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 : #include "utils/wait_event.h"
89 :
90 : #ifdef LWLOCK_STATS
91 : #include "utils/hsearch.h"
92 : #endif
93 :
94 :
95 : #define LW_FLAG_HAS_WAITERS ((uint32) 1 << 31)
96 : #define LW_FLAG_WAKE_IN_PROGRESS ((uint32) 1 << 30)
97 : #define LW_FLAG_LOCKED ((uint32) 1 << 29)
98 : #define LW_FLAG_BITS 3
99 : #define LW_FLAG_MASK (((1<<LW_FLAG_BITS)-1)<<(32-LW_FLAG_BITS))
100 :
101 : /* assumes MAX_BACKENDS is a (power of 2) - 1, checked below */
102 : #define LW_VAL_EXCLUSIVE (MAX_BACKENDS + 1)
103 : #define LW_VAL_SHARED 1
104 :
105 : /* already (power of 2)-1, i.e. suitable for a mask */
106 : #define LW_SHARED_MASK MAX_BACKENDS
107 : #define LW_LOCK_MASK (MAX_BACKENDS | LW_VAL_EXCLUSIVE)
108 :
109 :
110 : StaticAssertDecl(((MAX_BACKENDS + 1) & MAX_BACKENDS) == 0,
111 : "MAX_BACKENDS + 1 needs to be a power of 2");
112 :
113 : StaticAssertDecl((MAX_BACKENDS & LW_FLAG_MASK) == 0,
114 : "MAX_BACKENDS and LW_FLAG_MASK overlap");
115 :
116 : StaticAssertDecl((LW_VAL_EXCLUSIVE & LW_FLAG_MASK) == 0,
117 : "LW_VAL_EXCLUSIVE and LW_FLAG_MASK overlap");
118 :
119 : /*
120 : * There are three sorts of LWLock "tranches":
121 : *
122 : * 1. The individually-named locks defined in lwlocklist.h each have their
123 : * own tranche. We absorb the names of these tranches from there into
124 : * BuiltinTrancheNames here.
125 : *
126 : * 2. There are some predefined tranches for built-in groups of locks defined
127 : * in lwlocklist.h. We absorb the names of these tranches, too.
128 : *
129 : * 3. Extensions can create new tranches, via either RequestNamedLWLockTranche
130 : * or LWLockNewTrancheId. These names are stored in shared memory and can be
131 : * accessed via LWLockTrancheNames.
132 : *
133 : * All these names are user-visible as wait event names, so choose with care
134 : * ... and do not forget to update the documentation's list of wait events.
135 : */
136 : static const char *const BuiltinTrancheNames[] = {
137 : #define PG_LWLOCK(id, lockname) [id] = CppAsString(lockname),
138 : #define PG_LWLOCKTRANCHE(id, lockname) [LWTRANCHE_##id] = CppAsString(lockname),
139 : #include "storage/lwlocklist.h"
140 : #undef PG_LWLOCK
141 : #undef PG_LWLOCKTRANCHE
142 : };
143 :
144 : StaticAssertDecl(lengthof(BuiltinTrancheNames) ==
145 : LWTRANCHE_FIRST_USER_DEFINED,
146 : "missing entries in BuiltinTrancheNames[]");
147 :
148 : /*
149 : * This is indexed by tranche ID minus LWTRANCHE_FIRST_USER_DEFINED, and
150 : * points to the shared memory locations of the names of all
151 : * dynamically-created tranches. Backends inherit the pointer by fork from the
152 : * postmaster (except in the EXEC_BACKEND case, where we have special measures
153 : * to pass it down).
154 : */
155 : char **LWLockTrancheNames = NULL;
156 :
157 : /*
158 : * This points to the main array of LWLocks in shared memory. Backends inherit
159 : * the pointer by fork from the postmaster (except in the EXEC_BACKEND case,
160 : * where we have special measures to pass it down).
161 : */
162 : LWLockPadded *MainLWLockArray = NULL;
163 :
164 : /*
165 : * We use this structure to keep track of locked LWLocks for release
166 : * during error recovery. Normally, only a few will be held at once, but
167 : * occasionally the number can be much higher.
168 : */
169 : #define MAX_SIMUL_LWLOCKS 200
170 :
171 : /* struct representing the LWLocks we're holding */
172 : typedef struct LWLockHandle
173 : {
174 : LWLock *lock;
175 : LWLockMode mode;
176 : } LWLockHandle;
177 :
178 : static int num_held_lwlocks = 0;
179 : static LWLockHandle held_lwlocks[MAX_SIMUL_LWLOCKS];
180 :
181 : /* struct representing the LWLock tranche request for named tranche */
182 : typedef struct NamedLWLockTrancheRequest
183 : {
184 : char tranche_name[NAMEDATALEN];
185 : int num_lwlocks;
186 : } NamedLWLockTrancheRequest;
187 :
188 : /*
189 : * NamedLWLockTrancheRequests is the valid length of the request array. These
190 : * variables are non-static so that launch_backend.c can copy them to child
191 : * processes in EXEC_BACKEND builds.
192 : */
193 : int NamedLWLockTrancheRequests = 0;
194 : NamedLWLockTrancheRequest *NamedLWLockTrancheRequestArray = NULL;
195 :
196 : /* postmaster's local copy of the request array */
197 : static NamedLWLockTrancheRequest *LocalNamedLWLockTrancheRequestArray = NULL;
198 :
199 : /* shared memory counter of registered tranches */
200 : int *LWLockCounter = NULL;
201 :
202 : /* backend-local counter of registered tranches */
203 : static int LocalLWLockCounter;
204 :
205 : #define MAX_NAMED_TRANCHES 256
206 :
207 : static void InitializeLWLocks(void);
208 : static inline void LWLockReportWaitStart(LWLock *lock);
209 : static inline void LWLockReportWaitEnd(void);
210 : static const char *GetLWTrancheName(uint16 trancheId);
211 :
212 : #define T_NAME(lock) \
213 : GetLWTrancheName((lock)->tranche)
214 :
215 : #ifdef LWLOCK_STATS
216 : typedef struct lwlock_stats_key
217 : {
218 : int tranche;
219 : void *instance;
220 : } lwlock_stats_key;
221 :
222 : typedef struct lwlock_stats
223 : {
224 : lwlock_stats_key key;
225 : int sh_acquire_count;
226 : int ex_acquire_count;
227 : int block_count;
228 : int dequeue_self_count;
229 : int spin_delay_count;
230 : } lwlock_stats;
231 :
232 : static HTAB *lwlock_stats_htab;
233 : static lwlock_stats lwlock_stats_dummy;
234 : #endif
235 :
236 : #ifdef LOCK_DEBUG
237 : bool Trace_lwlocks = false;
238 :
239 : inline static void
240 : PRINT_LWDEBUG(const char *where, LWLock *lock, LWLockMode mode)
241 : {
242 : /* hide statement & context here, otherwise the log is just too verbose */
243 : if (Trace_lwlocks)
244 : {
245 : uint32 state = pg_atomic_read_u32(&lock->state);
246 :
247 : ereport(LOG,
248 : (errhidestmt(true),
249 : errhidecontext(true),
250 : errmsg_internal("%d: %s(%s %p): excl %u shared %u haswaiters %u waiters %u waking %d",
251 : MyProcPid,
252 : where, T_NAME(lock), lock,
253 : (state & LW_VAL_EXCLUSIVE) != 0,
254 : state & LW_SHARED_MASK,
255 : (state & LW_FLAG_HAS_WAITERS) != 0,
256 : pg_atomic_read_u32(&lock->nwaiters),
257 : (state & LW_FLAG_WAKE_IN_PROGRESS) != 0)));
258 : }
259 : }
260 :
261 : inline static void
262 : LOG_LWDEBUG(const char *where, LWLock *lock, const char *msg)
263 : {
264 : /* hide statement & context here, otherwise the log is just too verbose */
265 : if (Trace_lwlocks)
266 : {
267 : ereport(LOG,
268 : (errhidestmt(true),
269 : errhidecontext(true),
270 : errmsg_internal("%s(%s %p): %s", where,
271 : T_NAME(lock), lock, msg)));
272 : }
273 : }
274 :
275 : #else /* not LOCK_DEBUG */
276 : #define PRINT_LWDEBUG(a,b,c) ((void)0)
277 : #define LOG_LWDEBUG(a,b,c) ((void)0)
278 : #endif /* LOCK_DEBUG */
279 :
280 : #ifdef LWLOCK_STATS
281 :
282 : static void init_lwlock_stats(void);
283 : static void print_lwlock_stats(int code, Datum arg);
284 : static lwlock_stats * get_lwlock_stats_entry(LWLock *lock);
285 :
286 : static void
287 : init_lwlock_stats(void)
288 : {
289 : HASHCTL ctl;
290 : static MemoryContext lwlock_stats_cxt = NULL;
291 : static bool exit_registered = false;
292 :
293 : if (lwlock_stats_cxt != NULL)
294 : MemoryContextDelete(lwlock_stats_cxt);
295 :
296 : /*
297 : * The LWLock stats will be updated within a critical section, which
298 : * requires allocating new hash entries. Allocations within a critical
299 : * section are normally not allowed because running out of memory would
300 : * lead to a PANIC, but LWLOCK_STATS is debugging code that's not normally
301 : * turned on in production, so that's an acceptable risk. The hash entries
302 : * are small, so the risk of running out of memory is minimal in practice.
