomic_read_u32(&lock->state);
762 :
763 : /* loop until we've determined whether we could acquire the lock or not */
764 : while (true)
765 1049430 : {
766 : uint32 desired_state;
767 : bool lock_free;
768 :
769 758417680 : desired_state = old_state;
770 :
771 758417680 : if (mode == LW_EXCLUSIVE)
772 : {
773 464309204 : lock_free = (old_state & LW_LOCK_MASK) == 0;
774 464309204 : if (lock_free)
775 461802494 : desired_state += LW_VAL_EXCLUSIVE;
776 : }
777 : else
778 : {
779 294108476 : lock_free = (old_state & LW_VAL_EXCLUSIVE) == 0;
780 294108476 : if (lock_free)
781 289950938 : desired_state += LW_VAL_SHARED;
782 : }
783 :
784 : /*
785 : * Attempt to swap in the state we are expecting. If we didn't see
786 : * lock to be free, that's just the old value. If we saw it as free,
787 : * we'll attempt to mark it acquired. The reason that we always swap
788 : * in the value is that this doubles as a memory barrier. We could try
789 : * to be smarter and only swap in values if we saw the lock as free,
790 : * but benchmark haven't shown it as beneficial so far.
791 : *
792 : * Retry if the value changed since we last looked at it.
793 : */
794 758417680 : if (pg_atomic_compare_exchange_u32(&lock->state,
795 : &old_state, desired_state))
796 : {
797 757368250 : if (lock_free)
798 : {
799 : /* Great! Got the lock. */
800 : #ifdef LOCK_DEBUG
801 : if (mode == LW_EXCLUSIVE)
802 : lock->owner = MyProc;
803 : #endif
804 751036336 : return false;
805 : }
806 : else
807 6331914 : return true; /* somebody else has the lock */
808 : }
809 : }
810 : pg_unreachable();
811 : }
812 :
813 : /*
814 : * Lock the LWLock's wait list against concurrent activity.
815 : *
816 : * NB: even though the wait list is locked, non-conflicting lock operations
817 : * may still happen concurrently.
818 : *
819 : * Time spent holding mutex should be short!
820 : */
821 : static void
822 13258780 : LWLockWaitListLock(LWLock *lock)
823 : {
824 : uint32 old_state;
825 : #ifdef LWLOCK_STATS
826 : lwlock_stats *lwstats;
827 : uint32 delays = 0;
828 :
829 : lwstats = get_lwlock_stats_entry(lock);
830 : #endif
831 :
832 : while (true)
833 : {
834 : /* always try once to acquire lock directly */
835 13559194 : old_state = pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_LOCKED);
836 13559194 : if (!(old_state & LW_FLAG_LOCKED))
837 13258780 : break; /* got lock */
838 :
839 : /* and then spin without atomic operations until lock is released */
840 : {
841 : SpinDelayStatus delayStatus;
842 :
843 300414 : init_local_spin_delay(&delayStatus);
844 :
845 841142 : while (old_state & LW_FLAG_LOCKED)
846 : {
847 540728 : perform_spin_delay(&delayStatus);
848 540728 : old_state = pg_atomic_read_u32(&lock->state);
849 : }
850 : #ifdef LWLOCK_STATS
851 : delays += delayStatus.delays;
852 : #endif
853 300414 : finish_spin_delay(&delayStatus);
854 : }
855 :
856 : /*
857 : * Retry. The lock might obviously already be re-acquired by the time
858 : * we're attempting to get it again.
859 : */
860 : }
861 :
862 : #ifdef LWLOCK_STATS
863 : lwstats->spin_delay_count += delays;
864 : #endif
865 13258780 : }
866 :
867 : /*
868 : * Unlock the LWLock's wait list.
869 : *
870 : * Note that it can be more efficient to manipulate flags and release the
871 : * locks in a single atomic operation.
872 : */
873 : static void
874 9181302 : LWLockWaitListUnlock(LWLock *lock)
875 : {
876 : uint32 old_state PG_USED_FOR_ASSERTS_ONLY;
877 :
878 9181302 : old_state = pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_LOCKED);
879 :
880 : Assert(old_state & LW_FLAG_LOCKED);
881 9181302 : }
882 :
883 : /*
884 : * Wakeup all the lockers that currently have a chance to acquire the lock.
885 : */
886 : static void
887 4077478 : LWLockWakeup(LWLock *lock)
888 : {
889 : bool new_release_ok;
890 4077478 : bool wokeup_somebody = false;
891 : proclist_head wakeup;
892 : proclist_mutable_iter iter;
893 :
894 4077478 : proclist_init(&wakeup);
895 :
896 4077478 : new_release_ok = true;
897 :
898 : /* lock wait list while collecting backends to wake up */
899 4077478 : LWLockWaitListLock(lock);
900 :
901 6270994 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
902 : {
903 3275506 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
904 :
905 3275506 : if (wokeup_somebody && waiter->lwWaitMode == LW_EXCLUSIVE)
906 19656 : continue;
907 :
908 3255850 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
909 3255850 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
910 :
911 3255850 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
912 : {
913 : /*
914 : * Prevent additional wakeups until retryer gets to run. Backends
915 : * that are just waiting for the lock to become free don't retry
916 : * automatically.
917 : */
918 3068690 : new_release_ok = false;
919 :
920 : /*
921 : * Don't wakeup (further) exclusive locks.
922 : */
923 3068690 : wokeup_somebody = true;
924 : }
925 :
926 : /*
927 : * Signal that the process isn't on the wait list anymore. This allows
928 : * LWLockDequeueSelf() to remove itself of the waitlist with a
929 : * proclist_delete(), rather than having to check if it has been
930 : * removed from the list.
