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) 1996-2026, 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_WAKE_IN_PROGRESS ((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 waking %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_WAKE_IN_PROGRESS) != 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 3300 : NumLWLocksForNamedTranches(void)
383 : {
384 3300 : int numLocks = 0;
385 : int i;
386 :
387 3417 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
388 117 : numLocks += NamedLWLockTrancheRequestArray[i].num_lwlocks;
389 :
390 3300 : return numLocks;
391 : }
392 :
393 : /*
394 : * Compute shmem space needed for LWLocks and named tranches.
395 : */
396 : Size
397 3300 : LWLockShmemSize(void)
398 : {
399 : Size size;
400 3300 : 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 3300 : if (LocalNamedLWLockTrancheRequestArray)
409 0 : NamedLWLockTrancheRequestArray = LocalNamedLWLockTrancheRequestArray;
410 :
411 : /* Calculate total number of locks needed in the main array. */
412 3300 : numLocks += NumLWLocksForNamedTranches();
413 :
414 : /* Space for dynamic allocation counter. */
415 3300 : size = MAXALIGN(sizeof(int));
416 :
417 : /* Space for named tranches. */
418 3300 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, sizeof(char *)));
419 3300 : 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 3300 : size = add_size(size, mul_size(NamedLWLockTrancheRequests,
427 : sizeof(NamedLWLockTrancheRequest)));
428 :
429 : /* Space for the LWLock array, plus room for cache line alignment. */
430 3300 : size = add_size(size, LWLOCK_PADDED_SIZE);
431 3300 : size = add_size(size, mul_size(numLocks, sizeof(LWLockPadded)));
432 :
433 3300 : 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 1151 : CreateLWLocks(void)
442 : {
443 1151 : if (!IsUnderPostmaster)
444 : {
445 1151 : Size spaceLocks = LWLockShmemSize();
446 : char *ptr;
447 :
448 : /* Allocate space */
449 1151 : ptr = (char *) ShmemAlloc(spaceLocks);
450 :
451 : /* Initialize the dynamic-allocation counter for tranches */
452 1151 : LWLockCounter = (int *) ptr;
453 1151 : *LWLockCounter = LWTRANCHE_FIRST_USER_DEFINED;
454 1151 : ptr += MAXALIGN(sizeof(int));
455 :
456 : /* Initialize tranche names */
457 1151 : LWLockTrancheNames = (char **) ptr;
458 1151 : ptr += MAX_NAMED_TRANCHES * sizeof(char *);
459 295807 : for (int i = 0; i < MAX_NAMED_TRANCHES; i++)
460 : {
461 294656 : LWLockTrancheNames[i] = ptr;
462 294656 : 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 1151 : 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 8 : LocalNamedLWLockTrancheRequestArray = NamedLWLockTrancheRequestArray;
478 :
479 8 : memcpy(ptr, NamedLWLockTrancheRequestArray,
480 : NamedLWLockTrancheRequests * sizeof(NamedLWLockTrancheRequest));
481 8 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *) ptr;
482 8 : ptr += NamedLWLockTrancheRequests * sizeof(NamedLWLockTrancheRequest);
483 : }
484 :
485 : /* Ensure desired alignment of LWLock array */
486 1151 : ptr += LWLOCK_PADDED_SIZE - ((uintptr_t) ptr) % LWLOCK_PADDED_SIZE;
487 1151 : MainLWLockArray = (LWLockPadded *) ptr;
488 :
489 : /* Initialize all LWLocks */
490 1151 : InitializeLWLocks();
491 : }
492 1151 : }
493 :
494 : /*
495 : * Initialize LWLocks that are fixed and those belonging to named tranches.
496 : */
497 : static void
498 1151 : 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 65607 : for (id = 0, lock = MainLWLockArray; id < NUM_INDIVIDUAL_LWLOCKS; id++, lock++)
507 64456 : LWLockInitialize(&lock->lock, id);
508 :
509 : /* Initialize buffer mapping LWLocks in main array */
510 1151 : lock = MainLWLockArray + BUFFER_MAPPING_LWLOCK_OFFSET;
511 148479 : for (id = 0; id < NUM_BUFFER_PARTITIONS; id++, lock++)
512 147328 : LWLockInitialize(&lock->lock, LWTRANCHE_BUFFER_MAPPING);
513 :
514 : /* Initialize lmgrs' LWLocks in main array */
515 1151 : lock = MainLWLockArray + LOCK_MANAGER_LWLOCK_OFFSET;
516 19567 : for (id = 0; id < NUM_LOCK_PARTITIONS; id++, lock++)
517 18416 : LWLockInitialize(&lock->lock, LWTRANCHE_LOCK_MANAGER);
518 :
519 : /* Initialize predicate lmgrs' LWLocks in main array */
520 1151 : lock = MainLWLockArray + PREDICATELOCK_MANAGER_LWLOCK_OFFSET;
521 19567 : for (id = 0; id < NUM_PREDICATELOCK_PARTITIONS; id++, lock++)
522 18416 : 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 1151 : if (NamedLWLockTrancheRequests > 0)
529 : {
530 8 : lock = &MainLWLockArray[NUM_FIXED_LWLOCKS];
531 :
532 47 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
533 : {
534 : NamedLWLockTrancheRequest *request;
535 : int tranche;
536 :
537 39 : request = &NamedLWLockTrancheRequestArray[i];
538 39 : tranche = LWLockNewTrancheId(request->tranche_name);
539 :
540 78 : for (j = 0; j < request->num_lwlocks; j++, lock++)
541 39 : LWLockInitialize(&lock->lock, tranche);
542 : }
543 : }
544 1151 : }
545 :
546 : /*
547 : * InitLWLockAccess - initialize backend-local state needed to hold LWLocks
548 : */
549 : void
550 23339 : InitLWLockAccess(void)
551 : {
552 : #ifdef LWLOCK_STATS
553 : init_lwlock_stats();
554 : #endif
555 23339 : }
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 9 : 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 9 : lock_pos = NUM_FIXED_LWLOCKS;
577 41 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
578 : {
579 40 : if (strcmp(NamedLWLockTrancheRequestArray[i].tranche_name,
580 : tranche_name) == 0)
581 8 : return &MainLWLockArray[lock_pos];
582 :
583 32 : lock_pos += NamedLWLockTrancheRequestArray[i].num_lwlocks;
584 : }
585 :
586 1 : 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