303 : */
304 : lwlock_stats_cxt = AllocSetContextCreate(TopMemoryContext,
305 : "LWLock stats",
306 : ALLOCSET_DEFAULT_SIZES);
307 : MemoryContextAllowInCriticalSection(lwlock_stats_cxt, true);
308 :
309 : ctl.keysize = sizeof(lwlock_stats_key);
310 : ctl.entrysize = sizeof(lwlock_stats);
311 : ctl.hcxt = lwlock_stats_cxt;
312 : lwlock_stats_htab = hash_create("lwlock stats", 16384, &ctl,
313 : HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
314 : if (!exit_registered)
315 : {
316 : on_shmem_exit(print_lwlock_stats, 0);
317 : exit_registered = true;
318 : }
319 : }
320 :
321 : static void
322 : print_lwlock_stats(int code, Datum arg)
323 : {
324 : HASH_SEQ_STATUS scan;
325 : lwlock_stats *lwstats;
326 :
327 : hash_seq_init(&scan, lwlock_stats_htab);
328 :
329 : /* Grab an LWLock to keep different backends from mixing reports */
330 : LWLockAcquire(&MainLWLockArray[0].lock, LW_EXCLUSIVE);
331 :
332 : while ((lwstats = (lwlock_stats *) hash_seq_search(&scan)) != NULL)
333 : {
334 : fprintf(stderr,
335 : "PID %d lwlock %s %p: shacq %u exacq %u blk %u spindelay %u dequeue self %u\n",
336 : MyProcPid, GetLWTrancheName(lwstats->key.tranche),
337 : lwstats->key.instance, lwstats->sh_acquire_count,
338 : lwstats->ex_acquire_count, lwstats->block_count,
339 : lwstats->spin_delay_count, lwstats->dequeue_self_count);
340 : }
341 :
342 : LWLockRelease(&MainLWLockArray[0].lock);
343 : }
344 :
345 : static lwlock_stats *
346 : get_lwlock_stats_entry(LWLock *lock)
347 : {
348 : lwlock_stats_key key;
349 : lwlock_stats *lwstats;
350 : bool found;
351 :
352 : /*
353 : * During shared memory initialization, the hash table doesn't exist yet.
354 : * Stats of that phase aren't very interesting, so just collect operations
355 : * on all locks in a single dummy entry.
356 : */
357 : if (lwlock_stats_htab == NULL)
358 : return &lwlock_stats_dummy;
359 :
360 : /* Fetch or create the entry. */
361 : MemSet(&key, 0, sizeof(key));
362 : key.tranche = lock->tranche;
363 : key.instance = lock;
364 : lwstats = hash_search(lwlock_stats_htab, &key, HASH_ENTER, &found);
365 : if (!found)
366 : {
367 : lwstats->sh_acquire_count = 0;
368 : lwstats->ex_acquire_count = 0;
369 : lwstats->block_count = 0;
370 : lwstats->dequeue_self_count = 0;
371 : lwstats->spin_delay_count = 0;
372 : }
373 : return lwstats;
374 : }
375 : #endif /* LWLOCK_STATS */
376 :
377 :
378 : /*
379 : * Compute number of LWLocks required by named tranches. These will be
380 : * allocated in the main array.
381 : */
382 : static int
383 3387 : NumLWLocksForNamedTranches(void)
384 : {
385 3387 : int numLocks = 0;
386 : int i;
387 :
388 3504 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
389 117 : numLocks += NamedLWLockTrancheRequestArray[i].num_lwlocks;
390 :
391 3387 : return numLocks;
392 : }
393 :
394 : /*
395 : * Compute shmem space needed for LWLocks and named tranches.
396 : */
397 : Size
398 3387 : LWLockShmemSize(void)
399 : {
400 : Size size;
401 3387 : int numLocks = NUM_FIXED_LWLOCKS;
402 :
403 : /*
404 : * If re-initializing shared memory, the request array will no longer be
405 : * accessible, so switch to the copy in postmaster's local memory. We'll
406 : * copy it back into shared memory later when CreateLWLocks() is called
407 : * again.
408 : */
409 3387 : if (LocalNamedLWLockTrancheRequestArray)
410 0 : NamedLWLockTrancheRequestArray = LocalNamedLWLockTrancheRequestArray;
411 :
412 : /* Calculate total number of locks needed in the main array. */
413 3387 : numLocks += NumLWLocksForNamedTranches();
414 :
415 : /* Space for dynamic allocation counter. */
416 3387 : size = MAXALIGN(sizeof(int));
417 :
418 : /* Space for named tranches. */
419 3387 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, sizeof(char *)));
420 3387 : size = add_size(size, mul_size(MAX_NAMED_TRANCHES, NAMEDATALEN));
421 :
422 : /*
423 : * Make space for named tranche requests. This is done for the benefit of
424 : * EXEC_BACKEND builds, which otherwise wouldn't be able to call
425 : * GetNamedLWLockTranche() outside postmaster.
426 : */
427 3387 : size = add_size(size, mul_size(NamedLWLockTrancheRequests,
428 : sizeof(NamedLWLockTrancheRequest)));
429 :
430 : /* Space for the LWLock array, plus room for cache line alignment. */
431 3387 : size = add_size(size, LWLOCK_PADDED_SIZE);
432 3387 : size = add_size(size, mul_size(numLocks, sizeof(LWLockPadded)));
433 :
434 3387 : return size;
435 : }
436 :
437 : /*
438 : * Allocate shmem space for the main LWLock array and all tranches and
439 : * initialize it.
440 : */
441 : void
442 1180 : CreateLWLocks(void)
443 : {
444 1180 : if (!IsUnderPostmaster)
445 : {
446 1180 : Size spaceLocks = LWLockShmemSize();
447 : char *ptr;
448 :
449 : /* Allocate space */
450 1180 : ptr = (char *) ShmemAlloc(spaceLocks);
451 :
452 : /* Initialize the dynamic-allocation counter for tranches */
453 1180 : LWLockCounter = (int *) ptr;
454 1180 : *LWLockCounter = LWTRANCHE_FIRST_USER_DEFINED;
455 1180 : ptr += MAXALIGN(sizeof(int));
456 :
457 : /* Initialize tranche names */
458 1180 : LWLockTrancheNames = (char **) ptr;
459 1180 : ptr += MAX_NAMED_TRANCHES * sizeof(char *);
460 303260 : for (int i = 0; i < MAX_NAMED_TRANCHES; i++)
461 : {
462 302080 : LWLockTrancheNames[i] = ptr;
463 302080 : ptr += NAMEDATALEN;
464 : }
465 :
466 : /*
467 : * Move named tranche requests to shared memory. This is done for the
468 : * benefit of EXEC_BACKEND builds, which otherwise wouldn't be able to
469 : * call GetNamedLWLockTranche() outside postmaster.
470 : */
471 1180 : if (NamedLWLockTrancheRequests > 0)
472 : {
473 : /*
474 : * Save the pointer to the request array in postmaster's local
475 : * memory. We'll need it if we ever need to re-initialize shared
476 : * memory after a crash.
477 : */
478 8 : LocalNamedLWLockTrancheRequestArray = NamedLWLockTrancheRequestArray;
479 :
480 8 : memcpy(ptr, NamedLWLockTrancheRequestArray,
481 : NamedLWLockTrancheRequests * sizeof(NamedLWLockTrancheRequest));
482 8 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *) ptr;
483 8 : ptr += NamedLWLockTrancheRequests * sizeof(NamedLWLockTrancheRequest);
484 : }
485 :
486 : /* Ensure desired alignment of LWLock array */
487 1180 : ptr += LWLOCK_PADDED_SIZE - ((uintptr_t) ptr) % LWLOCK_PADDED_SIZE;
488 1180 : MainLWLockArray = (LWLockPadded *) ptr;
489 :
490 : /* Initialize all LWLocks */
491 1180 : InitializeLWLocks();
492 : }
493 1180 : }
494 :
495 : /*
496 : * Initialize LWLocks that are fixed and those belonging to named tranches.
