PostgreSQL Source Code  git master
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros
lwlock.c
Go to the documentation of this file.
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) 1996-2017, 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 it's
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 "pgstat.h"
81 #include "pg_trace.h"
82 #include "postmaster/postmaster.h"
83 #include "replication/slot.h"
84 #include "storage/ipc.h"
85 #include "storage/predicate.h"
86 #include "storage/proc.h"
87 #include "storage/proclist.h"
88 #include "storage/spin.h"
89 #include "utils/memutils.h"
90 
91 #ifdef LWLOCK_STATS
92 #include "utils/hsearch.h"
93 #endif
94 
95 
96 /* We use the ShmemLock spinlock to protect LWLockCounter */
97 extern slock_t *ShmemLock;
98 
99 #define LW_FLAG_HAS_WAITERS ((uint32) 1 << 30)
100 #define LW_FLAG_RELEASE_OK ((uint32) 1 << 29)
101 #define LW_FLAG_LOCKED ((uint32) 1 << 28)
102 
103 #define LW_VAL_EXCLUSIVE ((uint32) 1 << 24)
104 #define LW_VAL_SHARED 1
105 
106 #define LW_LOCK_MASK ((uint32) ((1 << 25)-1))
107 /* Must be greater than MAX_BACKENDS - which is 2^23-1, so we're fine. */
108 #define LW_SHARED_MASK ((uint32) ((1 << 24)-1))
109 
110 /*
111  * This is indexed by tranche ID and stores the names of all tranches known
112  * to the current backend.
113  */
114 static char **LWLockTrancheArray = NULL;
115 static int LWLockTranchesAllocated = 0;
116 
117 #define T_NAME(lock) \
118  (LWLockTrancheArray[(lock)->tranche])
119 
120 /*
121  * This points to the main array of LWLocks in shared memory. Backends inherit
122  * the pointer by fork from the postmaster (except in the EXEC_BACKEND case,
123  * where we have special measures to pass it down).
124  */
126 
127 /*
128  * We use this structure to keep track of locked LWLocks for release
129  * during error recovery. Normally, only a few will be held at once, but
130  * occasionally the number can be much higher; for example, the pg_buffercache
131  * extension locks all buffer partitions simultaneously.
132  */
133 #define MAX_SIMUL_LWLOCKS 200
134 
135 /* struct representing the LWLocks we're holding */
136 typedef struct LWLockHandle
137 {
140 } LWLockHandle;
141 
142 static int num_held_lwlocks = 0;
144 
145 /* struct representing the LWLock tranche request for named tranche */
147 {
151 
155 
157 
158 static bool lock_named_request_allowed = true;
159 
160 static void InitializeLWLocks(void);
161 static void RegisterLWLockTranches(void);
162 
163 static inline void LWLockReportWaitStart(LWLock *lock);
164 static inline void LWLockReportWaitEnd(void);
165 
166 #ifdef LWLOCK_STATS
167 typedef struct lwlock_stats_key
168 {
169  int tranche;
170  void *instance;
171 } lwlock_stats_key;
172 
173 typedef struct lwlock_stats
174 {
175  lwlock_stats_key key;
176  int sh_acquire_count;
177  int ex_acquire_count;
178  int block_count;
179  int dequeue_self_count;
180  int spin_delay_count;
181 } lwlock_stats;
182 
183 static HTAB *lwlock_stats_htab;
184 static lwlock_stats lwlock_stats_dummy;
185 #endif
186 
187 #ifdef LOCK_DEBUG
188 bool Trace_lwlocks = false;
189 
190 inline static void
191 PRINT_LWDEBUG(const char *where, LWLock *lock, LWLockMode mode)
192 {
193  /* hide statement & context here, otherwise the log is just too verbose */
194  if (Trace_lwlocks)
195  {
197 
198  ereport(LOG,
199  (errhidestmt(true),
200  errhidecontext(true),
201  errmsg_internal("%d: %s(%s %p): excl %u shared %u haswaiters %u waiters %u rOK %d",
202  MyProcPid,
203  where, T_NAME(lock), lock,
204  (state & LW_VAL_EXCLUSIVE) != 0,
205  state & LW_SHARED_MASK,
206  (state & LW_FLAG_HAS_WAITERS) != 0,
207  pg_atomic_read_u32(&lock->nwaiters),
208  (state & LW_FLAG_RELEASE_OK) != 0)));
209  }
210 }
211 
212 inline static void
213 LOG_LWDEBUG(const char *where, LWLock *lock, const char *msg)
214 {
215  /* hide statement & context here, otherwise the log is just too verbose */
216  if (Trace_lwlocks)
217  {
218  ereport(LOG,
219  (errhidestmt(true),
220  errhidecontext(true),
221  errmsg_internal("%s(%s %p): %s", where,
222  T_NAME(lock), lock, msg)));
223  }
224 }
225 
226 #else /* not LOCK_DEBUG */
227 #define PRINT_LWDEBUG(a,b,c) ((void)0)
228 #define LOG_LWDEBUG(a,b,c) ((void)0)
229 #endif /* LOCK_DEBUG */
230 
231 #ifdef LWLOCK_STATS
232 
233 static void init_lwlock_stats(void);
234 static void print_lwlock_stats(int code, Datum arg);
235 static lwlock_stats * get_lwlock_stats_entry(LWLock *lockid);
236 
237 static void
238 init_lwlock_stats(void)
239 {
240  HASHCTL ctl;
241  static MemoryContext lwlock_stats_cxt = NULL;
242  static bool exit_registered = false;
243 
244  if (lwlock_stats_cxt != NULL)
245  MemoryContextDelete(lwlock_stats_cxt);
246 
247  /*
248  * The LWLock stats will be updated within a critical section, which
249  * requires allocating new hash entries. Allocations within a critical
250  * section are normally not allowed because running out of memory would
251  * lead to a PANIC, but LWLOCK_STATS is debugging code that's not normally
252  * turned on in production, so that's an acceptable risk. The hash entries
253  * are small, so the risk of running out of memory is minimal in practice.
254  */
255  lwlock_stats_cxt = AllocSetContextCreate(TopMemoryContext,
256  "LWLock stats",
258  MemoryContextAllowInCriticalSection(lwlock_stats_cxt, true);
259 
260  MemSet(&ctl, 0, sizeof(ctl));
261  ctl.keysize = sizeof(lwlock_stats_key);
262  ctl.entrysize = sizeof(lwlock_stats);
263  ctl.hcxt = lwlock_stats_cxt;
264  lwlock_stats_htab = hash_create("lwlock stats", 16384, &ctl,
266  if (!exit_registered)
267  {
268  on_shmem_exit(print_lwlock_stats, 0);
269  exit_registered = true;
270  }
271 }
272 
273 static void
274 print_lwlock_stats(int code, Datum arg)
275 {
276  HASH_SEQ_STATUS scan;
277  lwlock_stats *lwstats;
278 
279  hash_seq_init(&scan, lwlock_stats_htab);
280 
281  /* Grab an LWLock to keep different backends from mixing reports */
282  LWLockAcquire(&MainLWLockArray[0].lock, LW_EXCLUSIVE);
283 
284  while ((lwstats = (lwlock_stats *) hash_seq_search(&scan)) != NULL)
285  {
286  fprintf(stderr,
287  "PID %d lwlock %s %p: shacq %u exacq %u blk %u spindelay %u dequeue self %u\n",
288  MyProcPid, LWLockTrancheArray[lwstats->key.tranche],
289  lwstats->key.instance, lwstats->sh_acquire_count,
290  lwstats->ex_acquire_count, lwstats->block_count,
291  lwstats->spin_delay_count, lwstats->dequeue_self_count);
292  }
293 
294  LWLockRelease(&MainLWLockArray[0].lock);
295 }
296 
297 static lwlock_stats *
298 get_lwlock_stats_entry(LWLock *lock)
299 {
300  lwlock_stats_key key;
301  lwlock_stats *lwstats;
302  bool found;
303 
304  /*
305  * During shared memory initialization, the hash table doesn't exist yet.