596 283 : LWLockNewTrancheId(const char *name)
597 : {
598 : int result;
599 :
600 283 : if (!name)
601 1 : ereport(ERROR,
602 : (errcode(ERRCODE_INVALID_NAME),
603 : errmsg("tranche name cannot be NULL")));
604 :
605 282 : if (strlen(name) >= NAMEDATALEN)
606 1 : 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 : */
616 281 : SpinLockAcquire(ShmemLock);
617 :
618 281 : if (*LWLockCounter - LWTRANCHE_FIRST_USER_DEFINED >= MAX_NAMED_TRANCHES)
619 : {
620 1 : SpinLockRelease(ShmemLock);
621 1 : 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 :
627 280 : result = (*LWLockCounter)++;
628 280 : LocalLWLockCounter = *LWLockCounter;
629 280 : strlcpy(LWLockTrancheNames[result - LWTRANCHE_FIRST_USER_DEFINED], name, NAMEDATALEN);
630 :
631 280 : SpinLockRelease(ShmemLock);
632 :
633 280 : 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 39 : RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
650 : {
651 : NamedLWLockTrancheRequest *request;
652 : static int NamedLWLockTrancheRequestsAllocated;
653 :
654 39 : if (!process_shmem_requests_in_progress)
655 0 : elog(FATAL, "cannot request additional LWLocks outside shmem_request_hook");
656 :
657 39 : if (!tranche_name)
658 0 : ereport(ERROR,
659 : (errcode(ERRCODE_INVALID_NAME),
660 : errmsg("tranche name cannot be NULL")));
661 :
662 39 : 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 39 : if (NamedLWLockTrancheRequestArray == NULL)
670 : {
671 8 : NamedLWLockTrancheRequestsAllocated = 16;
672 8 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
673 8 : MemoryContextAlloc(TopMemoryContext,
674 : NamedLWLockTrancheRequestsAllocated
675 : * sizeof(NamedLWLockTrancheRequest));
676 : }
677 :
678 39 : if (NamedLWLockTrancheRequests >= NamedLWLockTrancheRequestsAllocated)
679 : {
680 1 : int i = pg_nextpower2_32(NamedLWLockTrancheRequests + 1);
681 :
682 1 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
683 1 : repalloc(NamedLWLockTrancheRequestArray,
684 : i * sizeof(NamedLWLockTrancheRequest));
685 1 : NamedLWLockTrancheRequestsAllocated = i;
686 : }
687 :
688 39 : request = &NamedLWLockTrancheRequestArray[NamedLWLockTrancheRequests];
689 39 : strlcpy(request->tranche_name, tranche_name, NAMEDATALEN);
690 39 : request->num_lwlocks = num_lwlocks;
691 39 : NamedLWLockTrancheRequests++;
692 39 : }
693 :
694 : /*
695 : * LWLockInitialize - initialize a new lwlock; it's initially unlocked
696 : */
697 : void
698 2323443 : LWLockInitialize(LWLock *lock, int tranche_id)
699 : {
700 : /* verify the tranche_id is valid */
701 2323443 : (void) GetLWTrancheName(tranche_id);
702 :
703 2323442 : pg_atomic_init_u32(&lock->state, 0);
704 : #ifdef LOCK_DEBUG
705 : pg_atomic_init_u32(&lock->nwaiters, 0);
706 : #endif
707 2323442 : lock->tranche = tranche_id;
708 2323442 : proclist_init(&lock->waiters);
709 2323442 : }
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 4639432 : LWLockReportWaitStart(LWLock *lock)
720 : {
721 4639432 : pgstat_report_wait_start(PG_WAIT_LWLOCK | lock->tranche);
722 4639432 : }
723 :
724 : /*
725 : * Report end of wait event for light-weight locks.
726 : */
727 : static inline void
728 4639432 : LWLockReportWaitEnd(void)
729 : {
730 4639432 : pgstat_report_wait_end();
731 4639432 : }
732 :
733 : /*
734 : * Return the name of an LWLock tranche.
735 : */
736 : static const char *
737 2323722 : GetLWTrancheName(uint16 trancheId)
738 : {
739 : /* Built-in tranche or individual LWLock? */
740 2323722 : if (trancheId < LWTRANCHE_FIRST_USER_DEFINED)
741 2322953 : 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 : */
751 769 : if (trancheId >= LocalLWLockCounter)
752 : {
753 1 : SpinLockAcquire(ShmemLock);
754 1 : LocalLWLockCounter = *LWLockCounter;
755 1 : SpinLockRelease(ShmemLock);
756 :
757 1 : if (trancheId >= LocalLWLockCounter)
758 1 : elog(ERROR, "tranche %d is not registered", trancheId);
759 : }
760 :
761 : /*
762 : * It's an extension tranche, so look in LWLockTrancheNames.
763 : */
764 768 : trancheId -= LWTRANCHE_FIRST_USER_DEFINED;
765 :
766 768 : return LWLockTrancheNames[trancheId];
767 : }
768 :
769 : /*
770 : * Return an identifier for an LWLock based on the wait class and event.
771 : */
772 : const char *
773 279 : GetLWLockIdentifier(uint32 classId, uint16 eventId)
774 : {
775 : Assert(classId == PG_WAIT_LWLOCK);
776 : /* The event IDs are just tranche numbers. */
777 279 : 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 310087649 : 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 310087649 : 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 261519 : {
805 : uint32 desired_state;
806 : bool lock_free;
807 :
808 310349168 : desired_state = old_state;
809 :
810 310349168 : if (mode == LW_EXCLUSIVE)
811 : {
812 215221764 : lock_free = (old_state & LW_LOCK_MASK) == 0;
813 215221764 : if (lock_free)
814 213272252 : desired_state += LW_VAL_EXCLUSIVE;
815 : }
816 : else
817 : {
818 95127404 : lock_free = (old_state & LW_VAL_EXCLUSIVE) == 0;
819 95127404 : if (lock_free)
820 87839132 : 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 310349168 : if (pg_atomic_compare_exchange_u32(&lock->state,
834 : &old_state, desired_state))
835 : {
836 310087649 : 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 300936953 : return false;
844 : }
845 : else
846 9150696 : 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 11756141 : 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 : /*
874 : * Always try once to acquire the lock directly, without setting up
875 : * the spin-delay infrastructure. The work necessary for that shows up
876 : * in profiles and is rarely necessary.