497 : */
498 : static void
499 1180 : InitializeLWLocks(void)
500 : {
501 : int id;
502 : int i;
503 : int j;
504 : LWLockPadded *lock;
505 :
506 : /* Initialize all individual LWLocks in main array */
507 67260 : for (id = 0, lock = MainLWLockArray; id < NUM_INDIVIDUAL_LWLOCKS; id++, lock++)
508 66080 : LWLockInitialize(&lock->lock, id);
509 :
510 : /* Initialize buffer mapping LWLocks in main array */
511 1180 : lock = MainLWLockArray + BUFFER_MAPPING_LWLOCK_OFFSET;
512 152220 : for (id = 0; id < NUM_BUFFER_PARTITIONS; id++, lock++)
513 151040 : LWLockInitialize(&lock->lock, LWTRANCHE_BUFFER_MAPPING);
514 :
515 : /* Initialize lmgrs' LWLocks in main array */
516 1180 : lock = MainLWLockArray + LOCK_MANAGER_LWLOCK_OFFSET;
517 20060 : for (id = 0; id < NUM_LOCK_PARTITIONS; id++, lock++)
518 18880 : LWLockInitialize(&lock->lock, LWTRANCHE_LOCK_MANAGER);
519 :
520 : /* Initialize predicate lmgrs' LWLocks in main array */
521 1180 : lock = MainLWLockArray + PREDICATELOCK_MANAGER_LWLOCK_OFFSET;
522 20060 : for (id = 0; id < NUM_PREDICATELOCK_PARTITIONS; id++, lock++)
523 18880 : LWLockInitialize(&lock->lock, LWTRANCHE_PREDICATE_LOCK_MANAGER);
524 :
525 : /*
526 : * Copy the info about any named tranches into shared memory (so that
527 : * other processes can see it), and initialize the requested LWLocks.
528 : */
529 1180 : if (NamedLWLockTrancheRequests > 0)
530 : {
531 8 : lock = &MainLWLockArray[NUM_FIXED_LWLOCKS];
532 :
533 47 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
534 : {
535 : NamedLWLockTrancheRequest *request;
536 : int tranche;
537 :
538 39 : request = &NamedLWLockTrancheRequestArray[i];
539 39 : tranche = LWLockNewTrancheId(request->tranche_name);
540 :
541 78 : for (j = 0; j < request->num_lwlocks; j++, lock++)
542 39 : LWLockInitialize(&lock->lock, tranche);
543 : }
544 : }
545 1180 : }
546 :
547 : /*
548 : * InitLWLockAccess - initialize backend-local state needed to hold LWLocks
549 : */
550 : void
551 24305 : InitLWLockAccess(void)
552 : {
553 : #ifdef LWLOCK_STATS
554 : init_lwlock_stats();
555 : #endif
556 24305 : }
557 :
558 : /*
559 : * GetNamedLWLockTranche - returns the base address of LWLock from the
560 : * specified tranche.
561 : *
562 : * Caller needs to retrieve the requested number of LWLocks starting from
563 : * the base lock address returned by this API. This can be used for
564 : * tranches that are requested by using RequestNamedLWLockTranche() API.
565 : */
566 : LWLockPadded *
567 9 : GetNamedLWLockTranche(const char *tranche_name)
568 : {
569 : int lock_pos;
570 : int i;
571 :
572 : /*
573 : * Obtain the position of base address of LWLock belonging to requested
574 : * tranche_name in MainLWLockArray. LWLocks for named tranches are placed
575 : * in MainLWLockArray after fixed locks.
576 : */
577 9 : lock_pos = NUM_FIXED_LWLOCKS;
578 41 : for (i = 0; i < NamedLWLockTrancheRequests; i++)
579 : {
580 40 : if (strcmp(NamedLWLockTrancheRequestArray[i].tranche_name,
581 : tranche_name) == 0)
582 8 : return &MainLWLockArray[lock_pos];
583 :
584 32 : lock_pos += NamedLWLockTrancheRequestArray[i].num_lwlocks;
585 : }
586 :
587 1 : elog(ERROR, "requested tranche is not registered");
588 :
589 : /* just to keep compiler quiet */
590 : return NULL;
591 : }
592 :
593 : /*
594 : * Allocate a new tranche ID with the provided name.
595 : */
596 : int
597 283 : LWLockNewTrancheId(const char *name)
598 : {
599 : int result;
600 :
601 283 : if (!name)
602 1 : ereport(ERROR,
603 : (errcode(ERRCODE_INVALID_NAME),
604 : errmsg("tranche name cannot be NULL")));
605 :
606 282 : if (strlen(name) >= NAMEDATALEN)
607 1 : ereport(ERROR,
608 : (errcode(ERRCODE_NAME_TOO_LONG),
609 : errmsg("tranche name too long"),
610 : errdetail("LWLock tranche names must be no longer than %d bytes.",
611 : NAMEDATALEN - 1)));
612 :
613 : /*
614 : * We use the ShmemLock spinlock to protect LWLockCounter and
615 : * LWLockTrancheNames.
616 : */
617 281 : SpinLockAcquire(ShmemLock);
618 :
619 281 : if (*LWLockCounter - LWTRANCHE_FIRST_USER_DEFINED >= MAX_NAMED_TRANCHES)
620 : {
621 1 : SpinLockRelease(ShmemLock);
622 1 : ereport(ERROR,
623 : (errmsg("maximum number of tranches already registered"),
624 : errdetail("No more than %d tranches may be registered.",
625 : MAX_NAMED_TRANCHES)));
626 : }
627 :
628 280 : result = (*LWLockCounter)++;
629 280 : LocalLWLockCounter = *LWLockCounter;
630 280 : strlcpy(LWLockTrancheNames[result - LWTRANCHE_FIRST_USER_DEFINED], name, NAMEDATALEN);
631 :
632 280 : SpinLockRelease(ShmemLock);
633 :
634 280 : return result;
635 : }
636 :
637 : /*
638 : * RequestNamedLWLockTranche
639 : * Request that extra LWLocks be allocated during postmaster
640 : * startup.
641 : *
642 : * This may only be called via the shmem_request_hook of a library that is
643 : * loaded into the postmaster via shared_preload_libraries. Calls from
644 : * elsewhere will fail.
645 : *
646 : * The tranche name will be user-visible as a wait event name, so try to
647 : * use a name that fits the style for those.
648 : */
649 : void
650 39 : RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
651 : {
652 : NamedLWLockTrancheRequest *request;
653 : static int NamedLWLockTrancheRequestsAllocated;
654 :
655 39 : if (!process_shmem_requests_in_progress)
656 0 : elog(FATAL, "cannot request additional LWLocks outside shmem_request_hook");
657 :
658 39 : if (!tranche_name)
659 0 : ereport(ERROR,
660 : (errcode(ERRCODE_INVALID_NAME),
661 : errmsg("tranche name cannot be NULL")));
662 :
663 39 : if (strlen(tranche_name) >= NAMEDATALEN)
664 0 : ereport(ERROR,
665 : (errcode(ERRCODE_NAME_TOO_LONG),
666 : errmsg("tranche name too long"),
667 : errdetail("LWLock tranche names must be no longer than %d bytes.",
668 : NAMEDATALEN - 1)));
669 :
670 39 : if (NamedLWLockTrancheRequestArray == NULL)
671 : {
672 8 : NamedLWLockTrancheRequestsAllocated = 16;
673 8 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
674 8 : MemoryContextAlloc(TopMemoryContext,
675 : NamedLWLockTrancheRequestsAllocated
676 : * sizeof(NamedLWLockTrancheRequest));
677 : }
678 :
679 39 : if (NamedLWLockTrancheRequests >= NamedLWLockTrancheRequestsAllocated)
680 : {
681 1 : int i = pg_nextpower2_32(NamedLWLockTrancheRequests + 1);
682 :
683 1 : NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
684 1 : repalloc(NamedLWLockTrancheRequestArray,
685 : i * sizeof(NamedLWLockTrancheRequest));
686 1 : NamedLWLockTrancheRequestsAllocated = i;
687 : }
688 :
689 39 : request = &NamedLWLockTrancheRequestArray[NamedLWLockTrancheRequests];
690 39 : strlcpy(request->tranche_name, tranche_name, NAMEDATALEN);
691 39 : request->num_lwlocks = num_lwlocks;
692 39 : NamedLWLockTrancheRequests++;
693 39 : }
694 :
695 : /*
696 : * LWLockInitialize - initialize a new lwlock; it's initially unlocked
697 : */
698 : void
699 2450503 : LWLockInitialize(LWLock *lock, int tranche_id)
700 : {
701 : /* verify the tranche_id is valid */
702 2450503 : (void) GetLWTrancheName(tranche_id);
703 :
704 2450502 : pg_atomic_init_u32(&lock->state, 0);
705 : #ifdef LOCK_DEBUG
706 : pg_atomic_init_u32(&lock->nwaiters, 0);
707 : #endif
708 2450502 : lock->tranche = tranche_id;
709 2450502 : proclist_init(&lock->waiters);
710 2450502 : }
711 :
712 : /*
713 : * Report start of wait event for light-weight locks.
714 : *
715 : * This function will be used by all the light-weight lock calls which
716 : * needs to wait to acquire the lock. This function distinguishes wait
717 : * event based on tranche and lock id.
718 : */
719 : static inline void
720 5617796 : LWLockReportWaitStart(LWLock *lock)
721 : {
722 5617796 : pgstat_report_wait_start(PG_WAIT_LWLOCK | lock->tranche);
723 5617796 : }
724 :
725 : /*
726 : * Report end of wait event for light-weight locks.
727 : */
728 : static inline void
729 5617796 : LWLockReportWaitEnd(void)
730 : {
731 5617796 : pgstat_report_wait_end();
732 5617796 : }
733 :
734 : /*
735 : * Return the name of an LWLock tranche.