931 : */
932 : Assert(waiter->lwWaiting == LW_WS_WAITING);
933 3255850 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
934 :
935 : /*
936 : * Once we've woken up an exclusive lock, there's no point in waking
937 : * up anybody else.
938 : */
939 3255850 : if (waiter->lwWaitMode == LW_EXCLUSIVE)
940 1081990 : break;
941 : }
942 :
943 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS);
944 :
945 : /* unset required flags, and release lock, in one fell swoop */
946 : {
947 : uint32 old_state;
948 : uint32 desired_state;
949 :
950 4077478 : old_state = pg_atomic_read_u32(&lock->state);
951 : while (true)
952 : {
953 4243928 : desired_state = old_state;
954 :
955 : /* compute desired flags */
956 :
957 4243928 : if (new_release_ok)
958 1369390 : desired_state |= LW_FLAG_RELEASE_OK;
959 : else
960 2874538 : desired_state &= ~LW_FLAG_RELEASE_OK;
961 :
962 4243928 : if (proclist_is_empty(&wakeup))
963 1286380 : desired_state &= ~LW_FLAG_HAS_WAITERS;
964 :
965 4243928 : desired_state &= ~LW_FLAG_LOCKED; /* release lock */
966 :
967 4243928 : if (pg_atomic_compare_exchange_u32(&lock->state, &old_state,
968 : desired_state))
969 4077478 : break;
970 : }
971 : }
972 :
973 : /* Awaken any waiters I removed from the queue. */
974 7333328 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
975 : {
976 3255850 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
977 :
978 : LOG_LWDEBUG("LWLockRelease", lock, "release waiter");
979 3255850 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
980 :
981 : /*
982 : * Guarantee that lwWaiting being unset only becomes visible once the
983 : * unlink from the link has completed. Otherwise the target backend
984 : * could be woken up for other reason and enqueue for a new lock - if
985 : * that happens before the list unlink happens, the list would end up
986 : * being corrupted.
987 : *
988 : * The barrier pairs with the LWLockWaitListLock() when enqueuing for
989 : * another lock.
990 : */
991 3255850 : pg_write_barrier();
992 3255850 : waiter->lwWaiting = LW_WS_NOT_WAITING;
993 3255850 : PGSemaphoreUnlock(waiter->sem);
994 : }
995 4077478 : }
996 :
997 : /*
998 : * Add ourselves to the end of the queue.
999 : *
1000 : * NB: Mode can be LW_WAIT_UNTIL_FREE here!
1001 : */
1002 : static void
1003 3505912 : LWLockQueueSelf(LWLock *lock, LWLockMode mode)
1004 : {
1005 : /*
1006 : * If we don't have a PGPROC structure, there's no way to wait. This
1007 : * should never occur, since MyProc should only be null during shared
1008 : * memory initialization.
1009 : */
1010 3505912 : if (MyProc == NULL)
1011 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1012 :
1013 3505912 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
1014 0 : elog(PANIC, "queueing for lock while waiting on another one");
1015 :
1016 3505912 : LWLockWaitListLock(lock);
1017 :
1018 : /* setting the flag is protected by the spinlock */
1019 3505912 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_HAS_WAITERS);
1020 :
1021 3505912 : MyProc->lwWaiting = LW_WS_WAITING;
1022 3505912 : MyProc->lwWaitMode = mode;
1023 :
1024 : /* LW_WAIT_UNTIL_FREE waiters are always at the front of the queue */
1025 3505912 : if (mode == LW_WAIT_UNTIL_FREE)
1026 194334 : proclist_push_head(&lock->waiters, MyProcNumber, lwWaitLink);
1027 : else
1028 3311578 : proclist_push_tail(&lock->waiters, MyProcNumber, lwWaitLink);
1029 :
1030 : /* Can release the mutex now */
1031 3505912 : LWLockWaitListUnlock(lock);
1032 :
1033 : #ifdef LOCK_DEBUG
1034 : pg_atomic_fetch_add_u32(&lock->nwaiters, 1);
1035 : #endif
1036 3505912 : }
1037 :
1038 : /*
1039 : * Remove ourselves from the waitlist.
1040 : *
1041 : * This is used if we queued ourselves because we thought we needed to sleep
1042 : * but, after further checking, we discovered that we don't actually need to
1043 : * do so.
1044 : */
1045 : static void
1046 333082 : LWLockDequeueSelf(LWLock *lock)
1047 : {
1048 : bool on_waitlist;
1049 :
1050 : #ifdef LWLOCK_STATS
1051 : lwlock_stats *lwstats;
1052 :
1053 : lwstats = get_lwlock_stats_entry(lock);
1054 :
1055 : lwstats->dequeue_self_count++;
1056 : #endif
1057 :
1058 333082 : LWLockWaitListLock(lock);
1059 :
1060 : /*
1061 : * Remove ourselves from the waitlist, unless we've already been removed.
1062 : * The removal happens with the wait list lock held, so there's no race in
1063 : * this check.