877 : */
878 11805301 : old_state = pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_LOCKED);
879 11805301 : if (likely(!(old_state & LW_FLAG_LOCKED)))
880 11756141 : break; /* got lock */
881 :
882 : /* and then spin without atomic operations until lock is released */
883 : {
884 : SpinDelayStatus delayStatus;
885 :
886 49160 : init_local_spin_delay(&delayStatus);
887 :
888 175315 : while (old_state & LW_FLAG_LOCKED)
889 : {
890 126155 : perform_spin_delay(&delayStatus);
891 126155 : old_state = pg_atomic_read_u32(&lock->state);
892 : }
893 : #ifdef LWLOCK_STATS
894 : delays += delayStatus.delays;
895 : #endif
896 49160 : finish_spin_delay(&delayStatus);
897 : }
898 :
899 : /*
900 : * Retry. The lock might obviously already be re-acquired by the time
901 : * we're attempting to get it again.
902 : */
903 : }
904 :
905 : #ifdef LWLOCK_STATS
906 : lwstats->spin_delay_count += delays;
907 : #endif
908 11756141 : }
909 :
910 : /*
911 : * Unlock the LWLock's wait list.
912 : *
913 : * Note that it can be more efficient to manipulate flags and release the
914 : * locks in a single atomic operation.
915 : */
916 : static void
917 7397453 : LWLockWaitListUnlock(LWLock *lock)
918 : {
919 : uint32 old_state PG_USED_FOR_ASSERTS_ONLY;
920 :
921 7397453 : old_state = pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_LOCKED);
922 :
923 : Assert(old_state & LW_FLAG_LOCKED);
924 7397453 : }
925 :
926 : /*
927 : * Wakeup all the lockers that currently have a chance to acquire the lock.
928 : */
929 : static void
930 4358688 : LWLockWakeup(LWLock *lock)
931 : {
932 4358688 : bool new_wake_in_progress = false;
933 4358688 : bool wokeup_somebody = false;
934 : proclist_head wakeup;
935 : proclist_mutable_iter iter;
936 :
937 4358688 : proclist_init(&wakeup);
938 :
939 : /* lock wait list while collecting backends to wake up */
940 4358688 : LWLockWaitListLock(lock);
941 :
942 8099886 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
943 : {
944 4681903 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
945 :
946 4681903 : if (wokeup_somebody && waiter->lwWaitMode == LW_EXCLUSIVE)
947 18769 : continue;
948 :
949 4663134 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
950 4663134 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
951 :
952 4663134 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
953 : {
954 : /*
955 : * Prevent additional wakeups until retryer gets to run. Backends
956 : * that are just waiting for the lock to become free don't retry
957 : * automatically.
958 : */
959 4556948 : new_wake_in_progress = true;
960 :
961 : /*
962 : * Don't wakeup (further) exclusive locks.
963 : */
964 4556948 : wokeup_somebody = true;
965 : }
966 :
967 : /*
968 : * Signal that the process isn't on the wait list anymore. This allows
969 : * LWLockDequeueSelf() to remove itself of the waitlist with a
970 : * proclist_delete(), rather than having to check if it has been
971 : * removed from the list.
972 : */
973 : Assert(waiter->lwWaiting == LW_WS_WAITING);
974 4663134 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
975 :
976 : /*
977 : * Once we've woken up an exclusive lock, there's no point in waking
978 : * up anybody else.
979 : */
980 4663134 : if (waiter->lwWaitMode == LW_EXCLUSIVE)
981 940705 : break;
982 : }
983 :
984 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS);
985 :
986 : /* unset required flags, and release lock, in one fell swoop */
987 : {
988 : uint32 old_state;
989 : uint32 desired_state;
990 :
991 4358688 : old_state = pg_atomic_read_u32(&lock->state);
992 : while (true)
993 : {
994 4365210 : desired_state = old_state;
995 :
996 : /* compute desired flags */
997 :
998 4365210 : if (new_wake_in_progress)
999 4301952 : desired_state |= LW_FLAG_WAKE_IN_PROGRESS;
1000 : else
1001 63258 : desired_state &= ~LW_FLAG_WAKE_IN_PROGRESS;
1002 :
1003 4365210 : if (proclist_is_empty(&lock->waiters))
1004 4300965 : desired_state &= ~LW_FLAG_HAS_WAITERS;
1005 :
1006 4365210 : desired_state &= ~LW_FLAG_LOCKED; /* release lock */
1007 :
1008 4365210 : if (pg_atomic_compare_exchange_u32(&lock->state, &old_state,
1009 : desired_state))
1010 4358688 : break;
1011 : }
1012 : }
1013 :
1014 : /* Awaken any waiters I removed from the queue. */
1015 9021822 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1016 : {
1017 4663134 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1018 :
1019 : LOG_LWDEBUG("LWLockRelease", lock, "release waiter");
1020 4663134 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1021 :
1022 : /*
1023 : * Guarantee that lwWaiting being unset only becomes visible once the
1024 : * unlink from the link has completed. Otherwise the target backend
1025 : * could be woken up for other reason and enqueue for a new lock - if
1026 : * that happens before the list unlink happens, the list would end up
1027 : * being corrupted.
1028 : *
1029 : * The barrier pairs with the LWLockWaitListLock() when enqueuing for
1030 : * another lock.
1031 : */
1032 4663134 : pg_write_barrier();
1033 4663134 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1034 4663134 : PGSemaphoreUnlock(waiter->sem);
1035 : }
1036 4358688 : }
1037 :
1038 : /*
1039 : * Add ourselves to the end of the queue.
1040 : *
1041 : * NB: Mode can be LW_WAIT_UNTIL_FREE here!
1042 : */
1043 : static void
1044 4719427 : LWLockQueueSelf(LWLock *lock, LWLockMode mode)
1045 : {
1046 : /*
1047 : * If we don't have a PGPROC structure, there's no way to wait. This
1048 : * should never occur, since MyProc should only be null during shared
1049 : * memory initialization.
1050 : */
1051 4719427 : if (MyProc == NULL)
1052 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1053 :
1054 4719427 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
1055 0 : elog(PANIC, "queueing for lock while waiting on another one");
1056 :
1057 4719427 : LWLockWaitListLock(lock);
1058 :
1059 : /* setting the flag is protected by the spinlock */
1060 4719427 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_HAS_WAITERS);
1061 :
1062 4719427 : MyProc->lwWaiting = LW_WS_WAITING;
1063 4719427 : MyProc->lwWaitMode = mode;
1064 :
1065 : /* LW_WAIT_UNTIL_FREE waiters are always at the front of the queue */
1066 4719427 : if (mode == LW_WAIT_UNTIL_FREE)
1067 108151 : proclist_push_head(&lock->waiters, MyProcNumber, lwWaitLink);
1068 : else
1069 4611276 : proclist_push_tail(&lock->waiters, MyProcNumber, lwWaitLink);
1070 :
1071 : /* Can release the mutex now */
1072 4719427 : LWLockWaitListUnlock(lock);
1073 :
1074 : #ifdef LOCK_DEBUG
1075 : pg_atomic_fetch_add_u32(&lock->nwaiters, 1);
1076 : #endif
1077 4719427 : }
1078 :
1079 : /*
1080 : * Remove ourselves from the waitlist.