736 : */
737 : static const char *
738 2450788 : GetLWTrancheName(uint16 trancheId)
739 : {
740 : /* Built-in tranche or individual LWLock? */
741 2450788 : if (trancheId < LWTRANCHE_FIRST_USER_DEFINED)
742 2449590 : return BuiltinTrancheNames[trancheId];
743 :
744 : /*
745 : * We only ever add new entries to LWLockTrancheNames, so most lookups can
746 : * avoid taking the spinlock as long as the backend-local counter
747 : * (LocalLWLockCounter) is greater than the requested tranche ID. Else,
748 : * we need to first update the backend-local counter with ShmemLock held
749 : * before attempting the lookup again. In practice, the latter case is
750 : * probably rare.
751 : */
752 1198 : if (trancheId >= LocalLWLockCounter)
753 : {
754 1 : SpinLockAcquire(ShmemLock);
755 1 : LocalLWLockCounter = *LWLockCounter;
756 1 : SpinLockRelease(ShmemLock);
757 :
758 1 : if (trancheId >= LocalLWLockCounter)
759 1 : elog(ERROR, "tranche %d is not registered", trancheId);
760 : }
761 :
762 : /*
763 : * It's an extension tranche, so look in LWLockTrancheNames.
764 : */
765 1197 : trancheId -= LWTRANCHE_FIRST_USER_DEFINED;
766 :
767 1197 : return LWLockTrancheNames[trancheId];
768 : }
769 :
770 : /*
771 : * Return an identifier for an LWLock based on the wait class and event.
772 : */
773 : const char *
774 285 : GetLWLockIdentifier(uint32 classId, uint16 eventId)
775 : {
776 : Assert(classId == PG_WAIT_LWLOCK);
777 : /* The event IDs are just tranche numbers. */
778 285 : return GetLWTrancheName(eventId);
779 : }
780 :
781 : /*
782 : * Internal function that tries to atomically acquire the lwlock in the passed
783 : * in mode.
784 : *
785 : * This function will not block waiting for a lock to become free - that's the
786 : * caller's job.
787 : *
788 : * Returns true if the lock isn't free and we need to wait.
789 : */
790 : static bool
791 344092176 : LWLockAttemptLock(LWLock *lock, LWLockMode mode)
792 : {
793 : uint32 old_state;
794 :
795 : Assert(mode == LW_EXCLUSIVE || mode == LW_SHARED);
796 :
797 : /*
798 : * Read once outside the loop, later iterations will get the newer value
799 : * via compare & exchange.
800 : */
801 344092176 : old_state = pg_atomic_read_u32(&lock->state);
802 :
803 : /* loop until we've determined whether we could acquire the lock or not */
804 : while (true)
805 356552 : {
806 : uint32 desired_state;
807 : bool lock_free;
808 :
809 344448728 : desired_state = old_state;
810 :
811 344448728 : if (mode == LW_EXCLUSIVE)
812 : {
813 232485590 : lock_free = (old_state & LW_LOCK_MASK) == 0;
814 232485590 : if (lock_free)
815 230393298 : desired_state += LW_VAL_EXCLUSIVE;
816 : }
817 : else
818 : {
819 111963138 : lock_free = (old_state & LW_VAL_EXCLUSIVE) == 0;
820 111963138 : if (lock_free)
821 102799524 : desired_state += LW_VAL_SHARED;
822 : }
823 :
824 : /*
825 : * Attempt to swap in the state we are expecting. If we didn't see
826 : * lock to be free, that's just the old value. If we saw it as free,
827 : * we'll attempt to mark it acquired. The reason that we always swap
828 : * in the value is that this doubles as a memory barrier. We could try
829 : * to be smarter and only swap in values if we saw the lock as free,
830 : * but benchmark haven't shown it as beneficial so far.
831 : *
832 : * Retry if the value changed since we last looked at it.
833 : */
834 344448728 : if (pg_atomic_compare_exchange_u32(&lock->state,
835 : &old_state, desired_state))
836 : {
837 344092176 : if (lock_free)
838 : {
839 : /* Great! Got the lock. */
840 : #ifdef LOCK_DEBUG
841 : if (mode == LW_EXCLUSIVE)
842 : lock->owner = MyProc;
843 : #endif
844 332970095 : return false;
845 : }
846 : else
847 11122081 : return true; /* somebody else has the lock */
848 : }
849 : }
850 : pg_unreachable();
851 : }
852 :
853 : /*
854 : * Lock the LWLock's wait list against concurrent activity.
855 : *
856 : * NB: even though the wait list is locked, non-conflicting lock operations
857 : * may still happen concurrently.
858 : *
859 : * Time spent holding mutex should be short!
860 : */
861 : static void
862 13625567 : LWLockWaitListLock(LWLock *lock)
863 : {
864 : uint32 old_state;
865 : #ifdef LWLOCK_STATS
866 : lwlock_stats *lwstats;
867 : uint32 delays = 0;
868 :
869 : lwstats = get_lwlock_stats_entry(lock);
870 : #endif
871 :
872 : while (true)
873 : {
874 : /*
875 : * Always try once to acquire the lock directly, without setting up
876 : * the spin-delay infrastructure. The work necessary for that shows up
877 : * in profiles and is rarely necessary.
878 : */
879 13683699 : old_state = pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_LOCKED);
880 13683699 : if (likely(!(old_state & LW_FLAG_LOCKED)))
881 13625567 : break; /* got lock */
882 :
883 : /* and then spin without atomic operations until lock is released */
884 : {
885 : SpinDelayStatus delayStatus;
886 :
887 58132 : init_local_spin_delay(&delayStatus);
888 :
889 244299 : while (old_state & LW_FLAG_LOCKED)
890 : {
891 186167 : perform_spin_delay(&delayStatus);
892 186167 : old_state = pg_atomic_read_u32(&lock->state);
893 : }
894 : #ifdef LWLOCK_STATS
895 : delays += delayStatus.delays;
896 : #endif
897 58132 : finish_spin_delay(&delayStatus);
898 : }
899 :
900 : /*
901 : * Retry. The lock might obviously already be re-acquired by the time
902 : * we're attempting to get it again.
903 : */
904 : }
905 :
906 : #ifdef LWLOCK_STATS
907 : lwstats->spin_delay_count += delays;
908 : #endif
909 13625567 : }
910 :
911 : /*
912 : * Unlock the LWLock's wait list.
913 : *
914 : * Note that it can be more efficient to manipulate flags and release the
915 : * locks in a single atomic operation.
916 : */
917 : static void
918 8455373 : LWLockWaitListUnlock(LWLock *lock)
919 : {
920 : uint32 old_state PG_USED_FOR_ASSERTS_ONLY;
921 :
922 8455373 : old_state = pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_LOCKED);
923 :
924 : Assert(old_state & LW_FLAG_LOCKED);
925 8455373 : }
926 :
927 : /*
928 : * Wakeup all the lockers that currently have a chance to acquire the lock.
929 : */
930 : static void
931 5170194 : LWLockWakeup(LWLock *lock)
932 : {
933 5170194 : bool new_wake_in_progress = false;
934 5170194 : bool wokeup_somebody = false;
935 : proclist_head wakeup;
936 : proclist_mutable_iter iter;
937 :
938 5170194 : proclist_init(&wakeup);
939 :
940 : /* lock wait list while collecting backends to wake up */
941 5170194 : LWLockWaitListLock(lock);
942 :
943 9848116 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
944 : {
945 5669597 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
946 :
947 5669597 : if (wokeup_somebody && waiter->lwWaitMode == LW_EXCLUSIVE)
948 28658 : continue;
949 :
950 5640939 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
951 5640939 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
952 :
953 5640939 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
954 : {
955 : /*
956 : * Prevent additional wakeups until retryer gets to run. Backends
957 : * that are just waiting for the lock to become free don't retry
958 : * automatically.
959 : */
960 5531164 : new_wake_in_progress = true;
961 :
962 : /*
963 : * Don't wakeup (further) exclusive locks.
964 : */
965 5531164 : wokeup_somebody = true;
966 : }
967 :
968 : /*
969 : * Signal that the process isn't on the wait list anymore. This allows
970 : * LWLockDequeueSelf() to remove itself of the waitlist with a
971 : * proclist_delete(), rather than having to check if it has been
972 : * removed from the list.
973 : */
974 : Assert(waiter->lwWaiting == LW_WS_WAITING);
975 5640939 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
976 :
977 : /*
978 : * Once we've woken up an exclusive lock, there's no point in waking
979 : * up anybody else.