1064 : */
1065 333082 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
1066 333082 : if (on_waitlist)
1067 252434 : proclist_delete(&lock->waiters, MyProcNumber, lwWaitLink);
1068 :
1069 333082 : if (proclist_is_empty(&lock->waiters) &&
1070 315844 : (pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS) != 0)
1071 : {
1072 315348 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_HAS_WAITERS);
1073 : }
1074 :
1075 : /* XXX: combine with fetch_and above? */
1076 333082 : LWLockWaitListUnlock(lock);
1077 :
1078 : /* clear waiting state again, nice for debugging */
1079 333082 : if (on_waitlist)
1080 252434 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
1081 : else
1082 : {
1083 80648 : int extraWaits = 0;
1084 :
1085 : /*
1086 : * Somebody else dequeued us and has or will wake us up. Deal with the
1087 : * superfluous absorption of a wakeup.
1088 : */
1089 :
1090 : /*
1091 : * Reset RELEASE_OK flag if somebody woke us before we removed
1092 : * ourselves - they'll have set it to false.
1093 : */
1094 80648 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_RELEASE_OK);
1095 :
1096 : /*
1097 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
1098 : * get reset at some inconvenient point later. Most of the time this
1099 : * will immediately return.
1100 : */
1101 : for (;;)
1102 : {
1103 80648 : PGSemaphoreLock(MyProc->sem);
1104 80648 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
1105 80648 : break;
1106 0 : extraWaits++;
1107 : }
1108 :
1109 : /*
1110 : * Fix the process wait semaphore's count for any absorbed wakeups.
1111 : */
1112 80648 : while (extraWaits-- > 0)
1113 0 : PGSemaphoreUnlock(MyProc->sem);
1114 : }
1115 :
1116 : #ifdef LOCK_DEBUG
1117 : {
1118 : /* not waiting anymore */
1119 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1120 :
1121 : Assert(nwaiters < MAX_BACKENDS);
1122 : }
1123 : #endif
1124 333082 : }
1125 :
1126 : /*
1127 : * LWLockAcquire - acquire a lightweight lock in the specified mode
1128 : *
1129 : * If the lock is not available, sleep until it is. Returns true if the lock
1130 : * was available immediately, false if we had to sleep.
1131 : *
1132 : * Side effect: cancel/die interrupts are held off until lock release.
1133 : */
1134 : bool
1135 745964966 : LWLockAcquire(LWLock *lock, LWLockMode mode)
1136 : {
1137 745964966 : PGPROC *proc = MyProc;
1138 745964966 : bool result = true;
1139 745964966 : int extraWaits = 0;
1140 : #ifdef LWLOCK_STATS
1141 : lwlock_stats *lwstats;
1142 :
1143 : lwstats = get_lwlock_stats_entry(lock);
1144 : #endif
1145 :
1146 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1147 :
1148 : PRINT_LWDEBUG("LWLockAcquire", lock, mode);
1149 :
1150 : #ifdef LWLOCK_STATS
1151 : /* Count lock acquisition attempts */
1152 : if (mode == LW_EXCLUSIVE)
1153 : lwstats->ex_acquire_count++;
1154 : else
1155 : lwstats->sh_acquire_count++;
1156 : #endif /* LWLOCK_STATS */
1157 :
1158 : /*
1159 : * We can't wait if we haven't got a PGPROC. This should only occur
1160 : * during bootstrap or shared memory initialization. Put an Assert here
1161 : * to catch unsafe coding practices.
1162 : */
1163 : Assert(!(proc == NULL && IsUnderPostmaster));
1164 :
1165 : /* Ensure we will have room to remember the lock */
1166 745964966 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1167 0 : elog(ERROR, "too many LWLocks taken");
1168 :
1169 : /*
1170 : * Lock out cancel/die interrupts until we exit the code section protected
1171 : * by the LWLock. This ensures that interrupts will not interfere with
1172 : * manipulations of data structures in shared memory.
1173 : */
1174 745964966 : HOLD_INTERRUPTS();
1175 :
1176 : /*
1177 : * Loop here to try to acquire lock after each time we are signaled by
1178 : * LWLockRelease.
1179 : *
1180 : * NOTE: it might seem better to have LWLockRelease actually grant us the
1181 : * lock, rather than retrying and possibly having to go back to sleep. But
1182 : * in practice that is no good because it means a process swap for every
1183 : * lock acquisition when two or more processes are contending for the same
1184 : * lock. Since LWLocks are normally used to protect not-very-long
1185 : * sections of computation, a process needs to be able to acquire and
1186 : * release the same lock many times during a single CPU time slice, even
1187 : * in the presence of contention. The efficiency of being able to do that
1188 : * outweighs the inefficiency of sometimes wasting a process dispatch
1189 : * cycle because the lock is not free when a released waiter finally gets
1190 : * to run. See pgsql-hackers archives for 29-Dec-01.
1191 : */
1192 : for (;;)
1193 2984024 : {
1194 : bool mustwait;
1195 :
1196 : /*
1197 : * Try to grab the lock the first time, we're not in the waitqueue
1198 : * yet/anymore.
1199 : */
1200 748948990 : mustwait = LWLockAttemptLock(lock, mode);
1201 :
1202 748948990 : if (!mustwait)
1203 : {
1204 : LOG_LWDEBUG("LWLockAcquire", lock, "immediately acquired lock");
1205 745637412 : break; /* got the lock */
1206 : }
1207 :
1208 : /*
1209 : * Ok, at this point we couldn't grab the lock on the first try. We
1210 : * cannot simply queue ourselves to the end of the list and wait to be
1211 : * woken up because by now the lock could long have been released.
1212 : * Instead add us to the queue and try to grab the lock again. If we
1213 : * succeed we need to revert the queuing and be happy, otherwise we
1214 : * recheck the lock. If we still couldn't grab it, we know that the
1215 : * other locker will see our queue entries when releasing since they
1216 : * existed before we checked for the lock.