1081 : *
1082 : * This is used if we queued ourselves because we thought we needed to sleep
1083 : * but, after further checking, we discovered that we don't actually need to
1084 : * do so.
1085 : */
1086 : static void
1087 79995 : LWLockDequeueSelf(LWLock *lock)
1088 : {
1089 : bool on_waitlist;
1090 :
1091 : #ifdef LWLOCK_STATS
1092 : lwlock_stats *lwstats;
1093 :
1094 : lwstats = get_lwlock_stats_entry(lock);
1095 :
1096 : lwstats->dequeue_self_count++;
1097 : #endif
1098 :
1099 79995 : LWLockWaitListLock(lock);
1100 :
1101 : /*
1102 : * Remove ourselves from the waitlist, unless we've already been removed.
1103 : * The removal happens with the wait list lock held, so there's no race in
1104 : * this check.
1105 : */
1106 79995 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
1107 79995 : if (on_waitlist)
1108 56358 : proclist_delete(&lock->waiters, MyProcNumber, lwWaitLink);
1109 :
1110 79995 : if (proclist_is_empty(&lock->waiters) &&
1111 74621 : (pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS) != 0)
1112 : {
1113 54492 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_HAS_WAITERS);
1114 : }
1115 :
1116 : /* XXX: combine with fetch_and above? */
1117 79995 : LWLockWaitListUnlock(lock);
1118 :
1119 : /* clear waiting state again, nice for debugging */
1120 79995 : if (on_waitlist)
1121 56358 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
1122 : else
1123 : {
1124 23637 : int extraWaits = 0;
1125 :
1126 : /*
1127 : * Somebody else dequeued us and has or will wake us up. Deal with the
1128 : * superfluous absorption of a wakeup.
1129 : */
1130 :
1131 : /*
1132 : * Clear LW_FLAG_WAKE_IN_PROGRESS if somebody woke us before we
1133 : * removed ourselves - they'll have set it.
1134 : */
1135 23637 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_WAKE_IN_PROGRESS);
1136 :
1137 : /*
1138 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
1139 : * get reset at some inconvenient point later. Most of the time this
1140 : * will immediately return.
1141 : */
1142 : for (;;)
1143 : {
1144 23637 : PGSemaphoreLock(MyProc->sem);
1145 23637 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
1146 23637 : break;
1147 0 : extraWaits++;
1148 : }
1149 :
1150 : /*
1151 : * Fix the process wait semaphore's count for any absorbed wakeups.
1152 : */
1153 23637 : while (extraWaits-- > 0)
1154 0 : PGSemaphoreUnlock(MyProc->sem);
1155 : }
1156 :
1157 : #ifdef LOCK_DEBUG
1158 : {
1159 : /* not waiting anymore */
1160 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1161 :
1162 : Assert(nwaiters < MAX_BACKENDS);
1163 : }
1164 : #endif
1165 79995 : }
1166 :
1167 : /*
1168 : * LWLockAcquire - acquire a lightweight lock in the specified mode
1169 : *
1170 : * If the lock is not available, sleep until it is. Returns true if the lock
1171 : * was available immediately, false if we had to sleep.
1172 : *
1173 : * Side effect: cancel/die interrupts are held off until lock release.
1174 : */
1175 : bool
1176 300125907 : LWLockAcquire(LWLock *lock, LWLockMode mode)
1177 : {
1178 300125907 : PGPROC *proc = MyProc;
1179 300125907 : bool result = true;
1180 300125907 : int extraWaits = 0;
1181 : #ifdef LWLOCK_STATS
1182 : lwlock_stats *lwstats;
1183 :
1184 : lwstats = get_lwlock_stats_entry(lock);
1185 : #endif
1186 :
1187 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1188 :
1189 : PRINT_LWDEBUG("LWLockAcquire", lock, mode);
1190 :
1191 : #ifdef LWLOCK_STATS
1192 : /* Count lock acquisition attempts */
1193 : if (mode == LW_EXCLUSIVE)
1194 : lwstats->ex_acquire_count++;
1195 : else
1196 : lwstats->sh_acquire_count++;
1197 : #endif /* LWLOCK_STATS */
1198 :
1199 : /*
1200 : * We can't wait if we haven't got a PGPROC. This should only occur
1201 : * during bootstrap or shared memory initialization. Put an Assert here
1202 : * to catch unsafe coding practices.
1203 : */
1204 : Assert(!(proc == NULL && IsUnderPostmaster));
1205 :
1206 : /* Ensure we will have room to remember the lock */
1207 300125907 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1208 0 : elog(ERROR, "too many LWLocks taken");
1209 :
1210 : /*
1211 : * Lock out cancel/die interrupts until we exit the code section protected
1212 : * by the LWLock. This ensures that interrupts will not interfere with
1213 : * manipulations of data structures in shared memory.
1214 : */
1215 300125907 : HOLD_INTERRUPTS();
1216 :
1217 : /*
1218 : * Loop here to try to acquire lock after each time we are signaled by
1219 : * LWLockRelease.
1220 : *
1221 : * NOTE: it might seem better to have LWLockRelease actually grant us the
1222 : * lock, rather than retrying and possibly having to go back to sleep. But
1223 : * in practice that is no good because it means a process swap for every
1224 : * lock acquisition when two or more processes are contending for the same
1225 : * lock. Since LWLocks are normally used to protect not-very-long
1226 : * sections of computation, a process needs to be able to acquire and
1227 : * release the same lock many times during a single CPU time slice, even
1228 : * in the presence of contention. The efficiency of being able to do that
1229 : * outweighs the inefficiency of sometimes wasting a process dispatch
1230 : * cycle because the lock is not free when a released waiter finally gets
1231 : * to run. See pgsql-hackers archives for 29-Dec-01.
1232 : */
1233 : for (;;)
1234 4532815 : {
1235 : bool mustwait;
1236 :
1237 : /*
1238 : * Try to grab the lock the first time, we're not in the waitqueue
1239 : * yet/anymore.