306  * Stats of that phase aren't very interesting, so just collect operations
307  * on all locks in a single dummy entry.
308  */
309  if (lwlock_stats_htab == NULL)
310  return &lwlock_stats_dummy;
311 
312  /* Fetch or create the entry. */
313  key.tranche = lock->tranche;
314  key.instance = lock;
315  lwstats = hash_search(lwlock_stats_htab, &key, HASH_ENTER, &found);
316  if (!found)
317  {
318  lwstats->sh_acquire_count = 0;
319  lwstats->ex_acquire_count = 0;
320  lwstats->block_count = 0;
321  lwstats->dequeue_self_count = 0;
322  lwstats->spin_delay_count = 0;
323  }
324  return lwstats;
325 }
326 #endif /* LWLOCK_STATS */
327 
328 
329 /*
330  * Compute number of LWLocks required by named tranches. These will be
331  * allocated in the main array.
332  */
333 static int
335 {
336  int numLocks = 0;
337  int i;
338 
339  for (i = 0; i < NamedLWLockTrancheRequests; i++)
340  numLocks += NamedLWLockTrancheRequestArray[i].num_lwlocks;
341 
342  return numLocks;
343 }
344 
345 /*
346  * Compute shmem space needed for LWLocks and named tranches.
347  */
348 Size
350 {
351  Size size;
352  int i;
353  int numLocks = NUM_FIXED_LWLOCKS;
354 
355  numLocks += NumLWLocksByNamedTranches();
356 
357  /* Space for the LWLock array. */
358  size = mul_size(numLocks, sizeof(LWLockPadded));
359 
360  /* Space for dynamic allocation counter, plus room for alignment. */
361  size = add_size(size, sizeof(int) + LWLOCK_PADDED_SIZE);
362 
363  /* space for named tranches. */
365 
366  /* space for name of each tranche. */
367  for (i = 0; i < NamedLWLockTrancheRequests; i++)
368  size = add_size(size, strlen(NamedLWLockTrancheRequestArray[i].tranche_name) + 1);
369 
370  /* Disallow named LWLocks' requests after startup */
372 
373  return size;
374 }
375 
376 /*
377  * Allocate shmem space for the main LWLock array and all tranches and
378  * initialize it. We also register all the LWLock tranches here.
379  */
380 void
382 {
384  "MAX_BACKENDS too big for lwlock.c");
385 
387  sizeof(LWLock) <= LWLOCK_PADDED_SIZE,
388  "Miscalculated LWLock padding");
389 
390  if (!IsUnderPostmaster)
391  {
392  Size spaceLocks = LWLockShmemSize();
393  int *LWLockCounter;
394  char *ptr;
395 
396  /* Allocate space */
397  ptr = (char *) ShmemAlloc(spaceLocks);
398 
399  /* Leave room for dynamic allocation of tranches */
400  ptr += sizeof(int);
401 
402  /* Ensure desired alignment of LWLock array */
403  ptr += LWLOCK_PADDED_SIZE - ((uintptr_t) ptr) % LWLOCK_PADDED_SIZE;
404 
405  MainLWLockArray = (LWLockPadded *) ptr;
406 
407  /*
408  * Initialize the dynamic-allocation counter for tranches, which is
409  * stored just before the first LWLock.
410  */
411  LWLockCounter = (int *) ((char *) MainLWLockArray - sizeof(int));
412  *LWLockCounter = LWTRANCHE_FIRST_USER_DEFINED;
413 
414  /* Initialize all LWLocks */
416  }
417 
418  /* Register all LWLock tranches */
420 }
421 
422 /*
423  * Initialize LWLocks that are fixed and those belonging to named tranches.
424  */
425 static void
427 {
428  int numNamedLocks = NumLWLocksByNamedTranches();
429  int id;
430  int i;
431  int j;
432  LWLockPadded *lock;
433 
434  /* Initialize all individual LWLocks in main array */
435  for (id = 0, lock = MainLWLockArray; id < NUM_INDIVIDUAL_LWLOCKS; id++, lock++)
436  LWLockInitialize(&lock->lock, id);
437 
438  /* Initialize buffer mapping LWLocks in main array */
439  lock = MainLWLockArray + NUM_INDIVIDUAL_LWLOCKS;
440  for (id = 0; id < NUM_BUFFER_PARTITIONS; id++, lock++)
442 
443  /* Initialize lmgrs' LWLocks in main array */
444  lock = MainLWLockArray + NUM_INDIVIDUAL_LWLOCKS + NUM_BUFFER_PARTITIONS;
445  for (id = 0; id < NUM_LOCK_PARTITIONS; id++, lock++)
447 
448  /* Initialize predicate lmgrs' LWLocks in main array */
449  lock = MainLWLockArray + NUM_INDIVIDUAL_LWLOCKS +
450  NUM_BUFFER_PARTITIONS + NUM_LOCK_PARTITIONS;
451  for (id = 0; id < NUM_PREDICATELOCK_PARTITIONS; id++, lock++)
453 
454  /* Initialize named tranches. */
456  {
457  char *trancheNames;
458 
459  NamedLWLockTrancheArray = (NamedLWLockTranche *)
460  &MainLWLockArray[NUM_FIXED_LWLOCKS + numNamedLocks];
461 
462  trancheNames = (char *) NamedLWLockTrancheArray +
464  lock = &MainLWLockArray[NUM_FIXED_LWLOCKS];
465 
466  for (i = 0; i < NamedLWLockTrancheRequests; i++)
467  {
468  NamedLWLockTrancheRequest *request;
469  NamedLWLockTranche *tranche;
470  char *name;
471 
472  request = &NamedLWLockTrancheRequestArray[i];
473  tranche = &NamedLWLockTrancheArray[i];
474 
475  name = trancheNames;
476  trancheNames += strlen(request->tranche_name) + 1;
477  strcpy(name, request->tranche_name);
478  tranche->trancheId = LWLockNewTrancheId();
479  tranche->trancheName = name;
480 
481  for (j = 0; j < request->num_lwlocks; j++, lock++)
482  LWLockInitialize(&lock->lock, tranche->trancheId);
483  }
484  }
485 }
486 
487 /*
488  * Register named tranches and tranches for fixed LWLocks.