1217 : */
1218 :
1219 : /* add to the queue */
1220 3311578 : LWLockQueueSelf(lock, mode);
1221 :
1222 : /* we're now guaranteed to be woken up if necessary */
1223 3311578 : mustwait = LWLockAttemptLock(lock, mode);
1224 :
1225 : /* ok, grabbed the lock the second time round, need to undo queueing */
1226 3311578 : if (!mustwait)
1227 : {
1228 : LOG_LWDEBUG("LWLockAcquire", lock, "acquired, undoing queue");
1229 :
1230 327554 : LWLockDequeueSelf(lock);
1231 327554 : break;
1232 : }
1233 :
1234 : /*
1235 : * Wait until awakened.
1236 : *
1237 : * It is possible that we get awakened for a reason other than being
1238 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1239 : * we've gotten the LWLock, re-increment the sema by the number of
1240 : * additional signals received.
1241 : */
1242 : LOG_LWDEBUG("LWLockAcquire", lock, "waiting");
1243 :
1244 : #ifdef LWLOCK_STATS
1245 : lwstats->block_count++;
1246 : #endif
1247 :
1248 2984024 : LWLockReportWaitStart(lock);
1249 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1250 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1251 :
1252 : for (;;)
1253 : {
1254 2984024 : PGSemaphoreLock(proc->sem);
1255 2984024 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1256 2984024 : break;
1257 0 : extraWaits++;
1258 : }
1259 :
1260 : /* Retrying, allow LWLockRelease to release waiters again. */
1261 2984024 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_RELEASE_OK);
1262 :
1263 : #ifdef LOCK_DEBUG
1264 : {
1265 : /* not waiting anymore */
1266 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1267 :
1268 : Assert(nwaiters < MAX_BACKENDS);
1269 : }
1270 : #endif
1271 :
1272 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1273 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1274 2984024 : LWLockReportWaitEnd();
1275 :
1276 : LOG_LWDEBUG("LWLockAcquire", lock, "awakened");
1277 :
1278 : /* Now loop back and try to acquire lock again. */
1279 2984024 : result = false;
1280 : }
1281 :
1282 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_ENABLED())
1283 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE(T_NAME(lock), mode);
1284 :
1285 : /* Add lock to list of locks held by this backend */
1286 745964966 : held_lwlocks[num_held_lwlocks].lock = lock;
1287 745964966 : held_lwlocks[num_held_lwlocks++].mode = mode;
1288 :
1289 : /*
1290 : * Fix the process wait semaphore's count for any absorbed wakeups.
1291 : */
1292 745964966 : while (extraWaits-- > 0)
1293 0 : PGSemaphoreUnlock(proc->sem);
1294 :
1295 745964966 : return result;
1296 : }
1297 :
1298 : /*
1299 : * LWLockConditionalAcquire - acquire a lightweight lock in the specified mode
1300 : *
1301 : * If the lock is not available, return false with no side-effects.
1302 : *
1303 : * If successful, cancel/die interrupts are held off until lock release.
1304 : */
1305 : bool
1306 4826502 : LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
1307 : {
1308 : bool mustwait;
1309 :
1310 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1311 :
1312 : PRINT_LWDEBUG("LWLockConditionalAcquire", lock, mode);
1313 :
1314 : /* Ensure we will have room to remember the lock */
1315 4826502 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1316 0 : elog(ERROR, "too many LWLocks taken");
1317 :
1318 : /*
1319 : * Lock out cancel/die interrupts until we exit the code section protected
1320 : * by the LWLock. This ensures that interrupts will not interfere with
1321 : * manipulations of data structures in shared memory.
1322 : */
1323 4826502 : HOLD_INTERRUPTS();
1324 :
1325 : /* Check for the lock */
1326 4826502 : mustwait = LWLockAttemptLock(lock, mode);
1327 :
1328 4826502 : if (mustwait)
1329 : {
1330 : /* Failed to get lock, so release interrupt holdoff */
1331 12070 : RESUME_INTERRUPTS();
1332 :
1333 : LOG_LWDEBUG("LWLockConditionalAcquire", lock, "failed");
1334 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL_ENABLED())
1335 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL(T_NAME(lock), mode);
1336 : }
1337 : else
1338 : {
1339 : /* Add lock to list of locks held by this backend */
1340 4814432 : held_lwlocks[num_held_lwlocks].lock = lock;
1341 4814432 : held_lwlocks[num_held_lwlocks++].mode = mode;
1342 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_ENABLED())
1343 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE(T_NAME(lock), mode);
1344 : }
1345 4826502 : return !mustwait;
1346 : }
1347 :
1348 : /*
1349 : * LWLockAcquireOrWait - Acquire lock, or wait until it's free
1350 : *
1351 : * The semantics of this function are a bit funky. If the lock is currently
1352 : * free, it is acquired in the given mode, and the function returns true. If
1353 : * the lock isn't immediately free, the function waits until it is released
1354 : * and returns false, but does not acquire the lock.
1355 : *
1356 : * This is currently used for WALWriteLock: when a backend flushes the WAL,
1357 : * holding WALWriteLock, it can flush the commit records of many other
1358 : * backends as a side-effect. Those other backends need to wait until the
1359 : * flush finishes, but don't need to acquire the lock anymore. They can just
1360 : * wake up, observe that their records have already been flushed, and return.