1240 : */
1241 304658722 : mustwait = LWLockAttemptLock(lock, mode);
1242 :
1243 304658722 : if (!mustwait)
1244 : {
1245 : LOG_LWDEBUG("LWLockAcquire", lock, "immediately acquired lock");
1246 300047446 : break; /* got the lock */
1247 : }
1248 :
1249 : /*
1250 : * Ok, at this point we couldn't grab the lock on the first try. We
1251 : * cannot simply queue ourselves to the end of the list and wait to be
1252 : * woken up because by now the lock could long have been released.
1253 : * Instead add us to the queue and try to grab the lock again. If we
1254 : * succeed we need to revert the queuing and be happy, otherwise we
1255 : * recheck the lock. If we still couldn't grab it, we know that the
1256 : * other locker will see our queue entries when releasing since they
1257 : * existed before we checked for the lock.
1258 : */
1259 :
1260 : /* add to the queue */
1261 4611276 : LWLockQueueSelf(lock, mode);
1262 :
1263 : /* we're now guaranteed to be woken up if necessary */
1264 4611276 : mustwait = LWLockAttemptLock(lock, mode);
1265 :
1266 : /* ok, grabbed the lock the second time round, need to undo queueing */
1267 4611276 : if (!mustwait)
1268 : {
1269 : LOG_LWDEBUG("LWLockAcquire", lock, "acquired, undoing queue");
1270 :
1271 78461 : LWLockDequeueSelf(lock);
1272 78461 : break;
1273 : }
1274 :
1275 : /*
1276 : * Wait until awakened.
1277 : *
1278 : * It is possible that we get awakened for a reason other than being
1279 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1280 : * we've gotten the LWLock, re-increment the sema by the number of
1281 : * additional signals received.
1282 : */
1283 : LOG_LWDEBUG("LWLockAcquire", lock, "waiting");
1284 :
1285 : #ifdef LWLOCK_STATS
1286 : lwstats->block_count++;
1287 : #endif
1288 :
1289 4532815 : LWLockReportWaitStart(lock);
1290 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1291 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1292 :
1293 : for (;;)
1294 : {
1295 4532815 : PGSemaphoreLock(proc->sem);
1296 4532815 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1297 4532815 : break;
1298 0 : extraWaits++;
1299 : }
1300 :
1301 : /* Retrying, allow LWLockRelease to release waiters again. */
1302 4532815 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_WAKE_IN_PROGRESS);
1303 :
1304 : #ifdef LOCK_DEBUG
1305 : {
1306 : /* not waiting anymore */
1307 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1308 :
1309 : Assert(nwaiters < MAX_BACKENDS);
1310 : }
1311 : #endif
1312 :
1313 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1314 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1315 4532815 : LWLockReportWaitEnd();
1316 :
1317 : LOG_LWDEBUG("LWLockAcquire", lock, "awakened");
1318 :
1319 : /* Now loop back and try to acquire lock again. */
1320 4532815 : result = false;
1321 : }
1322 :
1323 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_ENABLED())
1324 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE(T_NAME(lock), mode);
1325 :
1326 : /* Add lock to list of locks held by this backend */
1327 300125907 : held_lwlocks[num_held_lwlocks].lock = lock;
1328 300125907 : held_lwlocks[num_held_lwlocks++].mode = mode;
1329 :
1330 : /*
1331 : * Fix the process wait semaphore's count for any absorbed wakeups.
1332 : */
1333 300125907 : while (extraWaits-- > 0)
1334 0 : PGSemaphoreUnlock(proc->sem);
1335 :
1336 300125907 : return result;
1337 : }
1338 :
1339 : /*
1340 : * LWLockConditionalAcquire - acquire a lightweight lock in the specified mode
1341 : *
1342 : * If the lock is not available, return false with no side-effects.
1343 : *
1344 : * If successful, cancel/die interrupts are held off until lock release.
1345 : */
1346 : bool
1347 676074 : LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
1348 : {
1349 : bool mustwait;
1350 :
1351 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1352 :
1353 : PRINT_LWDEBUG("LWLockConditionalAcquire", lock, mode);
1354 :
1355 : /* Ensure we will have room to remember the lock */
1356 676074 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1357 0 : elog(ERROR, "too many LWLocks taken");
1358 :
1359 : /*
1360 : * Lock out cancel/die interrupts until we exit the code section protected
1361 : * by the LWLock. This ensures that interrupts will not interfere with
1362 : * manipulations of data structures in shared memory.
1363 : */
1364 676074 : HOLD_INTERRUPTS();
1365 :
1366 : /* Check for the lock */
1367 676074 : mustwait = LWLockAttemptLock(lock, mode);
1368 :
1369 676074 : if (mustwait)
1370 : {
1371 : /* Failed to get lock, so release interrupt holdoff */
1372 189 : RESUME_INTERRUPTS();
1373 :
1374 : LOG_LWDEBUG("LWLockConditionalAcquire", lock, "failed");
1375 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL_ENABLED())
1376 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL(T_NAME(lock), mode);
1377 : }
1378 : else
1379 : {
1380 : /* Add lock to list of locks held by this backend */
1381 675885 : held_lwlocks[num_held_lwlocks].lock = lock;
1382 675885 : held_lwlocks[num_held_lwlocks++].mode = mode;
1383 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_ENABLED())
1384 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE(T_NAME(lock), mode);
1385 : }
1386 676074 : return !mustwait;
1387 : }
1388 :
1389 : /*
1390 : * LWLockAcquireOrWait - Acquire lock, or wait until it's free
1391 : *
1392 : * The semantics of this function are a bit funky. If the lock is currently
1393 : * free, it is acquired in the given mode, and the function returns true. If
1394 : * the lock isn't immediately free, the function waits until it is released
1395 : * and returns false, but does not acquire the lock.
1396 : *
1397 : * This is currently used for WALWriteLock: when a backend flushes the WAL,
1398 : * holding WALWriteLock, it can flush the commit records of many other
1399 : * backends as a side-effect. Those other backends need to wait until the
1400 : * flush finishes, but don't need to acquire the lock anymore. They can just
1401 : * wake up, observe that their records have already been flushed, and return.