489  */
490 static void
492 {
493  int i;
494 
495  if (LWLockTrancheArray == NULL)
496  {
498  LWLockTrancheArray = (char **)
500  LWLockTranchesAllocated * sizeof(char *));
502  }
503 
504  for (i = 0; i < NUM_INDIVIDUAL_LWLOCKS; ++i)
506 
510  "predicate_lock_manager");
512  "parallel_query_dsa");
514  "session_dsa");
516  "session_record_table");
518  "session_typmod_table");
520 
521  /* Register named tranches. */
522  for (i = 0; i < NamedLWLockTrancheRequests; i++)
523  LWLockRegisterTranche(NamedLWLockTrancheArray[i].trancheId,
524  NamedLWLockTrancheArray[i].trancheName);
525 }
526 
527 /*
528  * InitLWLockAccess - initialize backend-local state needed to hold LWLocks
529  */
530 void
532 {
533 #ifdef LWLOCK_STATS
534  init_lwlock_stats();
535 #endif
536 }
537 
538 /*
539  * GetNamedLWLockTranche - returns the base address of LWLock from the
540  * specified tranche.
541  *
542  * Caller needs to retrieve the requested number of LWLocks starting from
543  * the base lock address returned by this API. This can be used for
544  * tranches that are requested by using RequestNamedLWLockTranche() API.
545  */
546 LWLockPadded *
547 GetNamedLWLockTranche(const char *tranche_name)
548 {
549  int lock_pos;
550  int i;
551 
552  /*
553  * Obtain the position of base address of LWLock belonging to requested
554  * tranche_name in MainLWLockArray. LWLocks for named tranches are placed
555  * in MainLWLockArray after fixed locks.
556  */
557  lock_pos = NUM_FIXED_LWLOCKS;
558  for (i = 0; i < NamedLWLockTrancheRequests; i++)
559  {
560  if (strcmp(NamedLWLockTrancheRequestArray[i].tranche_name,
561  tranche_name) == 0)
562  return &MainLWLockArray[lock_pos];
563 
564  lock_pos += NamedLWLockTrancheRequestArray[i].num_lwlocks;
565  }
566 
567  if (i >= NamedLWLockTrancheRequests)
568  elog(ERROR, "requested tranche is not registered");
569 
570  /* just to keep compiler quiet */
571  return NULL;
572 }
573 
574 /*
575  * Allocate a new tranche ID.
576  */
577 int
579 {
580  int result;
581  int *LWLockCounter;
582 
583  LWLockCounter = (int *) ((char *) MainLWLockArray - sizeof(int));
585  result = (*LWLockCounter)++;
587 
588  return result;
589 }
590 
591 /*
592  * Register a tranche ID in the lookup table for the current process. This
593  * routine will save a pointer to the tranche name passed as an argument,
594  * so the name should be allocated in a backend-lifetime context
595  * (TopMemoryContext, static variable, or similar).
596  */
597 void
598 LWLockRegisterTranche(int tranche_id, char *tranche_name)
599 {
600  Assert(LWLockTrancheArray != NULL);
601 
602  if (tranche_id >= LWLockTranchesAllocated)
603  {
605  int j = LWLockTranchesAllocated;
606 
607  while (i <= tranche_id)
608  i *= 2;
609 
610  LWLockTrancheArray = (char **)
611  repalloc(LWLockTrancheArray, i * sizeof(char *));
613  while (j < LWLockTranchesAllocated)
614  LWLockTrancheArray[j++] = NULL;
615  }
616 
617  LWLockTrancheArray[tranche_id] = tranche_name;
618 }
619 
620 /*
621  * RequestNamedLWLockTranche
622  * Request that extra LWLocks be allocated during postmaster
623  * startup.
624  *
625  * This is only useful for extensions if called from the _PG_init hook
626  * of a library that is loaded into the postmaster via
627  * shared_preload_libraries. Once shared memory has been allocated, calls
628  * will be ignored. (We could raise an error, but it seems better to make
629  * it a no-op, so that libraries containing such calls can be reloaded if
630  * needed.)
631  */
632 void
633 RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
634 {
635  NamedLWLockTrancheRequest *request;
636 
638  return; /* too late */
639 
640  if (NamedLWLockTrancheRequestArray == NULL)
641  {
643  NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
646  * sizeof(NamedLWLockTrancheRequest));
647  }
648 
650  {
652 
653  while (i <= NamedLWLockTrancheRequests)
654  i *= 2;
655 
656  NamedLWLockTrancheRequestArray = (NamedLWLockTrancheRequest *)
657  repalloc(NamedLWLockTrancheRequestArray,
658  i * sizeof(NamedLWLockTrancheRequest));
660  }
661 
662  request = &NamedLWLockTrancheRequestArray[NamedLWLockTrancheRequests];
663  Assert(strlen(tranche_name) + 1 < NAMEDATALEN);
664  StrNCpy(request->tranche_name, tranche_name, NAMEDATALEN);
665  request->num_lwlocks = num_lwlocks;
667 }
668 
669 /*
670  * LWLockInitialize - initialize a new lwlock; it's initially unlocked
671  */
672 void
673 LWLockInitialize(LWLock *lock, int tranche_id)
674 {
676 #ifdef LOCK_DEBUG
677  pg_atomic_init_u32(&lock->nwaiters, 0);
678 #endif
679  lock->tranche = tranche_id;
680  proclist_init(&lock->waiters);
681 }
682 
683 /*
684  * Report start of wait event for light-weight locks.
685  *
686  * This function will be used by all the light-weight lock calls which
687  * needs to wait to acquire the lock. This function distinguishes wait
688  * event based on tranche and lock id.
689  */
690 static inline void
692 {
694 }
695 
696 /*
697  * Report end of wait event for light-weight locks.
698  */
699 static inline void
701 {
703 }
704 
705 /*
706  * Return an identifier for an LWLock based on the wait class and event.
707  */
708 const char *
710 {
711  Assert(classId == PG_WAIT_LWLOCK);
712 
713  /*
714  * It is quite possible that user has registered tranche in one of the
715  * backends (e.g. by allocating lwlocks in dynamic shared memory) but not
716  * all of them, so we can't assume the tranche is registered here.
717  */
718  if (eventId >= LWLockTranchesAllocated ||
719  LWLockTrancheArray[eventId] == NULL)
720  return "extension";
721 
722  return LWLockTrancheArray[eventId];
723 }
724 
725 /*
726  * Internal function that tries to atomically acquire the lwlock in the passed
727  * in mode.
728  *
729  * This function will not block waiting for a lock to become free - that's the
730  * callers job.
731  *
732  * Returns true if the lock isn't free and we need to wait.
733  */
734 static bool
736 {
737  uint32 old_state;
738 
739  AssertArg(mode == LW_EXCLUSIVE || mode == LW_SHARED);
740 
741  /*
742  * Read once outside the loop, later iterations will get the newer value
743  * via compare & exchange.
744  */
745  old_state = pg_atomic_read_u32(&lock->state);
746 
747  /* loop until we've determined whether we could acquire the lock or not */
748  while (true)
749  {
750  uint32 desired_state;
751  bool lock_free;
752 
753  desired_state = old_state;
754 
755  if (mode == LW_EXCLUSIVE)
756  {
757  lock_free = (old_state & LW_LOCK_MASK) == 0;
758  if (lock_free)
759  desired_state += LW_VAL_EXCLUSIVE;
760  }
761  else
762  {
763  lock_free = (old_state & LW_VAL_EXCLUSIVE) == 0;
764  if (lock_free)
765  desired_state += LW_VAL_SHARED;
766  }
767 
768  /*
769  * Attempt to swap in the state we are expecting. If we didn't see
770  * lock to be free, that's just the old value. If we saw it as free,
771  * we'll attempt to mark it acquired. The reason that we always swap
772  * in the value is that this doubles as a memory barrier. We could try
773  * to be smarter and only swap in values if we saw the lock as free,
774  * but benchmark haven't shown it as beneficial so far.