980 : */
981 5640939 : if (waiter->lwWaitMode == LW_EXCLUSIVE)
982 991675 : break;
983 : }
984 :
985 : Assert(proclist_is_empty(&wakeup) || pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS);
986 :
987 : /* unset required flags, and release lock, in one fell swoop */
988 : {
989 : uint32 old_state;
990 : uint32 desired_state;
991 :
992 5170194 : old_state = pg_atomic_read_u32(&lock->state);
993 : while (true)
994 : {
995 5178115 : desired_state = old_state;
996 :
997 : /* compute desired flags */
998 :
999 5178115 : if (new_wake_in_progress)
1000 5108795 : desired_state |= LW_FLAG_WAKE_IN_PROGRESS;
1001 : else
1002 69320 : desired_state &= ~LW_FLAG_WAKE_IN_PROGRESS;
1003 :
1004 5178115 : if (proclist_is_empty(&lock->waiters))
1005 5093688 : desired_state &= ~LW_FLAG_HAS_WAITERS;
1006 :
1007 5178115 : desired_state &= ~LW_FLAG_LOCKED; /* release lock */
1008 :
1009 5178115 : if (pg_atomic_compare_exchange_u32(&lock->state, &old_state,
1010 : desired_state))
1011 5170194 : break;
1012 : }
1013 : }
1014 :
1015 : /* Awaken any waiters I removed from the queue. */
1016 10811133 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1017 : {
1018 5640939 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1019 :
1020 : LOG_LWDEBUG("LWLockRelease", lock, "release waiter");
1021 5640939 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1022 :
1023 : /*
1024 : * Guarantee that lwWaiting being unset only becomes visible once the
1025 : * unlink from the link has completed. Otherwise the target backend
1026 : * could be woken up for other reason and enqueue for a new lock - if
1027 : * that happens before the list unlink happens, the list would end up
1028 : * being corrupted.
1029 : *
1030 : * The barrier pairs with the LWLockWaitListLock() when enqueuing for
1031 : * another lock.
1032 : */
1033 5640939 : pg_write_barrier();
1034 5640939 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1035 5640939 : PGSemaphoreUnlock(waiter->sem);
1036 : }
1037 5170194 : }
1038 :
1039 : /*
1040 : * Add ourselves to the end of the queue.
1041 : *
1042 : * NB: Mode can be LW_WAIT_UNTIL_FREE here!
1043 : */
1044 : static void
1045 5718495 : LWLockQueueSelf(LWLock *lock, LWLockMode mode)
1046 : {
1047 : /*
1048 : * If we don't have a PGPROC structure, there's no way to wait. This
1049 : * should never occur, since MyProc should only be null during shared
1050 : * memory initialization.
1051 : */
1052 5718495 : if (MyProc == NULL)
1053 0 : elog(PANIC, "cannot wait without a PGPROC structure");
1054 :
1055 5718495 : if (MyProc->lwWaiting != LW_WS_NOT_WAITING)
1056 0 : elog(PANIC, "queueing for lock while waiting on another one");
1057 :
1058 5718495 : LWLockWaitListLock(lock);
1059 :
1060 : /* setting the flag is protected by the spinlock */
1061 5718495 : pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_HAS_WAITERS);
1062 :
1063 5718495 : MyProc->lwWaiting = LW_WS_WAITING;
1064 5718495 : MyProc->lwWaitMode = mode;
1065 :
1066 : /* LW_WAIT_UNTIL_FREE waiters are always at the front of the queue */
1067 5718495 : if (mode == LW_WAIT_UNTIL_FREE)
1068 111984 : proclist_push_head(&lock->waiters, MyProcNumber, lwWaitLink);
1069 : else
1070 5606511 : proclist_push_tail(&lock->waiters, MyProcNumber, lwWaitLink);
1071 :
1072 : /* Can release the mutex now */
1073 5718495 : LWLockWaitListUnlock(lock);
1074 :
1075 : #ifdef LOCK_DEBUG
1076 : pg_atomic_fetch_add_u32(&lock->nwaiters, 1);
1077 : #endif
1078 5718495 : }
1079 :
1080 : /*
1081 : * Remove ourselves from the waitlist.
1082 : *
1083 : * This is used if we queued ourselves because we thought we needed to sleep
1084 : * but, after further checking, we discovered that we don't actually need to
1085 : * do so.
1086 : */
1087 : static void
1088 100699 : LWLockDequeueSelf(LWLock *lock)
1089 : {
1090 : bool on_waitlist;
1091 :
1092 : #ifdef LWLOCK_STATS
1093 : lwlock_stats *lwstats;
1094 :
1095 : lwstats = get_lwlock_stats_entry(lock);
1096 :
1097 : lwstats->dequeue_self_count++;
1098 : #endif
1099 :
1100 100699 : LWLockWaitListLock(lock);
1101 :
1102 : /*
1103 : * Remove ourselves from the waitlist, unless we've already been removed.
1104 : * The removal happens with the wait list lock held, so there's no race in
1105 : * this check.
1106 : */
1107 100699 : on_waitlist = MyProc->lwWaiting == LW_WS_WAITING;
1108 100699 : if (on_waitlist)
1109 76808 : proclist_delete(&lock->waiters, MyProcNumber, lwWaitLink);
1110 :
1111 100699 : if (proclist_is_empty(&lock->waiters) &&
1112 92239 : (pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS) != 0)
1113 : {
1114 74050 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_HAS_WAITERS);
1115 : }
1116 :
1117 : /* XXX: combine with fetch_and above? */
1118 100699 : LWLockWaitListUnlock(lock);
1119 :
1120 : /* clear waiting state again, nice for debugging */
1121 100699 : if (on_waitlist)
1122 76808 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
1123 : else
1124 : {
1125 23891 : int extraWaits = 0;
1126 :
1127 : /*
1128 : * Somebody else dequeued us and has or will wake us up. Deal with the
1129 : * superfluous absorption of a wakeup.
1130 : */
1131 :
1132 : /*
1133 : * Clear LW_FLAG_WAKE_IN_PROGRESS if somebody woke us before we
1134 : * removed ourselves - they'll have set it.
1135 : */
1136 23891 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_WAKE_IN_PROGRESS);
1137 :
1138 : /*
1139 : * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
1140 : * get reset at some inconvenient point later. Most of the time this
1141 : * will immediately return.
1142 : */
1143 : for (;;)
1144 : {
1145 23891 : PGSemaphoreLock(MyProc->sem);
1146 23891 : if (MyProc->lwWaiting == LW_WS_NOT_WAITING)
1147 23891 : break;
1148 0 : extraWaits++;
1149 : }
1150 :
1151 : /*
1152 : * Fix the process wait semaphore's count for any absorbed wakeups.
1153 : */
1154 23891 : while (extraWaits-- > 0)
1155 0 : PGSemaphoreUnlock(MyProc->sem);
1156 : }
1157 :
1158 : #ifdef LOCK_DEBUG
1159 : {
1160 : /* not waiting anymore */
1161 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1162 :
1163 : Assert(nwaiters < MAX_BACKENDS);
1164 : }
1165 : #endif
1166 100699 : }
1167 :
1168 : /*
1169 : * LWLockAcquire - acquire a lightweight lock in the specified mode
1170 : *
1171 : * If the lock is not available, sleep until it is. Returns true if the lock
1172 : * was available immediately, false if we had to sleep.
1173 : *
1174 : * Side effect: cancel/die interrupts are held off until lock release.
1175 : */
1176 : bool
1177 331397157 : LWLockAcquire(LWLock *lock, LWLockMode mode)
1178 : {
1179 331397157 : PGPROC *proc = MyProc;
1180 331397157 : bool result = true;
1181 331397157 : int extraWaits = 0;
1182 : #ifdef LWLOCK_STATS
1183 : lwlock_stats *lwstats;
1184 :
1185 : lwstats = get_lwlock_stats_entry(lock);
1186 : #endif
1187 :
1188 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1189 :
1190 : PRINT_LWDEBUG("LWLockAcquire", lock, mode);
1191 :
1192 : #ifdef LWLOCK_STATS
1193 : /* Count lock acquisition attempts */
1194 : if (mode == LW_EXCLUSIVE)
1195 : lwstats->ex_acquire_count++;
1196 : else
1197 : lwstats->sh_acquire_count++;
1198 : #endif /* LWLOCK_STATS */
1199 :
1200 : /*
1201 : * We can't wait if we haven't got a PGPROC. This should only occur
1202 : * during bootstrap or shared memory initialization. Put an Assert here
1203 : * to catch unsafe coding practices.
1204 : */
1205 : Assert(!(proc == NULL && IsUnderPostmaster));
1206 :
1207 : /* Ensure we will have room to remember the lock */
1208 331397157 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1209 0 : elog(ERROR, "too many LWLocks taken");
1210 :
1211 : /*
1212 : * Lock out cancel/die interrupts until we exit the code section protected
1213 : * by the LWLock. This ensures that interrupts will not interfere with
1214 : * manipulations of data structures in shared memory.
1215 : */
1216 331397157 : HOLD_INTERRUPTS();
1217 :
1218 : /*
1219 : * Loop here to try to acquire lock after each time we are signaled by
1220 : * LWLockRelease.
1221 : *
1222 : * NOTE: it might seem better to have LWLockRelease actually grant us the
1223 : * lock, rather than retrying and possibly having to go back to sleep. But
1224 : * in practice that is no good because it means a process swap for every
1225 : * lock acquisition when two or more processes are contending for the same
1226 : * lock. Since LWLocks are normally used to protect not-very-long
1227 : * sections of computation, a process needs to be able to acquire and
1228 : * release the same lock many times during a single CPU time slice, even
1229 : * in the presence of contention. The efficiency of being able to do that
1230 : * outweighs the inefficiency of sometimes wasting a process dispatch
1231 : * cycle because the lock is not free when a released waiter finally gets
1232 : * to run. See pgsql-hackers archives for 29-Dec-01.