1402 : */
1403 : bool
1404 138320 : LWLockAcquireOrWait(LWLock *lock, LWLockMode mode)
1405 : {
1406 138320 : PGPROC *proc = MyProc;
1407 : bool mustwait;
1408 138320 : int extraWaits = 0;
1409 : #ifdef LWLOCK_STATS
1410 : lwlock_stats *lwstats;
1411 :
1412 : lwstats = get_lwlock_stats_entry(lock);
1413 : #endif
1414 :
1415 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1416 :
1417 : PRINT_LWDEBUG("LWLockAcquireOrWait", lock, mode);
1418 :
1419 : /* Ensure we will have room to remember the lock */
1420 138320 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1421 0 : elog(ERROR, "too many LWLocks taken");
1422 :
1423 : /*
1424 : * Lock out cancel/die interrupts until we exit the code section protected
1425 : * by the LWLock. This ensures that interrupts will not interfere with
1426 : * manipulations of data structures in shared memory.
1427 : */
1428 138320 : HOLD_INTERRUPTS();
1429 :
1430 : /*
1431 : * NB: We're using nearly the same twice-in-a-row lock acquisition
1432 : * protocol as LWLockAcquire(). Check its comments for details.
1433 : */
1434 138320 : mustwait = LWLockAttemptLock(lock, mode);
1435 :
1436 138320 : if (mustwait)
1437 : {
1438 3257 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1439 :
1440 3257 : mustwait = LWLockAttemptLock(lock, mode);
1441 :
1442 3257 : if (mustwait)
1443 : {
1444 : /*
1445 : * Wait until awakened. Like in LWLockAcquire, be prepared for
1446 : * bogus wakeups.
1447 : */
1448 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "waiting");
1449 :
1450 : #ifdef LWLOCK_STATS
1451 : lwstats->block_count++;
1452 : #endif
1453 :
1454 3159 : LWLockReportWaitStart(lock);
1455 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1456 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1457 :
1458 : for (;;)
1459 : {
1460 3159 : PGSemaphoreLock(proc->sem);
1461 3159 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1462 3159 : break;
1463 0 : extraWaits++;
1464 : }
1465 :
1466 : #ifdef LOCK_DEBUG
1467 : {
1468 : /* not waiting anymore */
1469 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1470 :
1471 : Assert(nwaiters < MAX_BACKENDS);
1472 : }
1473 : #endif
1474 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1475 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1476 3159 : LWLockReportWaitEnd();
1477 :
1478 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "awakened");
1479 : }
1480 : else
1481 : {
1482 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "acquired, undoing queue");
1483 :
1484 : /*
1485 : * Got lock in the second attempt, undo queueing. We need to treat
1486 : * this as having successfully acquired the lock, otherwise we'd
1487 : * not necessarily wake up people we've prevented from acquiring
1488 : * the lock.
1489 : */
1490 98 : LWLockDequeueSelf(lock);
1491 : }
1492 : }
1493 :
1494 : /*
1495 : * Fix the process wait semaphore's count for any absorbed wakeups.
1496 : */
1497 138320 : while (extraWaits-- > 0)
1498 0 : PGSemaphoreUnlock(proc->sem);
1499 :
1500 138320 : if (mustwait)
1501 : {
1502 : /* Failed to get lock, so release interrupt holdoff */
1503 3159 : RESUME_INTERRUPTS();
1504 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "failed");
1505 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL_ENABLED())
1506 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL(T_NAME(lock), mode);
1507 : }
1508 : else
1509 : {
1510 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "succeeded");
1511 : /* Add lock to list of locks held by this backend */
1512 135161 : held_lwlocks[num_held_lwlocks].lock = lock;
1513 135161 : held_lwlocks[num_held_lwlocks++].mode = mode;
1514 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_ENABLED())
1515 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT(T_NAME(lock), mode);
1516 : }
1517 :
1518 138320 : return !mustwait;
1519 : }
1520 :
1521 : /*
1522 : * Does the lwlock in its current state need to wait for the variable value to
1523 : * change?
1524 : *
1525 : * If we don't need to wait, and it's because the value of the variable has
1526 : * changed, store the current value in newval.
1527 : *
1528 : * *result is set to true if the lock was free, and false otherwise.
1529 : */
1530 : static bool
1531 4045883 : LWLockConflictsWithVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1532 : uint64 *newval, bool *result)
1533 : {
1534 : bool mustwait;
1535 : uint64 value;
1536 :
1537 : /*
1538 : * Test first to see if it the slot is free right now.
1539 : *
1540 : * XXX: the unique caller of this routine, WaitXLogInsertionsToFinish()
1541 : * via LWLockWaitForVar(), uses an implied barrier with a spinlock before
1542 : * this, so we don't need a memory barrier here as far as the current
1543 : * usage is concerned. But that might not be safe in general.
1544 : */
1545 4045883 : mustwait = (pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE) != 0;
1546 :
1547 4045883 : if (!mustwait)
1548 : {
1549 2623953 : *result = true;
1550 2623953 : return false;
1551 : }
1552 :
1553 1421930 : *result = false;
1554 :
1555 : /*
1556 : * Reading this value atomically is safe even on platforms where uint64
1557 : * cannot be read without observing a torn value.
1558 : */
1559 1421930 : value = pg_atomic_read_u64(valptr);
1560 :
1561 1421930 : if (value != oldval)
1562 : {
1563 1213578 : mustwait = false;
1564 1213578 : *newval = value;
1565 : }
1566 : else
1567 : {
1568 208352 : mustwait = true;
1569 : }
1570 :
1571 1421930 : return mustwait;
1572 : }
1573 :
1574 : /*
1575 : * LWLockWaitForVar - Wait until lock is free, or a variable is updated.
1576 : *
1577 : * If the lock is held and *valptr equals oldval, waits until the lock is
1578 : * either freed, or the lock holder updates *valptr by calling
1579 : * LWLockUpdateVar. If the lock is free on exit (immediately or after
1580 : * waiting), returns true. If the lock is still held, but *valptr no longer
1581 : * matches oldval, returns false and sets *newval to the current value in
1582 : * *valptr.
1583 : *
1584 : * Note: this function ignores shared lock holders; if the lock is held
1585 : * in shared mode, returns 'true'.
1586 : *
1587 : * Be aware that LWLockConflictsWithVar() does not include a memory barrier,
1588 : * hence the caller of this function may want to rely on an explicit barrier or
1589 : * an implied barrier via spinlock or LWLock to avoid memory ordering issues.