775  *
776  * Retry if the value changed since we last looked at it.
777  */
779  &old_state, desired_state))
780  {
781  if (lock_free)
782  {
783  /* Great! Got the lock. */
784 #ifdef LOCK_DEBUG
785  if (mode == LW_EXCLUSIVE)
786  lock->owner = MyProc;
787 #endif
788  return false;
789  }
790  else
791  return true; /* somebody else has the lock */
792  }
793  }
794  pg_unreachable();
795 }
796 
797 /*
798  * Lock the LWLock's wait list against concurrent activity.
799  *
800  * NB: even though the wait list is locked, non-conflicting lock operations
801  * may still happen concurrently.
802  *
803  * Time spent holding mutex should be short!
804  */
805 static void
807 {
808  uint32 old_state;
809 #ifdef LWLOCK_STATS
810  lwlock_stats *lwstats;
811  uint32 delays = 0;
812 
813  lwstats = get_lwlock_stats_entry(lock);
814 #endif
815 
816  while (true)
817  {
818  /* always try once to acquire lock directly */
819  old_state = pg_atomic_fetch_or_u32(&lock->state, LW_FLAG_LOCKED);
820  if (!(old_state & LW_FLAG_LOCKED))
821  break; /* got lock */
822 
823  /* and then spin without atomic operations until lock is released */
824  {
825  SpinDelayStatus delayStatus;
826 
827  init_local_spin_delay(&delayStatus);
828 
829  while (old_state & LW_FLAG_LOCKED)
830  {
831  perform_spin_delay(&delayStatus);
832  old_state = pg_atomic_read_u32(&lock->state);
833  }
834 #ifdef LWLOCK_STATS
835  delays += delayStatus.delays;
836 #endif
837  finish_spin_delay(&delayStatus);
838  }
839 
840  /*
841  * Retry. The lock might obviously already be re-acquired by the time
842  * we're attempting to get it again.
843  */
844  }
845 
846 #ifdef LWLOCK_STATS
847  lwstats->spin_delay_count += delays;
848 #endif
849 }
850 
851 /*
852  * Unlock the LWLock's wait list.
853  *
854  * Note that it can be more efficient to manipulate flags and release the
855  * locks in a single atomic operation.
856  */
857 static void
859 {
861 
862  old_state = pg_atomic_fetch_and_u32(&lock->state, ~LW_FLAG_LOCKED);
863 
864  Assert(old_state & LW_FLAG_LOCKED);
865 }
866 
867 /*
868  * Wakeup all the lockers that currently have a chance to acquire the lock.
869  */
870 static void
872 {
873  bool new_release_ok;
874  bool wokeup_somebody = false;
875  proclist_head wakeup;
877 
878  proclist_init(&wakeup);
879 
880  new_release_ok = true;
881 
882  /* lock wait list while collecting backends to wake up */
883  LWLockWaitListLock(lock);
884 
885  proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
886  {
887  PGPROC *waiter = GetPGProcByNumber(iter.cur);
888 
889  if (wokeup_somebody && waiter->lwWaitMode == LW_EXCLUSIVE)
890  continue;
891 
892  proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
893  proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
894 
895  if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
896  {
897  /*
898  * Prevent additional wakeups until retryer gets to run. Backends
899  * that are just waiting for the lock to become free don't retry
900  * automatically.
901  */
902  new_release_ok = false;
903 
904  /*
905  * Don't wakeup (further) exclusive locks.
906  */
907  wokeup_somebody = true;
908  }
909 
910  /*
911  * Once we've woken up an exclusive lock, there's no point in waking
912  * up anybody else.
913  */
914  if (waiter->lwWaitMode == LW_EXCLUSIVE)
915  break;
916  }
917 
919 
920  /* unset required flags, and release lock, in one fell swoop */
921  {
922  uint32 old_state;
923  uint32 desired_state;
924 
925  old_state = pg_atomic_read_u32(&lock->state);
926  while (true)
927  {
928  desired_state = old_state;
929 
930  /* compute desired flags */
931 
932  if (new_release_ok)
933  desired_state |= LW_FLAG_RELEASE_OK;
934  else
935  desired_state &= ~LW_FLAG_RELEASE_OK;
936 
937  if (proclist_is_empty(&wakeup))
938  desired_state &= ~LW_FLAG_HAS_WAITERS;
939 
940  desired_state &= ~LW_FLAG_LOCKED; /* release lock */
941 
942  if (pg_atomic_compare_exchange_u32(&lock->state, &old_state,
943  desired_state))
944  break;
945  }
946  }
947 
948  /* Awaken any waiters I removed from the queue. */
949  proclist_foreach_modify(iter, &wakeup, lwWaitLink)
950  {
951  PGPROC *waiter = GetPGProcByNumber(iter.cur);
952 
953  LOG_LWDEBUG("LWLockRelease", lock, "release waiter");
954  proclist_delete(&wakeup, iter.cur, lwWaitLink);
955 
956  /*
957  * Guarantee that lwWaiting being unset only becomes visible once the
958  * unlink from the link has completed. Otherwise the target backend
959  * could be woken up for other reason and enqueue for a new lock - if
960  * that happens before the list unlink happens, the list would end up
961  * being corrupted.
962  *
963  * The barrier pairs with the LWLockWaitListLock() when enqueuing for
964  * another lock.
965  */
967  waiter->lwWaiting = false;
968  PGSemaphoreUnlock(waiter->sem);
969  }
970 }
971 
972 /*
973  * Add ourselves to the end of the queue.
974  *
975  * NB: Mode can be LW_WAIT_UNTIL_FREE here!
976  */
977 static void
979 {
980  /*
981  * If we don't have a PGPROC structure, there's no way to wait. This
982  * should never occur, since MyProc should only be null during shared
983  * memory initialization.
984  */
985  if (MyProc == NULL)
986  elog(PANIC, "cannot wait without a PGPROC structure");
987 
988  if (MyProc->lwWaiting)
989  elog(PANIC, "queueing for lock while waiting on another one");
990 
991  LWLockWaitListLock(lock);
992 
993  /* setting the flag is protected by the spinlock */
995 
996  MyProc->lwWaiting = true;
997  MyProc->lwWaitMode = mode;
998 
999  /* LW_WAIT_UNTIL_FREE waiters are always at the front of the queue */
1000  if (mode == LW_WAIT_UNTIL_FREE)
1001  proclist_push_head(&lock->waiters, MyProc->pgprocno, lwWaitLink);
1002  else
1003  proclist_push_tail(&lock->waiters, MyProc->pgprocno, lwWaitLink);
1004 
1005  /* Can release the mutex now */
1006  LWLockWaitListUnlock(lock);
1007 
1008 #ifdef LOCK_DEBUG
1009  pg_atomic_fetch_add_u32(&lock->nwaiters, 1);
1010 #endif
1011 
1012 }
1013 
1014 /*
1015  * Remove ourselves from the waitlist.
1016  *
1017  * This is used if we queued ourselves because we thought we needed to sleep
1018  * but, after further checking, we discovered that we don't actually need to
1019  * do so.