1233 : */
1234 : for (;;)
1235 5507378 : {
1236 : bool mustwait;
1237 :
1238 : /*
1239 : * Try to grab the lock the first time, we're not in the waitqueue
1240 : * yet/anymore.
1241 : */
1242 336904535 : mustwait = LWLockAttemptLock(lock, mode);
1243 :
1244 336904535 : if (!mustwait)
1245 : {
1246 : LOG_LWDEBUG("LWLockAcquire", lock, "immediately acquired lock");
1247 331298024 : break; /* got the lock */
1248 : }
1249 :
1250 : /*
1251 : * Ok, at this point we couldn't grab the lock on the first try. We
1252 : * cannot simply queue ourselves to the end of the list and wait to be
1253 : * woken up because by now the lock could long have been released.
1254 : * Instead add us to the queue and try to grab the lock again. If we
1255 : * succeed we need to revert the queuing and be happy, otherwise we
1256 : * recheck the lock. If we still couldn't grab it, we know that the
1257 : * other locker will see our queue entries when releasing since they
1258 : * existed before we checked for the lock.
1259 : */
1260 :
1261 : /* add to the queue */
1262 5606511 : LWLockQueueSelf(lock, mode);
1263 :
1264 : /* we're now guaranteed to be woken up if necessary */
1265 5606511 : mustwait = LWLockAttemptLock(lock, mode);
1266 :
1267 : /* ok, grabbed the lock the second time round, need to undo queueing */
1268 5606511 : if (!mustwait)
1269 : {
1270 : LOG_LWDEBUG("LWLockAcquire", lock, "acquired, undoing queue");
1271 :
1272 99133 : LWLockDequeueSelf(lock);
1273 99133 : break;
1274 : }
1275 :
1276 : /*
1277 : * Wait until awakened.
1278 : *
1279 : * It is possible that we get awakened for a reason other than being
1280 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1281 : * we've gotten the LWLock, re-increment the sema by the number of
1282 : * additional signals received.
1283 : */
1284 : LOG_LWDEBUG("LWLockAcquire", lock, "waiting");
1285 :
1286 : #ifdef LWLOCK_STATS
1287 : lwstats->block_count++;
1288 : #endif
1289 :
1290 5507378 : LWLockReportWaitStart(lock);
1291 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1292 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1293 :
1294 : for (;;)
1295 : {
1296 5507378 : PGSemaphoreLock(proc->sem);
1297 5507378 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1298 5507378 : break;
1299 0 : extraWaits++;
1300 : }
1301 :
1302 : /* Retrying, allow LWLockRelease to release waiters again. */
1303 5507378 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_WAKE_IN_PROGRESS);
1304 :
1305 : #ifdef LOCK_DEBUG
1306 : {
1307 : /* not waiting anymore */
1308 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1309 :
1310 : Assert(nwaiters < MAX_BACKENDS);
1311 : }
1312 : #endif
1313 :
1314 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1315 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1316 5507378 : LWLockReportWaitEnd();
1317 :
1318 : LOG_LWDEBUG("LWLockAcquire", lock, "awakened");
1319 :
1320 : /* Now loop back and try to acquire lock again. */
1321 5507378 : result = false;
1322 : }
1323 :
1324 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_ENABLED())
1325 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE(T_NAME(lock), mode);
1326 :
1327 : /* Add lock to list of locks held by this backend */
1328 331397157 : held_lwlocks[num_held_lwlocks].lock = lock;
1329 331397157 : held_lwlocks[num_held_lwlocks++].mode = mode;
1330 :
1331 : /*
1332 : * Fix the process wait semaphore's count for any absorbed wakeups.
1333 : */
1334 331397157 : while (extraWaits-- > 0)
1335 0 : PGSemaphoreUnlock(proc->sem);
1336 :
1337 331397157 : return result;
1338 : }
1339 :
1340 : /*
1341 : * LWLockConditionalAcquire - acquire a lightweight lock in the specified mode
1342 : *
1343 : * If the lock is not available, return false with no side-effects.
1344 : *
1345 : * If successful, cancel/die interrupts are held off until lock release.
1346 : */
1347 : bool
1348 1419091 : LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
1349 : {
1350 : bool mustwait;
1351 :
1352 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1353 :
1354 : PRINT_LWDEBUG("LWLockConditionalAcquire", lock, mode);
1355 :
1356 : /* Ensure we will have room to remember the lock */
1357 1419091 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1358 0 : elog(ERROR, "too many LWLocks taken");
1359 :
1360 : /*
1361 : * Lock out cancel/die interrupts until we exit the code section protected
1362 : * by the LWLock. This ensures that interrupts will not interfere with
1363 : * manipulations of data structures in shared memory.
1364 : */
1365 1419091 : HOLD_INTERRUPTS();
1366 :
1367 : /* Check for the lock */
1368 1419091 : mustwait = LWLockAttemptLock(lock, mode);
1369 :
1370 1419091 : if (mustwait)
1371 : {
1372 : /* Failed to get lock, so release interrupt holdoff */
1373 1364 : RESUME_INTERRUPTS();
1374 :
1375 : LOG_LWDEBUG("LWLockConditionalAcquire", lock, "failed");
1376 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL_ENABLED())
1377 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL(T_NAME(lock), mode);
1378 : }
1379 : else
1380 : {
1381 : /* Add lock to list of locks held by this backend */
1382 1417727 : held_lwlocks[num_held_lwlocks].lock = lock;
1383 1417727 : held_lwlocks[num_held_lwlocks++].mode = mode;
1384 : if (TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_ENABLED())
1385 : TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE(T_NAME(lock), mode);
1386 : }
1387 1419091 : return !mustwait;
1388 : }
1389 :
1390 : /*
1391 : * LWLockAcquireOrWait - Acquire lock, or wait until it's free
1392 : *
1393 : * The semantics of this function are a bit funky. If the lock is currently
1394 : * free, it is acquired in the given mode, and the function returns true. If
1395 : * the lock isn't immediately free, the function waits until it is released
1396 : * and returns false, but does not acquire the lock.
1397 : *
1398 : * This is currently used for WALWriteLock: when a backend flushes the WAL,
1399 : * holding WALWriteLock, it can flush the commit records of many other
1400 : * backends as a side-effect. Those other backends need to wait until the
1401 : * flush finishes, but don't need to acquire the lock anymore. They can just
1402 : * wake up, observe that their records have already been flushed, and return.
1403 : */
1404 : bool
1405 158579 : LWLockAcquireOrWait(LWLock *lock, LWLockMode mode)
1406 : {
1407 158579 : PGPROC *proc = MyProc;
1408 : bool mustwait;
1409 158579 : int extraWaits = 0;
1410 : #ifdef LWLOCK_STATS
1411 : lwlock_stats *lwstats;
1412 :
1413 : lwstats = get_lwlock_stats_entry(lock);
1414 : #endif
1415 :
1416 : Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1417 :
1418 : PRINT_LWDEBUG("LWLockAcquireOrWait", lock, mode);
1419 :
1420 : /* Ensure we will have room to remember the lock */
1421 158579 : if (num_held_lwlocks >= MAX_SIMUL_LWLOCKS)
1422 0 : elog(ERROR, "too many LWLocks taken");
1423 :
1424 : /*
1425 : * Lock out cancel/die interrupts until we exit the code section protected
1426 : * by the LWLock. This ensures that interrupts will not interfere with
1427 : * manipulations of data structures in shared memory.
1428 : */
1429 158579 : HOLD_INTERRUPTS();
1430 :
1431 : /*
1432 : * NB: We're using nearly the same twice-in-a-row lock acquisition
1433 : * protocol as LWLockAcquire(). Check its comments for details.
1434 : */
1435 158579 : mustwait = LWLockAttemptLock(lock, mode);
1436 :
1437 158579 : if (mustwait)
1438 : {
1439 3460 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1440 :
1441 3460 : mustwait = LWLockAttemptLock(lock, mode);
1442 :
1443 3460 : if (mustwait)
1444 : {
1445 : /*
1446 : * Wait until awakened. Like in LWLockAcquire, be prepared for
1447 : * bogus wakeups.
1448 : */
1449 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "waiting");
1450 :
1451 : #ifdef LWLOCK_STATS
1452 : lwstats->block_count++;
1453 : #endif
1454 :
1455 3368 : LWLockReportWaitStart(lock);
1456 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1457 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1458 :
1459 : for (;;)
1460 : {
1461 3368 : PGSemaphoreLock(proc->sem);
1462 3368 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1463 3368 : break;
1464 0 : extraWaits++;
1465 : }
1466 :
1467 : #ifdef LOCK_DEBUG
1468 : {
1469 : /* not waiting anymore */
1470 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1471 :
1472 : Assert(nwaiters < MAX_BACKENDS);
1473 : }
1474 : #endif
1475 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1476 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1477 3368 : LWLockReportWaitEnd();
1478 :
1479 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "awakened");
1480 : }
1481 : else
1482 : {
1483 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "acquired, undoing queue");
1484 :
1485 : /*
1486 : * Got lock in the second attempt, undo queueing. We need to treat
1487 : * this as having successfully acquired the lock, otherwise we'd
1488 : * not necessarily wake up people we've prevented from acquiring
1489 : * the lock.