1590 : */
1591 : bool
1592 3837531 : LWLockWaitForVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1593 : uint64 *newval)
1594 : {
1595 3837531 : PGPROC *proc = MyProc;
1596 3837531 : int extraWaits = 0;
1597 3837531 : bool result = false;
1598 : #ifdef LWLOCK_STATS
1599 : lwlock_stats *lwstats;
1600 :
1601 : lwstats = get_lwlock_stats_entry(lock);
1602 : #endif
1603 :
1604 : PRINT_LWDEBUG("LWLockWaitForVar", lock, LW_WAIT_UNTIL_FREE);
1605 :
1606 : /*
1607 : * Lock out cancel/die interrupts while we sleep on the lock. There is no
1608 : * cleanup mechanism to remove us from the wait queue if we got
1609 : * interrupted.
1610 : */
1611 3837531 : HOLD_INTERRUPTS();
1612 :
1613 : /*
1614 : * Loop here to check the lock's status after each time we are signaled.
1615 : */
1616 : for (;;)
1617 103458 : {
1618 : bool mustwait;
1619 :
1620 3940989 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1621 : &result);
1622 :
1623 3940989 : if (!mustwait)
1624 3836095 : break; /* the lock was free or value didn't match */
1625 :
1626 : /*
1627 : * Add myself to wait queue. Note that this is racy, somebody else
1628 : * could wakeup before we're finished queuing. NB: We're using nearly
1629 : * the same twice-in-a-row lock acquisition protocol as
1630 : * LWLockAcquire(). Check its comments for details. The only
1631 : * difference is that we also have to check the variable's values when
1632 : * checking the state of the lock.
1633 : */
1634 104894 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1635 :
1636 : /*
1637 : * Clear LW_FLAG_WAKE_IN_PROGRESS flag, to make sure we get woken up
1638 : * as soon as the lock is released.
1639 : */
1640 104894 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_WAKE_IN_PROGRESS);
1641 :
1642 : /*
1643 : * We're now guaranteed to be woken up if necessary. Recheck the lock
1644 : * and variables state.
1645 : */
1646 104894 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1647 : &result);
1648 :
1649 : /* Ok, no conflict after we queued ourselves. Undo queueing. */
1650 104894 : if (!mustwait)
1651 : {
1652 : LOG_LWDEBUG("LWLockWaitForVar", lock, "free, undoing queue");
1653 :
1654 1436 : LWLockDequeueSelf(lock);
1655 1436 : break;
1656 : }
1657 :
1658 : /*
1659 : * Wait until awakened.
1660 : *
1661 : * It is possible that we get awakened for a reason other than being
1662 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1663 : * we've gotten the LWLock, re-increment the sema by the number of
1664 : * additional signals received.
1665 : */
1666 : LOG_LWDEBUG("LWLockWaitForVar", lock, "waiting");
1667 :
1668 : #ifdef LWLOCK_STATS
1669 : lwstats->block_count++;
1670 : #endif
1671 :
1672 103458 : LWLockReportWaitStart(lock);
1673 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1674 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), LW_EXCLUSIVE);
1675 :
1676 : for (;;)
1677 : {
1678 103458 : PGSemaphoreLock(proc->sem);
1679 103458 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1680 103458 : break;
1681 0 : extraWaits++;
1682 : }
1683 :
1684 : #ifdef LOCK_DEBUG
1685 : {
1686 : /* not waiting anymore */
1687 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1688 :
1689 : Assert(nwaiters < MAX_BACKENDS);
1690 : }
1691 : #endif
1692 :
1693 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1694 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), LW_EXCLUSIVE);
1695 103458 : LWLockReportWaitEnd();
1696 :
1697 : LOG_LWDEBUG("LWLockWaitForVar", lock, "awakened");
1698 :
1699 : /* Now loop back and check the status of the lock again. */
1700 : }
1701 :
1702 : /*
1703 : * Fix the process wait semaphore's count for any absorbed wakeups.
1704 : */
1705 3837531 : while (extraWaits-- > 0)
1706 0 : PGSemaphoreUnlock(proc->sem);
1707 :
1708 : /*
1709 : * Now okay to allow cancel/die interrupts.
1710 : */
1711 3837531 : RESUME_INTERRUPTS();
1712 :
1713 3837531 : return result;
1714 : }
1715 :
1716 :
1717 : /*
1718 : * LWLockUpdateVar - Update a variable and wake up waiters atomically
1719 : *
1720 : * Sets *valptr to 'val', and wakes up all processes waiting for us with
1721 : * LWLockWaitForVar(). It first sets the value atomically and then wakes up
1722 : * waiting processes so that any process calling LWLockWaitForVar() on the same
1723 : * lock is guaranteed to see the new value, and act accordingly.
1724 : *
1725 : * The caller must be holding the lock in exclusive mode.
1726 : */
1727 : void
1728 2598031 : LWLockUpdateVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1729 : {
1730 : proclist_head wakeup;
1731 : proclist_mutable_iter iter;
1732 :
1733 : PRINT_LWDEBUG("LWLockUpdateVar", lock, LW_EXCLUSIVE);
1734 :
1735 : /*
1736 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1737 : * that the variable is updated before waking up waiters.
1738 : */
1739 2598031 : pg_atomic_exchange_u64(valptr, val);
1740 :
1741 2598031 : proclist_init(&wakeup);
1742 :
1743 2598031 : LWLockWaitListLock(lock);
1744 :
1745 : Assert(pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE);
1746 :
1747 : /*
1748 : * See if there are any LW_WAIT_UNTIL_FREE waiters that need to be woken
1749 : * up. They are always in the front of the queue.
1750 : */
1751 2598877 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
1752 : {
1753 38878 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1754 :
1755 38878 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
1756 38032 : break;
1757 :
1758 846 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
1759 846 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
1760 :
1761 : /* see LWLockWakeup() */
1762 : Assert(waiter->lwWaiting == LW_WS_WAITING);
1763 846 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
1764 : }
1765 :
1766 : /* We are done updating shared state of the lock itself. */
1767 2598031 : LWLockWaitListUnlock(lock);
1768 :
1769 : /*
1770 : * Awaken any waiters I removed from the queue.
1771 : */
1772 2598877 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1773 : {
1774 846 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1775 :
1776 846 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1777 : /* check comment in LWLockWakeup() about this barrier */
1778 846 : pg_write_barrier();
1779 846 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1780 846 : PGSemaphoreUnlock(waiter->sem);
1781 : }
1782 2598031 : }
1783 :
1784 :
1785 : /*
1786 : * LWLockRelease - release a previously acquired lock
1787 : *
1788 : * NB: This will leave lock->owner pointing to the current backend (if
1789 : * LOCK_DEBUG is set). This is somewhat intentional, as it makes it easier to
1790 : * debug cases of missing wakeups during lock release.