1020  */
1021 static void
1023 {
1024  bool found = false;
1025  proclist_mutable_iter iter;
1026 
1027 #ifdef LWLOCK_STATS
1028  lwlock_stats *lwstats;
1029 
1030  lwstats = get_lwlock_stats_entry(lock);
1031 
1032  lwstats->dequeue_self_count++;
1033 #endif
1034 
1035  LWLockWaitListLock(lock);
1036 
1037  /*
1038  * Can't just remove ourselves from the list, but we need to iterate over
1039  * all entries as somebody else could have dequeued us.
1040  */
1041  proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
1042  {
1043  if (iter.cur == MyProc->pgprocno)
1044  {
1045  found = true;
1046  proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
1047  break;
1048  }
1049  }
1050 
1051  if (proclist_is_empty(&lock->waiters) &&
1052  (pg_atomic_read_u32(&lock->state) & LW_FLAG_HAS_WAITERS) != 0)
1053  {
1055  }
1056 
1057  /* XXX: combine with fetch_and above? */
1058  LWLockWaitListUnlock(lock);
1059 
1060  /* clear waiting state again, nice for debugging */
1061  if (found)
1062  MyProc->lwWaiting = false;
1063  else
1064  {
1065  int extraWaits = 0;
1066 
1067  /*
1068  * Somebody else dequeued us and has or will wake us up. Deal with the
1069  * superfluous absorption of a wakeup.
1070  */
1071 
1072  /*
1073  * Reset releaseOk if somebody woke us before we removed ourselves -
1074  * they'll have set it to false.
1075  */
1077 
1078  /*
1079  * Now wait for the scheduled wakeup, otherwise our ->lwWaiting would
1080  * get reset at some inconvenient point later. Most of the time this
1081  * will immediately return.
1082  */
1083  for (;;)
1084  {
1086  if (!MyProc->lwWaiting)
1087  break;
1088  extraWaits++;
1089  }
1090 
1091  /*
1092  * Fix the process wait semaphore's count for any absorbed wakeups.
1093  */
1094  while (extraWaits-- > 0)
1096  }
1097 
1098 #ifdef LOCK_DEBUG
1099  {
1100  /* not waiting anymore */
1101  uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1102 
1103  Assert(nwaiters < MAX_BACKENDS);
1104  }
1105 #endif
1106 }
1107 
1108 /*
1109  * LWLockAcquire - acquire a lightweight lock in the specified mode
1110  *
1111  * If the lock is not available, sleep until it is. Returns true if the lock
1112  * was available immediately, false if we had to sleep.
1113  *
1114  * Side effect: cancel/die interrupts are held off until lock release.
1115  */
1116 bool
1118 {
1119  PGPROC *proc = MyProc;
1120  bool result = true;
1121  int extraWaits = 0;
1122 #ifdef LWLOCK_STATS
1123  lwlock_stats *lwstats;
1124 
1125  lwstats = get_lwlock_stats_entry(lock);
1126 #endif
1127 
1128  AssertArg(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1129 
1130  PRINT_LWDEBUG("LWLockAcquire", lock, mode);
1131 
1132 #ifdef LWLOCK_STATS
1133  /* Count lock acquisition attempts */
1134  if (mode == LW_EXCLUSIVE)
1135  lwstats->ex_acquire_count++;
1136  else
1137  lwstats->sh_acquire_count++;
1138 #endif /* LWLOCK_STATS */
1139 
1140  /*
1141  * We can't wait if we haven't got a PGPROC. This should only occur
1142  * during bootstrap or shared memory initialization. Put an Assert here
1143  * to catch unsafe coding practices.
1144  */
1145  Assert(!(proc == NULL && IsUnderPostmaster));
1146 
1147  /* Ensure we will have room to remember the lock */
1149  elog(ERROR, "too many LWLocks taken");
1150 
1151  /*
1152  * Lock out cancel/die interrupts until we exit the code section protected
1153  * by the LWLock. This ensures that interrupts will not interfere with
1154  * manipulations of data structures in shared memory.
1155  */
1156  HOLD_INTERRUPTS();
1157 
1158  /*
1159  * Loop here to try to acquire lock after each time we are signaled by
1160  * LWLockRelease.
1161  *
1162  * NOTE: it might seem better to have LWLockRelease actually grant us the
1163  * lock, rather than retrying and possibly having to go back to sleep. But
1164  * in practice that is no good because it means a process swap for every
1165  * lock acquisition when two or more processes are contending for the same
1166  * lock. Since LWLocks are normally used to protect not-very-long
1167  * sections of computation, a process needs to be able to acquire and
1168  * release the same lock many times during a single CPU time slice, even
1169  * in the presence of contention. The efficiency of being able to do that
1170  * outweighs the inefficiency of sometimes wasting a process dispatch
1171  * cycle because the lock is not free when a released waiter finally gets
1172  * to run. See pgsql-hackers archives for 29-Dec-01.
1173  */
1174  for (;;)
1175  {
1176  bool mustwait;
1177 
1178  /*
1179  * Try to grab the lock the first time, we're not in the waitqueue
1180  * yet/anymore.
1181  */
1182  mustwait = LWLockAttemptLock(lock, mode);
1183 
1184  if (!mustwait)
1185  {
1186  LOG_LWDEBUG("LWLockAcquire", lock, "immediately acquired lock");
1187  break; /* got the lock */
1188  }
1189 
1190  /*
1191  * Ok, at this point we couldn't grab the lock on the first try. We
1192  * cannot simply queue ourselves to the end of the list and wait to be
1193  * woken up because by now the lock could long have been released.
1194  * Instead add us to the queue and try to grab the lock again. If we
1195  * succeed we need to revert the queuing and be happy, otherwise we
1196  * recheck the lock. If we still couldn't grab it, we know that the
1197  * other locker will see our queue entries when releasing since they
1198  * existed before we checked for the lock.
1199  */
1200 
1201  /* add to the queue */
1202  LWLockQueueSelf(lock, mode);
1203 
1204  /* we're now guaranteed to be woken up if necessary */
1205  mustwait = LWLockAttemptLock(lock, mode);
1206 
1207  /* ok, grabbed the lock the second time round, need to undo queueing */
1208  if (!mustwait)
1209  {
1210  LOG_LWDEBUG("LWLockAcquire", lock, "acquired, undoing queue");
1211 
1212  LWLockDequeueSelf(lock);
1213  break;
1214  }
1215 
1216  /*
1217  * Wait until awakened.
1218  *
1219  * Since we share the process wait semaphore with the regular lock
1220  * manager and ProcWaitForSignal, and we may need to acquire an LWLock
1221  * while one of those is pending, it is possible that we get awakened
1222  * for a reason other than being signaled by LWLockRelease. If so,
1223  * loop back and wait again. Once we've gotten the LWLock,
1224  * re-increment the sema by the number of additional signals received,
1225  * so that the lock manager or signal manager will see the received
1226  * signal when it next waits.