1361 : */
1362 : bool
1363 268926 : LWLockAcquireOrWait(LWLock *lock, LWLockMode mode)
1364 : {
1365 268926 : PGPROC *proc = MyProc;
1366 : bool mustwait;
1367 268926 : int extraWaits = 0;
1368 : #ifdef LWLOCK_STATS
1369 : lwlock_stats *lwstats;
1370 :
1371 : lwstats = get_lwlock_stats_entry(lock);
1372 : #endif
1373 :
1374 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1375 :
1376 : PRINT_LWDEBUG("LWLockAcquireOrWait", lock, mode);
1377 :
1378 : /* Ensure we will have room to remember the lock */
1379 268926 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1380 0 : elog(ERROR, "too many LWLocks taken");
1381 :
1382 : /*
1383 : * Lock out cancel/die interrupts until we exit the code section protected
1384 : * by the LWLock. This ensures that interrupts will not interfere with
1385 : * manipulations of data structures in shared memory.
1386 : */
1387 268926 : HOLD_INTERRUPTS();
1388 :
1389 : /*
1390 : * NB: We're using nearly the same twice-in-a-row lock acquisition
1391 : * protocol as LWLockAcquire(). Check its comments for details.
1392 : */
1393 268926 : mustwait = LWLockAttemptLock(lock, mode);
1394 :
1395 268926 : if (mustwait)
1396 : {
1397 12254 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1398 :
1399 12254 : mustwait = LWLockAttemptLock(lock, mode);
1400 :
1401 12254 : if (mustwait)
1402 : {
1403 : /*
1404 : * Wait until awakened. Like in LWLockAcquire, be prepared for
1405 : * bogus wakeups.
1406 : */
1407 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "waiting");
1408 :
1409 : #ifdef LWLOCK_STATS
1410 : lwstats->block_count++;
1411 : #endif
1412 :
1413 11988 : LWLockReportWaitStart(lock);
1414 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1415 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1416 :
1417 : for (;;)
1418 : {
1419 11988 : PGSemaphoreLock(proc->sem);
1420 11988 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1421 11988 : break;
1422 0 : extraWaits++;
1423 : }
1424 :
1425 : #ifdef LOCK_DEBUG
1426 : {
1427 : /* not waiting anymore */
1428 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1429 :
1430 : Assert(nwaiters < MAX_BACKENDS);
1431 : }
1432 : #endif
1433 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1434 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1435 11988 : LWLockReportWaitEnd();
1436 :
1437 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "awakened");
1438 : }
1439 : else
1440 : {
1441 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "acquired, undoing queue");
1442 :
1443 : /*
1444 : * Got lock in the second attempt, undo queueing. We need to treat
1445 : * this as having successfully acquired the lock, otherwise we'd
1446 : * not necessarily wake up people we've prevented from acquiring
1447 : * the lock.
1448 : */
1449 266 : LWLockDequeueSelf(lock);
1450 : }
1451 : }
1452 :
1453 : /*
1454 : * Fix the process wait semaphore's count for any absorbed wakeups.
1455 : */
1456 268926 : while (extraWaits-- > 0)
1457 0 : PGSemaphoreUnlock(proc->sem);
1458 :
1459 268926 : if (mustwait)
1460 : {
1461 : /* Failed to get lock, so release interrupt holdoff */
1462 11988 : RESUME_INTERRUPTS();
1463 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "failed");
1464 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL_ENABLED())
1465 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL(T_NAME(lock), mode);
1466 : }
1467 : else
1468 : {
1469 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "succeeded");
1470 : /* Add lock to list of locks held by this backend */
1471 256938 : held_lwlocks[num_held_lwlocks].lock = lock;
1472 256938 : held_lwlocks[num_held_lwlocks++].mode = mode;
1473 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_ENABLED())
1474 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT(T_NAME(lock), mode);
1475 : }
1476 :
1477 268926 : return !mustwait;
1478 : }
1479 :
1480 : /*
1481 : * Does the lwlock in its current state need to wait for the variable value to
1482 : * change?
1483 : *
1484 : * If we don't need to wait, and it's because the value of the variable has
1485 : * changed, store the current value in newval.
1486 : *
1487 : * *result is set to true if the lock was free, and false otherwise.
1488 : */
1489 : static bool
1490 7596340 : LWLockConflictsWithVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1491 : uint64 *newval, bool *result)
1492 : {
1493 : bool mustwait;
1494 : uint64 value;
1495 :
1496 : /*
1497 : * Test first to see if it the slot is free right now.
1498 : *
1499 : * XXX: the unique caller of this routine, WaitXLogInsertionsToFinish()
1500 : * via LWLockWaitForVar(), uses an implied barrier with a spinlock before
1501 : * this, so we don't need a memory barrier here as far as the current
1502 : * usage is concerned. But that might not be safe in general.
1503 : */
1504 7596340 : mustwait = (pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE) != 0;
1505 :
1506 7596340 : if (!mustwait)
1507 : {
1508 5081660 : *result = true;
1509 5081660 : return false;
1510 : }
1511 :
1512 2514680 : *result = false;
1513 :
1514 : /*
1515 : * Reading this value atomically is safe even on platforms where uint64
1516 : * cannot be read without observing a torn value.
1517 : */
1518 2514680 : value = pg_atomic_read_u64(valptr);
1519 :
1520 2514680 : if (value != oldval)
1521 : {
1522 2155782 : mustwait = false;
1523 2155782 : *newval = value;
1524 : }
1525 : else
1526 : {
1527 358898 : mustwait = true;
1528 : }
1529 :
1530 2514680 : return mustwait;
1531 : }
1532 :
1533 : /*
1534 : * LWLockWaitForVar - Wait until lock is free, or a variable is updated.