1490 : */
1491 92 : LWLockDequeueSelf(lock);
1492 : }
1493 : }
1494 :
1495 : /*
1496 : * Fix the process wait semaphore's count for any absorbed wakeups.
1497 : */
1498 158579 : while (extraWaits-- > 0)
1499 0 : PGSemaphoreUnlock(proc->sem);
1500 :
1501 158579 : if (mustwait)
1502 : {
1503 : /* Failed to get lock, so release interrupt holdoff */
1504 3368 : RESUME_INTERRUPTS();
1505 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "failed");
1506 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL_ENABLED())
1507 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL(T_NAME(lock), mode);
1508 : }
1509 : else
1510 : {
1511 : LOG_LWDEBUG("LWLockAcquireOrWait", lock, "succeeded");
1512 : /* Add lock to list of locks held by this backend */
1513 155211 : held_lwlocks[num_held_lwlocks].lock = lock;
1514 155211 : held_lwlocks[num_held_lwlocks++].mode = mode;
1515 : if (TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_ENABLED())
1516 : TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT(T_NAME(lock), mode);
1517 : }
1518 :
1519 158579 : return !mustwait;
1520 : }
1521 :
1522 : /*
1523 : * Does the lwlock in its current state need to wait for the variable value to
1524 : * change?
1525 : *
1526 : * If we don't need to wait, and it's because the value of the variable has
1527 : * changed, store the current value in newval.
1528 : *
1529 : * *result is set to true if the lock was free, and false otherwise.
1530 : */
1531 : static bool
1532 4267137 : LWLockConflictsWithVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1533 : uint64 *newval, bool *result)
1534 : {
1535 : bool mustwait;
1536 : uint64 value;
1537 :
1538 : /*
1539 : * Test first to see if it the slot is free right now.
1540 : *
1541 : * XXX: the unique caller of this routine, WaitXLogInsertionsToFinish()
1542 : * via LWLockWaitForVar(), uses an implied barrier with a spinlock before
1543 : * this, so we don't need a memory barrier here as far as the current
1544 : * usage is concerned. But that might not be safe in general.
1545 : */
1546 4267137 : mustwait = (pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE) != 0;
1547 :
1548 4267137 : if (!mustwait)
1549 : {
1550 2813688 : *result = true;
1551 2813688 : return false;
1552 : }
1553 :
1554 1453449 : *result = false;
1555 :
1556 : /*
1557 : * Reading this value atomically is safe even on platforms where uint64
1558 : * cannot be read without observing a torn value.
1559 : */
1560 1453449 : value = pg_atomic_read_u64(valptr);
1561 :
1562 1453449 : if (value != oldval)
1563 : {
1564 1237875 : mustwait = false;
1565 1237875 : *newval = value;
1566 : }
1567 : else
1568 : {
1569 215574 : mustwait = true;
1570 : }
1571 :
1572 1453449 : return mustwait;
1573 : }
1574 :
1575 : /*
1576 : * LWLockWaitForVar - Wait until lock is free, or a variable is updated.
1577 : *
1578 : * If the lock is held and *valptr equals oldval, waits until the lock is
1579 : * either freed, or the lock holder updates *valptr by calling
1580 : * LWLockUpdateVar. If the lock is free on exit (immediately or after
1581 : * waiting), returns true. If the lock is still held, but *valptr no longer
1582 : * matches oldval, returns false and sets *newval to the current value in
1583 : * *valptr.
1584 : *
1585 : * Note: this function ignores shared lock holders; if the lock is held
1586 : * in shared mode, returns 'true'.
1587 : *
1588 : * Be aware that LWLockConflictsWithVar() does not include a memory barrier,
1589 : * hence the caller of this function may want to rely on an explicit barrier or
1590 : * an implied barrier via spinlock or LWLock to avoid memory ordering issues.
1591 : */
1592 : bool
1593 4051563 : LWLockWaitForVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 oldval,
1594 : uint64 *newval)
1595 : {
1596 4051563 : PGPROC *proc = MyProc;
1597 4051563 : int extraWaits = 0;
1598 4051563 : bool result = false;
1599 : #ifdef LWLOCK_STATS
1600 : lwlock_stats *lwstats;
1601 :
1602 : lwstats = get_lwlock_stats_entry(lock);
1603 : #endif
1604 :
1605 : PRINT_LWDEBUG("LWLockWaitForVar", lock, LW_WAIT_UNTIL_FREE);
1606 :
1607 : /*
1608 : * Lock out cancel/die interrupts while we sleep on the lock. There is no
1609 : * cleanup mechanism to remove us from the wait queue if we got
1610 : * interrupted.
1611 : */
1612 4051563 : HOLD_INTERRUPTS();
1613 :
1614 : /*
1615 : * Loop here to check the lock's status after each time we are signaled.
1616 : */
1617 : for (;;)
1618 107050 : {
1619 : bool mustwait;
1620 :
1621 4158613 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1622 : &result);
1623 :
1624 4158613 : if (!mustwait)
1625 4050089 : break; /* the lock was free or value didn't match */
1626 :
1627 : /*
1628 : * Add myself to wait queue. Note that this is racy, somebody else
1629 : * could wakeup before we're finished queuing. NB: We're using nearly
1630 : * the same twice-in-a-row lock acquisition protocol as
1631 : * LWLockAcquire(). Check its comments for details. The only
1632 : * difference is that we also have to check the variable's values when
1633 : * checking the state of the lock.
1634 : */
1635 108524 : LWLockQueueSelf(lock, LW_WAIT_UNTIL_FREE);
1636 :
1637 : /*
1638 : * Clear LW_FLAG_WAKE_IN_PROGRESS flag, to make sure we get woken up
1639 : * as soon as the lock is released.
1640 : */
1641 108524 : pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_WAKE_IN_PROGRESS);
1642 :
1643 : /*
1644 : * We're now guaranteed to be woken up if necessary. Recheck the lock
1645 : * and variables state.
1646 : */
1647 108524 : mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1648 : &result);
1649 :
1650 : /* Ok, no conflict after we queued ourselves. Undo queueing. */
1651 108524 : if (!mustwait)
1652 : {
1653 : LOG_LWDEBUG("LWLockWaitForVar", lock, "free, undoing queue");
1654 :
1655 1474 : LWLockDequeueSelf(lock);
1656 1474 : break;
1657 : }
1658 :
1659 : /*
1660 : * Wait until awakened.
1661 : *
1662 : * It is possible that we get awakened for a reason other than being
1663 : * signaled by LWLockRelease. If so, loop back and wait again. Once
1664 : * we've gotten the LWLock, re-increment the sema by the number of
1665 : * additional signals received.
1666 : */
1667 : LOG_LWDEBUG("LWLockWaitForVar", lock, "waiting");
1668 :
1669 : #ifdef LWLOCK_STATS
1670 : lwstats->block_count++;
1671 : #endif
1672 :
1673 107050 : LWLockReportWaitStart(lock);
1674 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_START_ENABLED())
1675 : TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), LW_EXCLUSIVE);
1676 :
1677 : for (;;)
1678 : {
1679 107050 : PGSemaphoreLock(proc->sem);
1680 107050 : if (proc->lwWaiting == LW_WS_NOT_WAITING)
1681 107050 : break;
1682 0 : extraWaits++;
1683 : }
1684 :
1685 : #ifdef LOCK_DEBUG
1686 : {
1687 : /* not waiting anymore */
1688 : uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1689 :
1690 : Assert(nwaiters < MAX_BACKENDS);
1691 : }
1692 : #endif
1693 :
1694 : if (TRACE_POSTGRESQL_LWLOCK_WAIT_DONE_ENABLED())
1695 : TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), LW_EXCLUSIVE);
1696 107050 : LWLockReportWaitEnd();
1697 :
1698 : LOG_LWDEBUG("LWLockWaitForVar", lock, "awakened");
1699 :
1700 : /* Now loop back and check the status of the lock again. */
1701 : }
1702 :
1703 : /*
1704 : * Fix the process wait semaphore's count for any absorbed wakeups.
1705 : */
1706 4051563 : while (extraWaits-- > 0)
1707 0 : PGSemaphoreUnlock(proc->sem);
1708 :
1709 : /*
1710 : * Now okay to allow cancel/die interrupts.
1711 : */
1712 4051563 : RESUME_INTERRUPTS();
1713 :
1714 4051563 : return result;
1715 : }
1716 :
1717 :
1718 : /*
1719 : * LWLockUpdateVar - Update a variable and wake up waiters atomically
1720 : *
1721 : * Sets *valptr to 'val', and wakes up all processes waiting for us with
1722 : * LWLockWaitForVar(). It first sets the value atomically and then wakes up
1723 : * waiting processes so that any process calling LWLockWaitForVar() on the same
1724 : * lock is guaranteed to see the new value, and act accordingly.