1227  */
1228  LOG_LWDEBUG("LWLockAcquire", lock, "waiting");
1229 
1230 #ifdef LWLOCK_STATS
1231  lwstats->block_count++;
1232 #endif
1233 
1234  LWLockReportWaitStart(lock);
1235  TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1236 
1237  for (;;)
1238  {
1239  PGSemaphoreLock(proc->sem);
1240  if (!proc->lwWaiting)
1241  break;
1242  extraWaits++;
1243  }
1244 
1245  /* Retrying, allow LWLockRelease to release waiters again. */
1247 
1248 #ifdef LOCK_DEBUG
1249  {
1250  /* not waiting anymore */
1251  uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1252 
1253  Assert(nwaiters < MAX_BACKENDS);
1254  }
1255 #endif
1256 
1257  TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1259 
1260  LOG_LWDEBUG("LWLockAcquire", lock, "awakened");
1261 
1262  /* Now loop back and try to acquire lock again. */
1263  result = false;
1264  }
1265 
1266  TRACE_POSTGRESQL_LWLOCK_ACQUIRE(T_NAME(lock), mode);
1267 
1268  /* Add lock to list of locks held by this backend */
1269  held_lwlocks[num_held_lwlocks].lock = lock;
1270  held_lwlocks[num_held_lwlocks++].mode = mode;
1271 
1272  /*
1273  * Fix the process wait semaphore's count for any absorbed wakeups.
1274  */
1275  while (extraWaits-- > 0)
1276  PGSemaphoreUnlock(proc->sem);
1277 
1278  return result;
1279 }
1280 
1281 /*
1282  * LWLockConditionalAcquire - acquire a lightweight lock in the specified mode
1283  *
1284  * If the lock is not available, return FALSE with no side-effects.
1285  *
1286  * If successful, cancel/die interrupts are held off until lock release.
1287  */
1288 bool
1290 {
1291  bool mustwait;
1292 
1293  AssertArg(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1294 
1295  PRINT_LWDEBUG("LWLockConditionalAcquire", lock, mode);
1296 
1297  /* Ensure we will have room to remember the lock */
1299  elog(ERROR, "too many LWLocks taken");
1300 
1301  /*
1302  * Lock out cancel/die interrupts until we exit the code section protected
1303  * by the LWLock. This ensures that interrupts will not interfere with
1304  * manipulations of data structures in shared memory.
1305  */
1306  HOLD_INTERRUPTS();
1307 
1308  /* Check for the lock */
1309  mustwait = LWLockAttemptLock(lock, mode);
1310 
1311  if (mustwait)
1312  {
1313  /* Failed to get lock, so release interrupt holdoff */
1315 
1316  LOG_LWDEBUG("LWLockConditionalAcquire", lock, "failed");
1317  TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE_FAIL(T_NAME(lock), mode);
1318  }
1319  else
1320  {
1321  /* Add lock to list of locks held by this backend */
1322  held_lwlocks[num_held_lwlocks].lock = lock;
1323  held_lwlocks[num_held_lwlocks++].mode = mode;
1324  TRACE_POSTGRESQL_LWLOCK_CONDACQUIRE(T_NAME(lock), mode);
1325  }
1326  return !mustwait;
1327 }
1328 
1329 /*
1330  * LWLockAcquireOrWait - Acquire lock, or wait until it's free
1331  *
1332  * The semantics of this function are a bit funky. If the lock is currently
1333  * free, it is acquired in the given mode, and the function returns true. If
1334  * the lock isn't immediately free, the function waits until it is released
1335  * and returns false, but does not acquire the lock.
1336  *
1337  * This is currently used for WALWriteLock: when a backend flushes the WAL,
1338  * holding WALWriteLock, it can flush the commit records of many other
1339  * backends as a side-effect. Those other backends need to wait until the
1340  * flush finishes, but don't need to acquire the lock anymore. They can just
1341  * wake up, observe that their records have already been flushed, and return.
1342  */
1343 bool
1345 {
1346  PGPROC *proc = MyProc;
1347  bool mustwait;
1348  int extraWaits = 0;
1349 #ifdef LWLOCK_STATS
1350  lwlock_stats *lwstats;
1351 
1352  lwstats = get_lwlock_stats_entry(lock);
1353 #endif
1354 
1355  Assert(mode == LW_SHARED || mode == LW_EXCLUSIVE);
1356 
1357  PRINT_LWDEBUG("LWLockAcquireOrWait", lock, mode);
1358 
1359  /* Ensure we will have room to remember the lock */
1361  elog(ERROR, "too many LWLocks taken");
1362 
1363  /*
1364  * Lock out cancel/die interrupts until we exit the code section protected
1365  * by the LWLock. This ensures that interrupts will not interfere with
1366  * manipulations of data structures in shared memory.
1367  */
1368  HOLD_INTERRUPTS();
1369 
1370  /*
1371  * NB: We're using nearly the same twice-in-a-row lock acquisition
1372  * protocol as LWLockAcquire(). Check its comments for details.
1373  */
1374  mustwait = LWLockAttemptLock(lock, mode);
1375 
1376  if (mustwait)
1377  {
1379 
1380  mustwait = LWLockAttemptLock(lock, mode);
1381 
1382  if (mustwait)
1383  {
1384  /*
1385  * Wait until awakened. Like in LWLockAcquire, be prepared for
1386  * bogus wakeups, because we share the semaphore with
1387  * ProcWaitForSignal.
1388  */
1389  LOG_LWDEBUG("LWLockAcquireOrWait", lock, "waiting");
1390 
1391 #ifdef LWLOCK_STATS
1392  lwstats->block_count++;
1393 #endif
1394 
1395  LWLockReportWaitStart(lock);
1396  TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), mode);
1397 
1398  for (;;)
1399  {
1400  PGSemaphoreLock(proc->sem);
1401  if (!proc->lwWaiting)
1402  break;
1403  extraWaits++;
1404  }
1405 
1406 #ifdef LOCK_DEBUG
1407  {
1408  /* not waiting anymore */
1409  uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1410 
1411  Assert(nwaiters < MAX_BACKENDS);
1412  }
1413 #endif
1414  TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), mode);
1416 
1417  LOG_LWDEBUG("LWLockAcquireOrWait", lock, "awakened");
1418  }
1419  else
1420  {
1421  LOG_LWDEBUG("LWLockAcquireOrWait", lock, "acquired, undoing queue");
1422 
1423  /*
1424  * Got lock in the second attempt, undo queueing. We need to treat
1425  * this as having successfully acquired the lock, otherwise we'd
1426  * not necessarily wake up people we've prevented from acquiring
1427  * the lock.
1428  */
1429  LWLockDequeueSelf(lock);
1430  }
1431  }
1432 
1433  /*
1434  * Fix the process wait semaphore's count for any absorbed wakeups.
1435  */
1436  while (extraWaits-- > 0)
1437  PGSemaphoreUnlock(proc->sem);
1438 
1439  if (mustwait)
1440  {
1441  /* Failed to get lock, so release interrupt holdoff */
1443  LOG_LWDEBUG("LWLockAcquireOrWait", lock, "failed");
1444  TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT_FAIL(T_NAME(lock), mode);
1445  }
1446  else
1447  {
1448  LOG_LWDEBUG("LWLockAcquireOrWait", lock, "succeeded");
1449  /* Add lock to list of locks held by this backend */
1450  held_lwlocks[num_held_lwlocks].lock = lock;
1451  held_lwlocks[num_held_lwlocks++].mode = mode;
1452  TRACE_POSTGRESQL_LWLOCK_ACQUIRE_OR_WAIT(T_NAME(lock), mode);
1453  }
1454 
1455  return !mustwait;
1456 }
1457 
1458 /*
1459  * Does the lwlock in its current state need to wait for the variable value to
1460  * change?
1461  *
1462  * If we don't need to wait, and it's because the value of the variable has
1463  * changed, store the current value in newval.