1535 : *
1536 : * If the lock is held and *valptr equals oldval, waits until the lock is
1537 : * either freed, or the lock holder updates *valptr by calling
1538 : * LWLockUpdateVar. If the lock is free on exit (immediately or after
1539 : * waiting), returns true. If the lock is still held, but *valptr no longer
1540 : * matches oldval, returns false and sets *newval to the current value in
1541 : * *valptr.
1542 : *
1543 : * Note: this function ignores shared lock holders; if the lock is held
1544 : * in shared mode, returns 'true'.
1545 : *
1546 : * Be aware that LWLockConflictsWithVar() does not include a memory barrier,
1547 : * hence the caller of this function may want to rely on an explicit barrier or
1548 : * an implied barrier via spinlock or LWLock to avoid memory ordering issues.
1549 : */
1550 : bool
1551 7237442 : LWLockWaitForVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1552 : uint64 *newval)
1553 : {
1554 7237442 : PGPROC *proc = MyProc;
1555 7237442 : int extraWaits = 0;
1556 7237442 : bool result = false;
1557 : #ifdef LWLOCK_STATS
1558 : lwlock_stats *lwstats;
1559 :
1560 : lwstats = get_lwlock_stats_entry(lock);
1561 : #endif
1562 :
1563 : PRINT_LWDEBUG("LWLockWaitForVar", lock, LW_WAIT_UNTIL_FREE);
1564 :
1565 : /*
1566 : * Lock out cancel/die interrupts while we sleep on the lock. There is no
1567 : * cleanup mechanism to remove us from the wait queue if we got
1568 : * interrupted.
1569 : */
1570 7237442 : HOLD_INTERRUPTS();
1571 :
1572 : /*
1573 : * Loop here to check the lock's status after each time we are signaled.
1574 : */
1575 : for (;;)
1576 176818 : {
1577 : bool mustwait;
1578 :
1579 7414260 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1580 : &result);
1581 :
1582 7414260 : if (!mustwait)
1583 7232180 : break; /* the lock was free or value didn't match */
1584 :
1585 : /*
1586 : * Add myself to wait queue. Note that this is racy, somebody else
1587 : * could wakeup before we're finished queuing. NB: We're using nearly
1588 : * the same twice-in-a-row lock acquisition protocol as
1589 : * LWLockAcquire(). Check its comments for details. The only
1590 : * difference is that we also have to check the variable's values when
1591 : * checking the state of the lock.
1592 : */
1593 182080 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1594 :
1595 : /*
1596 : * Set RELEASE_OK flag, to make sure we get woken up as soon as the
1597 : * lock is released.
1598 : */
1599 182080 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_RELEASE_OK);
1600 :
1601 : /*
1602 : * We're now guaranteed to be woken up if necessary. Recheck the lock
1603 : * and variables state.
1604 : */
1605 182080 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1606 : &result);
1607 :
1608 : /* Ok, no conflict after we queued ourselves. Undo queueing. */
1609 182080 : if (!mustwait)
1610 : {
1611 : LOG_LWDEBUG("LWLockWaitForVar", lock, "free, undoing queue");
1612 :
1613 5262 : LWLockDequeueSelf(lock);
1614 5262 : break;
1615 : }
1616 :
1617 : /*
1618 : * Wait until awakened.
1619 : *
1620 : * It is possible that we get awakened for a reason other than being
1621 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1622 : * we've gotten the LWLock, re-increment the sema by the number of
1623 : * additional signals received.
1624 : */
1625 : LOG_LWDEBUG("LWLockWaitForVar", lock, "waiting");
1626 :
1627 : #ifdef LWLOCK_STATS
1628 : lwstats->block_count++;
1629 : #endif
1630 :
1631 176818 : LWLockReportWaitStart(lock);
1632 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1633 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), LW_EXCLUSIVE);
1634 :
1635 : for (;;)
1636 : {
1637 176818 : PGSemaphoreLock(proc->sem);
1638 176818 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1639 176818 : break;
1640 0 : extraWaits++;
1641 : }
1642 :
1643 : #ifdef LOCK_DEBUG
1644 : {
1645 : /* not waiting anymore */
1646 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1647 :
1648 : Assert(nwaiters < MAX_BACKENDS);
1649 : }
1650 : #endif
1651 :
1652 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1653 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), LW_EXCLUSIVE);
1654 176818 : LWLockReportWaitEnd();
1655 :
1656 : LOG_LWDEBUG("LWLockWaitForVar", lock, "awakened");
1657 :
1658 : /* Now loop back and check the status of the lock again. */
1659 : }
1660 :
1661 : /*
1662 : * Fix the process wait semaphore's count for any absorbed wakeups.
1663 : */
1664 7237442 : while (extraWaits-- > 0)
1665 0 : PGSemaphoreUnlock(proc->sem);
1666 :
1667 : /*
1668 : * Now okay to allow cancel/die interrupts.
1669 : */
1670 7237442 : RESUME_INTERRUPTS();
1671 :
1672 7237442 : return result;
1673 : }
1674 :
1675 :
1676 : /*
1677 : * LWLockUpdateVar - Update a variable and wake up waiters atomically
1678 : *
1679 : * Sets *valptr to 'val', and wakes up all processes waiting for us with
1680 : * LWLockWaitForVar(). It first sets the value atomically and then wakes up
1681 : * waiting processes so that any process calling LWLockWaitForVar() on the same
1682 : * lock is guaranteed to see the new value, and act accordingly.
1683 : *
1684 : * The caller must be holding the lock in exclusive mode.