1725 : *
1726 : * The caller must be holding the lock in exclusive mode.
1727 : */
1728 : void
1729 2636179 : LWLockUpdateVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1730 : {
1731 : proclist_head wakeup;
1732 : proclist_mutable_iter iter;
1733 :
1734 : PRINT_LWDEBUG("LWLockUpdateVar", lock, LW_EXCLUSIVE);
1735 :
1736 : /*
1737 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1738 : * that the variable is updated before waking up waiters.
1739 : */
1740 2636179 : pg_atomic_exchange_u64(valptr, val);
1741 :
1742 2636179 : proclist_init(&wakeup);
1743 :
1744 2636179 : LWLockWaitListLock(lock);
1745 :
1746 : Assert(pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE);
1747 :
1748 : /*
1749 : * See if there are any LW_WAIT_UNTIL_FREE waiters that need to be woken
1750 : * up. They are always in the front of the queue.
1751 : */
1752 2637163 : proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
1753 : {
1754 39759 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1755 :
1756 39759 : if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
1757 38775 : break;
1758 :
1759 984 : proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
1760 984 : proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
1761 :
1762 : /* see LWLockWakeup() */
1763 : Assert(waiter->lwWaiting == LW_WS_WAITING);
1764 984 : waiter->lwWaiting = LW_WS_PENDING_WAKEUP;
1765 : }
1766 :
1767 : /* We are done updating shared state of the lock itself. */
1768 2636179 : LWLockWaitListUnlock(lock);
1769 :
1770 : /*
1771 : * Awaken any waiters I removed from the queue.
1772 : */
1773 2637163 : proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1774 : {
1775 984 : PGPROC *waiter = GetPGProcByNumber(iter.cur);
1776 :
1777 984 : proclist_delete(&wakeup, iter.cur, lwWaitLink);
1778 : /* check comment in LWLockWakeup() about this barrier */
1779 984 : pg_write_barrier();
1780 984 : waiter->lwWaiting = LW_WS_NOT_WAITING;
1781 984 : PGSemaphoreUnlock(waiter->sem);
1782 : }
1783 2636179 : }
1784 :
1785 :
1786 : /*
1787 : * LWLockRelease - release a previously acquired lock
1788 : *
1789 : * NB: This will leave lock->owner pointing to the current backend (if
1790 : * LOCK_DEBUG is set). This is somewhat intentional, as it makes it easier to
1791 : * debug cases of missing wakeups during lock release.
1792 : */
1793 : void
1794 332970095 : LWLockRelease(LWLock *lock)
1795 : {
1796 : LWLockMode mode;
1797 : uint32 oldstate;
1798 : bool check_waiters;
1799 : int i;
1800 :
1801 : /*
1802 : * Remove lock from list of locks held. Usually, but not always, it will
1803 : * be the latest-acquired lock; so search array backwards.
1804 : */
1805 376095780 : for (i = num_held_lwlocks; --i >= 0;)
1806 376095780 : if (lock == held_lwlocks[i].lock)
1807 332970095 : break;
1808 :
1809 332970095 : if (i < 0)
1810 0 : elog(ERROR, "lock %s is not held", T_NAME(lock));
1811 :
1812 332970095 : mode = held_lwlocks[i].mode;
1813 :
1814 332970095 : num_held_lwlocks--;
1815 376095780 : for (; i < num_held_lwlocks; i++)
1816 43125685 : held_lwlocks[i] = held_lwlocks[i + 1];
1817 :
1818 : PRINT_LWDEBUG("LWLockRelease", lock, mode);
1819 :
1820 : /*
1821 : * Release my hold on lock, after that it can immediately be acquired by
1822 : * others, even if we still have to wakeup other waiters.
1823 : */
1824 332970095 : if (mode == LW_EXCLUSIVE)
1825 230280807 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_EXCLUSIVE);
1826 : else
1827 102689288 : oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_SHARED);
1828 :
1829 : /* nobody else can have that kind of lock */
1830 : Assert(!(oldstate & LW_VAL_EXCLUSIVE));
1831 :
1832 : if (TRACE_POSTGRESQL_LWLOCK_RELEASE_ENABLED())
1833 : TRACE_POSTGRESQL_LWLOCK_RELEASE(T_NAME(lock));
1834 :
1835 : /*
1836 : * Check if we're still waiting for backends to get scheduled, if so,
1837 : * don't wake them up again.
1838 : */
1839 332970095 : if ((oldstate & LW_FLAG_HAS_WAITERS) &&
1840 6077976 : !(oldstate & LW_FLAG_WAKE_IN_PROGRESS) &&
1841 5173024 : (oldstate & LW_LOCK_MASK) == 0)
1842 5170194 : check_waiters = true;
1843 : else
1844 327799901 : check_waiters = false;
1845 :
1846 : /*
1847 : * As waking up waiters requires the spinlock to be acquired, only do so
1848 : * if necessary.
1849 : */
1850 332970095 : if (check_waiters)
1851 : {
1852 : /* XXX: remove before commit? */
1853 : LOG_LWDEBUG("LWLockRelease", lock, "releasing waiters");
1854 5170194 : LWLockWakeup(lock);
1855 : }
1856 :
1857 : /*
1858 : * Now okay to allow cancel/die interrupts.
1859 : */
1860 332970095 : RESUME_INTERRUPTS();
1861 332970095 : }
1862 :
1863 : /*
1864 : * LWLockReleaseClearVar - release a previously acquired lock, reset variable
1865 : */
1866 : void
1867 18436352 : LWLockReleaseClearVar(LWLock *lock, pg_atomic_uint64 *valptr, uint64 val)
1868 : {
1869 : /*
1870 : * Note that pg_atomic_exchange_u64 is a full barrier, so we're guaranteed
1871 : * that the variable is updated before releasing the lock.
1872 : */
1873 18436352 : pg_atomic_exchange_u64(valptr, val);
1874 :
1875 18436352 : LWLockRelease(lock);
1876 18436352 : }
1877 :
1878 :
1879 : /*
1880 : * LWLockReleaseAll - release all currently-held locks
1881 : *
1882 : * Used to clean up after ereport(ERROR). An important difference between this
1883 : * function and retail LWLockRelease calls is that InterruptHoldoffCount is
1884 : * unchanged by this operation. This is necessary since InterruptHoldoffCount
1885 : * has been set to an appropriate level earlier in error recovery. We could
1886 : * decrement it below zero if we allow it to drop for each released lock!
1887 : *
1888 : * Note that this function must be safe to call even before the LWLock
1889 : * subsystem has been initialized (e.g., during early startup failures).
1890 : * In that case, num_held_lwlocks will be 0 and we do nothing.
1891 : */
1892 : void
1893 120936 : LWLockReleaseAll(void)
1894 : {
1895 121031 : while (num_held_lwlocks > 0)
1896 : {
1897 95 : HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
1898 :
1899 95 : LWLockRelease(held_lwlocks[num_held_lwlocks - 1].lock);
1900 : }
1901 :
1902 : Assert(num_held_lwlocks == 0);
1903 120936 : }
1904 :
1905 :
1906 : /*
1907 : * LWLockHeldByMe - test whether my process holds a lock in any mode
1908 : *
1909 : * This is meant as debug support only.
1910 : */
1911 : bool
1912 0 : LWLockHeldByMe(LWLock *lock)
1913 : {
1914 : int i;
1915 :
1916 0 : for (i = 0; i < num_held_lwlocks; i++)
1917 : {
1918 0 : if (held_lwlocks[i].lock == lock)
1919 0 : return true;
1920 : }
1921 0 : return false;
1922 : }
1923 :
1924 : /*
1925 : * LWLockAnyHeldByMe - test whether my process holds any of an array of locks
1926 : *
1927 : * This is meant as debug support only.
1928 : */
1929 : bool
1930 0 : LWLockAnyHeldByMe(LWLock *lock, int nlocks, size_t stride)
1931 : {
1932 : char *held_lock_addr;
1933 : char *begin;
1934 : char *end;
1935 : int i;
1936 :
1937 0 : begin = (char *) lock;
1938 0 : end = begin + nlocks * stride;
1939 0 : for (i = 0; i < num_held_lwlocks; i++)
1940 : {
1941 0 : held_lock_addr = (char *) held_lwlocks[i].lock;
1942 0 : if (held_lock_addr >= begin &&
1943 0 : held_lock_addr < end &&
1944 0 : (held_lock_addr - begin) % stride == 0)
1945 0 : return true;
1946 : }
1947 0 : return false;
1948 : }
1949 :
1950 : /*
1951 : * LWLockHeldByMeInMode - test whether my process holds a lock in given mode
1952 : *
1953 : * This is meant as debug support only.
1954 : */
1955 : bool
1956 0 : LWLockHeldByMeInMode(LWLock *lock, LWLockMode mode)
1957 : {
1958 : int i;
1959 :
1960 0 : for (i = 0; i < num_held_lwlocks; i++)
1961 : {
1962 0 : if (held_lwlocks[i].lock == lock && held_lwlocks[i].mode == mode)
1963 0 : return true;
1964 : }
1965 0 : return false;
1966 : }
|