1464  *
1465  * *result is set to true if the lock was free, and false otherwise.
1466  */
1467 static bool
1469  uint64 *valptr, uint64 oldval, uint64 *newval,
1470  bool *result)
1471 {
1472  bool mustwait;
1473  uint64 value;
1474 
1475  /*
1476  * Test first to see if it the slot is free right now.
1477  *
1478  * XXX: the caller uses a spinlock before this, so we don't need a memory
1479  * barrier here as far as the current usage is concerned. But that might
1480  * not be safe in general.
1481  */
1482  mustwait = (pg_atomic_read_u32(&lock->state) & LW_VAL_EXCLUSIVE) != 0;
1483 
1484  if (!mustwait)
1485  {
1486  *result = true;
1487  return false;
1488  }
1489 
1490  *result = false;
1491 
1492  /*
1493  * Read value using the lwlock's wait list lock, as we can't generally
1494  * rely on atomic 64 bit reads/stores. TODO: On platforms with a way to
1495  * do atomic 64 bit reads/writes the spinlock should be optimized away.
1496  */
1497  LWLockWaitListLock(lock);
1498  value = *valptr;
1499  LWLockWaitListUnlock(lock);
1500 
1501  if (value != oldval)
1502  {
1503  mustwait = false;
1504  *newval = value;
1505  }
1506  else
1507  {
1508  mustwait = true;
1509  }
1510 
1511  return mustwait;
1512 }
1513 
1514 /*
1515  * LWLockWaitForVar - Wait until lock is free, or a variable is updated.
1516  *
1517  * If the lock is held and *valptr equals oldval, waits until the lock is
1518  * either freed, or the lock holder updates *valptr by calling
1519  * LWLockUpdateVar. If the lock is free on exit (immediately or after
1520  * waiting), returns true. If the lock is still held, but *valptr no longer
1521  * matches oldval, returns false and sets *newval to the current value in
1522  * *valptr.
1523  *
1524  * Note: this function ignores shared lock holders; if the lock is held
1525  * in shared mode, returns 'true'.
1526  */
1527 bool
1528 LWLockWaitForVar(LWLock *lock, uint64 *valptr, uint64 oldval, uint64 *newval)
1529 {
1530  PGPROC *proc = MyProc;
1531  int extraWaits = 0;
1532  bool result = false;
1533 #ifdef LWLOCK_STATS
1534  lwlock_stats *lwstats;
1535 
1536  lwstats = get_lwlock_stats_entry(lock);
1537 #endif
1538 
1539  PRINT_LWDEBUG("LWLockWaitForVar", lock, LW_WAIT_UNTIL_FREE);
1540 
1541  /*
1542  * Lock out cancel/die interrupts while we sleep on the lock. There is no
1543  * cleanup mechanism to remove us from the wait queue if we got
1544  * interrupted.
1545  */
1546  HOLD_INTERRUPTS();
1547 
1548  /*
1549  * Loop here to check the lock's status after each time we are signaled.
1550  */
1551  for (;;)
1552  {
1553  bool mustwait;
1554 
1555  mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1556  &result);
1557 
1558  if (!mustwait)
1559  break; /* the lock was free or value didn't match */
1560 
1561  /*
1562  * Add myself to wait queue. Note that this is racy, somebody else
1563  * could wakeup before we're finished queuing. NB: We're using nearly
1564  * the same twice-in-a-row lock acquisition protocol as
1565  * LWLockAcquire(). Check its comments for details. The only
1566  * difference is that we also have to check the variable's values when
1567  * checking the state of the lock.
1568  */
1570 
1571  /*
1572  * Set RELEASE_OK flag, to make sure we get woken up as soon as the
1573  * lock is released.
1574  */
1576 
1577  /*
1578  * We're now guaranteed to be woken up if necessary. Recheck the lock
1579  * and variables state.
1580  */
1581  mustwait = LWLockConflictsWithVar(lock, valptr, oldval, newval,
1582  &result);
1583 
1584  /* Ok, no conflict after we queued ourselves. Undo queueing. */
1585  if (!mustwait)
1586  {
1587  LOG_LWDEBUG("LWLockWaitForVar", lock, "free, undoing queue");
1588 
1589  LWLockDequeueSelf(lock);
1590  break;
1591  }
1592 
1593  /*
1594  * Wait until awakened.
1595  *
1596  * Since we share the process wait semaphore with the regular lock
1597  * manager and ProcWaitForSignal, and we may need to acquire an LWLock
1598  * while one of those is pending, it is possible that we get awakened
1599  * for a reason other than being signaled by LWLockRelease. If so,
1600  * loop back and wait again. Once we've gotten the LWLock,
1601  * re-increment the sema by the number of additional signals received,
1602  * so that the lock manager or signal manager will see the received
1603  * signal when it next waits.
1604  */
1605  LOG_LWDEBUG("LWLockWaitForVar", lock, "waiting");
1606 
1607 #ifdef LWLOCK_STATS
1608  lwstats->block_count++;
1609 #endif
1610 
1611  LWLockReportWaitStart(lock);
1612  TRACE_POSTGRESQL_LWLOCK_WAIT_START(T_NAME(lock), LW_EXCLUSIVE);
1613 
1614  for (;;)
1615  {
1616  PGSemaphoreLock(proc->sem);
1617  if (!proc->lwWaiting)
1618  break;
1619  extraWaits++;
1620  }
1621 
1622 #ifdef LOCK_DEBUG
1623  {
1624  /* not waiting anymore */
1625  uint32 nwaiters PG_USED_FOR_ASSERTS_ONLY = pg_atomic_fetch_sub_u32(&lock->nwaiters, 1);
1626 
1627  Assert(nwaiters < MAX_BACKENDS);
1628  }
1629 #endif
1630 
1631  TRACE_POSTGRESQL_LWLOCK_WAIT_DONE(T_NAME(lock), LW_EXCLUSIVE);
1633 
1634  LOG_LWDEBUG("LWLockWaitForVar", lock, "awakened");
1635 
1636  /* Now loop back and check the status of the lock again. */
1637  }
1638 
1639  TRACE_POSTGRESQL_LWLOCK_ACQUIRE(T_NAME(lock), LW_EXCLUSIVE);
1640 
1641  /*
1642  * Fix the process wait semaphore's count for any absorbed wakeups.
1643  */
1644  while (extraWaits-- > 0)
1645  PGSemaphoreUnlock(proc->sem);
1646 
1647  /*
1648  * Now okay to allow cancel/die interrupts.
1649  */
1651 
1652  return result;
1653 }
1654 
1655 
1656 /*
1657  * LWLockUpdateVar - Update a variable and wake up waiters atomically
1658  *
1659  * Sets *valptr to 'val', and wakes up all processes waiting for us with
1660  * LWLockWaitForVar(). Setting the value and waking up the processes happen
1661  * atomically so that any process calling LWLockWaitForVar() on the same lock
1662  * is guaranteed to see the new value, and act accordingly.
1663  *
1664  * The caller must be holding the lock in exclusive mode.