1685 : */
1686 : void
1687 5342308 : LWLockUpdateVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1688 : {
1689 : proclist_head wakeup;
1690 : proclist_mutable_iter iter;
1691 :
1692 : PRINT_LWDEBUG("LWLockUpdateVar", lock, LW_EXCLUSIVE);
1693 :
1694 : /*
1695 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1696 : * that the variable is updated before waking up waiters.
1697 : */
1698 5342308 : pg_atomic_exchange_u64(valptr, val);
1699 :
1700 5342308 : proclist_init(&wakeup);
1701 :
1702 5342308 : LWLockWaitListLock(lock);
1703 :
1704 : Assert(pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE);
1705 :
1706 : /*
1707 : * See if there are any LW_WAIT_UNTIL_FREE waiters that need to be woken
1708 : * up. They are always in the front of the queue.
1709 : */
1710 5345646 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
1711 : {
1712 129832 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1713 :
1714 129832 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
1715 126494 : break;
1716 :
1717 3338 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
1718 3338 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
1719 :
1720 : /* see LWLockWakeup() */
1721 : Assert(waiter->lwWaiting == LW_WS_WAITING);
1722 3338 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
1723 : }
1724 :
1725 : /* We are done updating shared state of the lock itself. */
1726 5342308 : LWLockWaitListUnlock(lock);
1727 :
1728 : /*
1729 : * Awaken any waiters I removed from the queue.
1730 : */
1731 5345646 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1732 : {
1733 3338 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1734 :
1735 3338 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1736 : /* check comment in LWLockWakeup() about this barrier */
1737 3338 : pg_write_barrier();
1738 3338 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1739 3338 : PGSemaphoreUnlock(waiter->sem);
1740 : }
1741 5342308 : }
1742 :
1743 :
1744 : /*
1745 : * Stop treating lock as held by current backend.
1746 : *
1747 : * This is the code that can be shared between actually releasing a lock
1748 : * (LWLockRelease()) and just not tracking ownership of the lock anymore
1749 : * without releasing the lock (LWLockDisown()).
1750 : *
1751 : * Returns the mode in which the lock was held by the current backend.
1752 : *
1753 : * NB: This does not call RESUME_INTERRUPTS(), but leaves that responsibility
1754 : * of the caller.
1755 : *
1756 : * NB: This will leave lock->owner pointing to the current backend (if
1757 : * LOCK_DEBUG is set). This is somewhat intentional, as it makes it easier to
1758 : * debug cases of missing wakeups during lock release.
1759 : */
1760 : static inline LWLockMode
1761 751036336 : LWLockDisownInternal(LWLock *lock)
1762 : {
1763 : LWLockMode mode;
1764 : int i;
1765 :
1766 : /*
1767 : * Remove lock from list of locks held. Usually, but not always, it will
1768 : * be the latest-acquired lock; so search array backwards.
1769 : */
1770 834120548 : for (i = num_held_lwlocks; --i >= 0;)
1771 834120548 : if (lock == held_lwlocks[i].lock)
1772 751036336 : break;
1773 :
1774 751036336 : if (i < 0)
1775 0 : elog(ERROR, "lock %s is not held", T_NAME(lock));
1776 :
1777 751036336 : mode = held_lwlocks[i].mode;
1778 :
1779 751036336 : num_held_lwlocks--;
1780 834120548 : for (; i < num_held_lwlocks; i++)
1781 83084212 : held_lwlocks[i] = held_lwlocks[i + 1];
1782 :
1783 751036336 : return mode;
1784 : }
1785 :
1786 : /*
1787 : * Helper function to release lock, shared between LWLockRelease() and
1788 : * LWLockReleaseDisowned().
1789 : */
1790 : static void
1791 751036336 : LWLockReleaseInternal(LWLock *lock, LWLockMode mode)
1792 : {
1793 : uint32 oldstate;
1794 : bool check_waiters;
1795 :
1796 : /*
1797 : * Release my hold on lock, after that it can immediately be acquired by
1798 : * others, even if we still have to wakeup other waiters.
1799 : */
1800 751036336 : if (mode == LW_EXCLUSIVE)
1801 461431018 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_EXCLUSIVE);
1802 : else
1803 289605318 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_SHARED);
1804 :
1805 : /* nobody else can have that kind of lock */
1806 : Assert(!(oldstate & LW_VAL_EXCLUSIVE));
1807 :
1808 : if (TRACE_POSTGRESQL_LWLOCK_RELEASE_ENABLED())
1809 : TRACE_POSTGRESQL_LWLOCK_RELEASE(T_NAME(lock));
1810 :
1811 : /*
1812 : * We're still waiting for backends to get scheduled, don't wake them up
1813 : * again.
1814 : */
1815 751036336 : if ((oldstate & (LW_FLAG_HAS_WAITERS | LW_FLAG_RELEASE_OK)) ==
1816 4127172 : (LW_FLAG_HAS_WAITERS | LW_FLAG_RELEASE_OK) &&
1817 4127172 : (oldstate & LW_LOCK_MASK) == 0)
1818 4077478 : check_waiters = true;
1819 : else
1820 746958858 : check_waiters = false;
1821 :
1822 : /*
1823 : * As waking up waiters requires the spinlock to be acquired, only do so
1824 : * if necessary.
1825 : */
1826 751036336 : if (check_waiters)
1827 : {
1828 : /* XXX: remove before commit? */
1829 : LOG_LWDEBUG("LWLockRelease", lock, "releasing waiters");
1830 4077478 : LWLockWakeup(lock);
1831 : }
1832 751036336 : }
1833 :
1834 :
1835 : /*
1836 : * Stop treating lock as held by current backend.