1791 : */
1792 : void
1793 300936953 : LWLockRelease(LWLock *lock)
1794 : {
1795 : LWLockMode mode;
1796 : uint32 oldstate;
1797 : bool check_waiters;
1798 : int i;
1799 :
1800 : /*
1801 : * Remove lock from list of locks held. Usually, but not always, it will
1802 : * be the latest-acquired lock; so search array backwards.
1803 : */
1804 343097039 : for (i = num_held_lwlocks; --i >= 0;)
1805 343097039 : if (lock == held_lwlocks[i].lock)
1806 300936953 : break;
1807 :
1808 300936953 : if (i < 0)
1809 0 : elog(ERROR, "lock %s is not held", T_NAME(lock));
1810 :
1811 300936953 : mode = held_lwlocks[i].mode;
1812 :
1813 300936953 : num_held_lwlocks--;
1814 343097039 : for (; i < num_held_lwlocks; i++)
1815 42160086 : held_lwlocks[i] = held_lwlocks[i + 1];
1816 :
1817 : PRINT_LWDEBUG("LWLockRelease", lock, mode);
1818 :
1819 : /*
1820 : * Release my hold on lock, after that it can immediately be acquired by
1821 : * others, even if we still have to wakeup other waiters.
1822 : */
1823 300936953 : if (mode == LW_EXCLUSIVE)
1824 213194519 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_EXCLUSIVE);
1825 : else
1826 87742434 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_SHARED);
1827 :
1828 : /* nobody else can have that kind of lock */
1829 : Assert(!(oldstate & LW_VAL_EXCLUSIVE));
1830 :
1831 : if (TRACE_POSTGRESQL_LWLOCK_RELEASE_ENABLED())
1832 : TRACE_POSTGRESQL_LWLOCK_RELEASE(T_NAME(lock));
1833 :
1834 : /*
1835 : * Check if we're still waiting for backends to get scheduled, if so,
1836 : * don't wake them up again.
1837 : */
1838 300936953 : if ((oldstate & LW_FLAG_HAS_WAITERS) &&
1839 5505979 : !(oldstate & LW_FLAG_WAKE_IN_PROGRESS) &&
1840 4361150 : (oldstate & LW_LOCK_MASK) == 0)
1841 4358688 : check_waiters = true;
1842 : else
1843 296578265 : check_waiters = false;
1844 :
1845 : /*
1846 : * As waking up waiters requires the spinlock to be acquired, only do so
1847 : * if necessary.
1848 : */
1849 300936953 : if (check_waiters)
1850 : {
1851 : /* XXX: remove before commit? */
1852 : LOG_LWDEBUG("LWLockRelease", lock, "releasing waiters");
1853 4358688 : LWLockWakeup(lock);
1854 : }
1855 :
1856 : /*
1857 : * Now okay to allow cancel/die interrupts.
1858 : */
1859 300936953 : RESUME_INTERRUPTS();
1860 300936953 : }
1861 :
1862 : /*
1863 : * LWLockReleaseClearVar - release a previously acquired lock, reset variable
1864 : */
1865 : void
1866 15553493 : LWLockReleaseClearVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1867 : {
1868 : /*
1869 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1870 : * that the variable is updated before releasing the lock.
1871 : */
1872 15553493 : pg_atomic_exchange_u64(valptr, val);
1873 :
1874 15553493 : LWLockRelease(lock);
1875 15553493 : }
1876 :
1877 :
1878 : /*
1879 : * LWLockReleaseAll - release all currently-held locks
1880 : *
1881 : * Used to clean up after ereport(ERROR). An important difference between this
1882 : * function and retail LWLockRelease calls is that InterruptHoldoffCount is
1883 : * unchanged by this operation. This is necessary since InterruptHoldoffCount
1884 : * has been set to an appropriate level earlier in error recovery. We could
1885 : * decrement it below zero if we allow it to drop for each released lock!
1886 : *
1887 : * Note that this function must be safe to call even before the LWLock
1888 : * subsystem has been initialized (e.g., during early startup failures).
1889 : * In that case, num_held_lwlocks will be 0 and we do nothing.
1890 : */
1891 : void
1892 108515 : LWLockReleaseAll(void)
1893 : {
1894 108607 : while (num_held_lwlocks > 0)
1895 : {
1896 92 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
1897 :
1898 92 : LWLockRelease(held_lwlocks[num_held_lwlocks - 1].lock);
1899 : }
1900 :
1901 : Assert(num_held_lwlocks == 0);
1902 108515 : }
1903 :
1904 :
1905 : /*
1906 : * LWLockHeldByMe - test whether my process holds a lock in any mode
1907 : *
1908 : * This is meant as debug support only.
1909 : */
1910 : bool
1911 0 : LWLockHeldByMe(LWLock *lock)
1912 : {
1913 : int i;
1914 :
1915 0 : for (i = 0; i < num_held_lwlocks; i++)
1916 : {
1917 0 : if (held_lwlocks[i].lock == lock)
1918 0 : return true;
1919 : }
1920 0 : return false;
1921 : }
1922 :
1923 : /*
1924 : * LWLockAnyHeldByMe - test whether my process holds any of an array of locks
1925 : *
1926 : * This is meant as debug support only.
1927 : */
1928 : bool
1929 0 : LWLockAnyHeldByMe(LWLock *lock, int nlocks, size_t stride)
1930 : {
1931 : char *held_lock_addr;
1932 : char *begin;
1933 : char *end;
1934 : int i;
1935 :
1936 0 : begin = (char *) lock;
1937 0 : end = begin + nlocks * stride;
1938 0 : for (i = 0; i < num_held_lwlocks; i++)
1939 : {
1940 0 : held_lock_addr = (char *) held_lwlocks[i].lock;
1941 0 : if (held_lock_addr >= begin &&
1942 0 : held_lock_addr < end &&
1943 0 : (held_lock_addr - begin) % stride == 0)
1944 0 : return true;
1945 : }
1946 0 : return false;
1947 : }
1948 :
1949 : /*
1950 : * LWLockHeldByMeInMode - test whether my process holds a lock in given mode
1951 : *
1952 : * This is meant as debug support only.
1953 : */
1954 : bool
1955 0 : LWLockHeldByMeInMode(LWLock *lock, LWLockMode mode)
1956 : {
1957 : int i;
1958 :
1959 0 : for (i = 0; i < num_held_lwlocks; i++)
1960 : {
1961 0 : if (held_lwlocks[i].lock == lock && held_lwlocks[i].mode == mode)
1962 0 : return true;
1963 : }
1964 0 : return false;
1965 : }
|