1837 : *
1838 : * After calling this function it's the callers responsibility to ensure that
1839 : * the lock gets released (via LWLockReleaseDisowned()), even in case of an
1840 : * error. This only is desirable if the lock is going to be released in a
1841 : * different process than the process that acquired it.
1842 : */
1843 : void
1844 0 : LWLockDisown(LWLock *lock)
1845 : {
1846 0 : LWLockDisownInternal(lock);
1847 :
1848 0 : RESUME_INTERRUPTS();
1849 0 : }
1850 :
1851 : /*
1852 : * LWLockRelease - release a previously acquired lock
1853 : */
1854 : void
1855 751036336 : LWLockRelease(LWLock *lock)
1856 : {
1857 : LWLockMode mode;
1858 :
1859 751036336 : mode = LWLockDisownInternal(lock);
1860 :
1861 : PRINT_LWDEBUG("LWLockRelease", lock, mode);
1862 :
1863 751036336 : LWLockReleaseInternal(lock, mode);
1864 :
1865 : /*
1866 : * Now okay to allow cancel/die interrupts.
1867 : */
1868 751036336 : RESUME_INTERRUPTS();
1869 751036336 : }
1870 :
1871 : /*
1872 : * Release lock previously disowned with LWLockDisown().
1873 : */
1874 : void
1875 0 : LWLockReleaseDisowned(LWLock *lock, LWLockMode mode)
1876 : {
1877 0 : LWLockReleaseInternal(lock, mode);
1878 0 : }
1879 :
1880 : /*
1881 : * LWLockReleaseClearVar - release a previously acquired lock, reset variable
1882 : */
1883 : void
1884 29188408 : LWLockReleaseClearVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1885 : {
1886 : /*
1887 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1888 : * that the variable is updated before releasing the lock.
1889 : */
1890 29188408 : pg_atomic_exchange_u64(valptr, val);
1891 :
1892 29188408 : LWLockRelease(lock);
1893 29188408 : }
1894 :
1895 :
1896 : /*
1897 : * LWLockReleaseAll - release all currently-held locks
1898 : *
1899 : * Used to clean up after ereport(ERROR). An important difference between this
1900 : * function and retail LWLockRelease calls is that InterruptHoldoffCount is
1901 : * unchanged by this operation. This is necessary since InterruptHoldoffCount
1902 : * has been set to an appropriate level earlier in error recovery. We could
1903 : * decrement it below zero if we allow it to drop for each released lock!
1904 : */
1905 : void
1906 112364 : LWLockReleaseAll(void)
1907 : {
1908 112760 : while (num_held_lwlocks > 0)
1909 : {
1910 396 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
1911 :
1912 396 : LWLockRelease(held_lwlocks[num_held_lwlocks - 1].lock);
1913 : }
1914 112364 : }
1915 :
1916 :
1917 : /*
1918 : * ForEachLWLockHeldByMe - run a callback for each held lock
1919 : *
1920 : * This is meant as debug support only.
1921 : */
1922 : void
1923 0 : ForEachLWLockHeldByMe(void (*callback) (LWLock *, LWLockMode, void *),
1924 : void *context)
1925 : {
1926 : int i;
1927 :
1928 0 : for (i = 0; i < num_held_lwlocks; i++)
1929 0 : callback(held_lwlocks[i].lock, held_lwlocks[i].mode, context);
1930 0 : }
1931 :
1932 : /*
1933 : * LWLockHeldByMe - test whether my process holds a lock in any mode
1934 : *
1935 : * This is meant as debug support only.
1936 : */
1937 : bool
1938 0 : LWLockHeldByMe(LWLock *lock)
1939 : {
1940 : int i;
1941 :
1942 0 : for (i = 0; i < num_held_lwlocks; i++)
1943 : {
1944 0 : if (held_lwlocks[i].lock == lock)
1945 0 : return true;
1946 : }
1947 0 : return false;
1948 : }
1949 :
1950 : /*
1951 : * LWLockAnyHeldByMe - test whether my process holds any of an array of locks
1952 : *
1953 : * This is meant as debug support only.
1954 : */
1955 : bool
1956 0 : LWLockAnyHeldByMe(LWLock *lock, int nlocks, size_t stride)
1957 : {
1958 : char *held_lock_addr;
1959 : char *begin;
1960 : char *end;
1961 : int i;
1962 :
1963 0 : begin = (char *) lock;
1964 0 : end = begin + nlocks * stride;
1965 0 : for (i = 0; i < num_held_lwlocks; i++)
1966 : {
1967 0 : held_lock_addr = (char *) held_lwlocks[i].lock;
1968 0 : if (held_lock_addr >= begin &&
1969 0 : held_lock_addr < end &&
1970 0 : (held_lock_addr - begin) % stride == 0)
1971 0 : return true;
1972 : }
1973 0 : return false;
1974 : }
1975 :
1976 : /*
1977 : * LWLockHeldByMeInMode - test whether my process holds a lock in given mode
1978 : *
1979 : * This is meant as debug support only.
1980 : */
1981 : bool
1982 0 : LWLockHeldByMeInMode(LWLock *lock, LWLockMode mode)
1983 : {
1984 : int i;
1985 :
1986 0 : for (i = 0; i < num_held_lwlocks; i++)
1987 : {
1988 0 : if (held_lwlocks[i].lock == lock && held_lwlocks[i].mode == mode)
1989 0 : return true;
1990 : }
1991 0 : return false;
1992 : }
|