1665  */
1666 void
1667 LWLockUpdateVar(LWLock *lock, uint64 *valptr, uint64 val)
1668 {
1669  proclist_head wakeup;
1670  proclist_mutable_iter iter;
1671 
1672  PRINT_LWDEBUG("LWLockUpdateVar", lock, LW_EXCLUSIVE);
1673 
1674  proclist_init(&wakeup);
1675 
1676  LWLockWaitListLock(lock);
1677 
1679 
1680  /* Update the lock's value */
1681  *valptr = val;
1682 
1683  /*
1684  * See if there are any LW_WAIT_UNTIL_FREE waiters that need to be woken
1685  * up. They are always in the front of the queue.
1686  */
1687  proclist_foreach_modify(iter, &lock->waiters, lwWaitLink)
1688  {
1689  PGPROC *waiter = GetPGProcByNumber(iter.cur);
1690 
1691  if (waiter->lwWaitMode != LW_WAIT_UNTIL_FREE)
1692  break;
1693 
1694  proclist_delete(&lock->waiters, iter.cur, lwWaitLink);
1695  proclist_push_tail(&wakeup, iter.cur, lwWaitLink);
1696  }
1697 
1698  /* We are done updating shared state of the lock itself. */
1699  LWLockWaitListUnlock(lock);
1700 
1701  /*
1702  * Awaken any waiters I removed from the queue.
1703  */
1704  proclist_foreach_modify(iter, &wakeup, lwWaitLink)
1705  {
1706  PGPROC *waiter = GetPGProcByNumber(iter.cur);
1707 
1708  proclist_delete(&wakeup, iter.cur, lwWaitLink);
1709  /* check comment in LWLockWakeup() about this barrier */
1710  pg_write_barrier();
1711  waiter->lwWaiting = false;
1712  PGSemaphoreUnlock(waiter->sem);
1713  }
1714 }
1715 
1716 
1717 /*
1718  * LWLockRelease - release a previously acquired lock
1719  */
1720 void
1722 {
1723  LWLockMode mode;
1724  uint32 oldstate;
1725  bool check_waiters;
1726  int i;
1727 
1728  /*
1729  * Remove lock from list of locks held. Usually, but not always, it will
1730  * be the latest-acquired lock; so search array backwards.
1731  */
1732  for (i = num_held_lwlocks; --i >= 0;)
1733  if (lock == held_lwlocks[i].lock)
1734  break;
1735 
1736  if (i < 0)
1737  elog(ERROR, "lock %s is not held", T_NAME(lock));
1738 
1739  mode = held_lwlocks[i].mode;
1740 
1741  num_held_lwlocks--;
1742  for (; i < num_held_lwlocks; i++)
1743  held_lwlocks[i] = held_lwlocks[i + 1];
1744 
1745  PRINT_LWDEBUG("LWLockRelease", lock, mode);
1746 
1747  /*
1748  * Release my hold on lock, after that it can immediately be acquired by
1749  * others, even if we still have to wakeup other waiters.
1750  */
1751  if (mode == LW_EXCLUSIVE)
1752  oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_EXCLUSIVE);
1753  else
1754  oldstate = pg_atomic_sub_fetch_u32(&lock->state, LW_VAL_SHARED);
1755 
1756  /* nobody else can have that kind of lock */
1757  Assert(!(oldstate & LW_VAL_EXCLUSIVE));
1758 
1759 
1760  /*
1761  * We're still waiting for backends to get scheduled, don't wake them up
1762  * again.
1763  */
1764  if ((oldstate & (LW_FLAG_HAS_WAITERS | LW_FLAG_RELEASE_OK)) ==
1766  (oldstate & LW_LOCK_MASK) == 0)
1767  check_waiters = true;
1768  else
1769  check_waiters = false;
1770 
1771  /*
1772  * As waking up waiters requires the spinlock to be acquired, only do so
1773  * if necessary.
1774  */
1775  if (check_waiters)
1776  {
1777  /* XXX: remove before commit? */
1778  LOG_LWDEBUG("LWLockRelease", lock, "releasing waiters");
1779  LWLockWakeup(lock);
1780  }
1781 
1782  TRACE_POSTGRESQL_LWLOCK_RELEASE(T_NAME(lock));
1783 
1784  /*
1785  * Now okay to allow cancel/die interrupts.
1786  */
1788 }
1789 
1790 /*
1791  * LWLockReleaseClearVar - release a previously acquired lock, reset variable
1792  */
1793 void
1794 LWLockReleaseClearVar(LWLock *lock, uint64 *valptr, uint64 val)
1795 {
1796  LWLockWaitListLock(lock);
1797 
1798  /*
1799  * Set the variable's value before releasing the lock, that prevents race
1800  * a race condition wherein a new locker acquires the lock, but hasn't yet
1801  * set the variables value.
1802  */
1803  *valptr = val;
1804  LWLockWaitListUnlock(lock);
1805 
1806  LWLockRelease(lock);
1807 }
1808 
1809 
1810 /*
1811  * LWLockReleaseAll - release all currently-held locks
1812  *
1813  * Used to clean up after ereport(ERROR). An important difference between this
1814  * function and retail LWLockRelease calls is that InterruptHoldoffCount is
1815  * unchanged by this operation. This is necessary since InterruptHoldoffCount
1816  * has been set to an appropriate level earlier in error recovery. We could
1817  * decrement it below zero if we allow it to drop for each released lock!
1818  */
1819 void
1821 {
1822  while (num_held_lwlocks > 0)
1823  {
1824  HOLD_INTERRUPTS(); /* match the upcoming RESUME_INTERRUPTS */
1825 
1826  LWLockRelease(held_lwlocks[num_held_lwlocks - 1].lock);
1827  }
1828 }
1829 
1830 
1831 /*
1832  * LWLockHeldByMe - test whether my process holds a lock in any mode
1833  *
1834  * This is meant as debug support only.
1835  */
1836 bool
1838 {
1839  int i;
1840 
1841  for (i = 0; i < num_held_lwlocks; i++)
1842  {
1843  if (held_lwlocks[i].lock == l)
1844  return true;
1845  }
1846  return false;
1847 }
1848 
1849 /*
1850  * LWLockHeldByMeInMode - test whether my process holds a lock in given mode
1851  *
1852  * This is meant as debug support only.
1853  */
1854 bool
1856 {
1857  int i;
1858 
1859  for (i = 0; i < num_held_lwlocks; i++)
1860  {
1861  if (held_lwlocks[i].lock == l && held_lwlocks[i].mode == mode)
1862  return true;
1863  }
1864  return false;
1865 }
#define T_NAME(lock)
Definition: lwlock.c:117
int slock_t
Definition: s_lock.h:888
#define init_local_spin_delay(status)
Definition: s_lock.h:997
static uint32 pg_atomic_fetch_sub_u32(volatile pg_atomic_uint32 *ptr, int32 sub_)
Definition: atomics.h:354
Definition: lwlock.h:32
#define pg_unreachable()
Definition: c.h:952
void MemoryContextDelete(MemoryContext context)
Definition: mcxt.c:200
bool LWLockHeldByMeInMode(LWLock *l, LWLockMode mode)
Definition: lwlock.c:1855
#define GetPGProcByNumber(n)
Definition: proc.h:277
static bool LWLockAttemptLock(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:735
int MyProcPid
Definition: globals.c:39
#define LW_FLAG_LOCKED
Definition: lwlock.c:101
int LWLockNewTrancheId(void)
Definition: lwlock.c:578
LWLockMode
Definition: lwlock.h:132
#define HASH_CONTEXT
Definition: hsearch.h:93
void MemoryContextAllowInCriticalSection(MemoryContext context, bool allow)
Definition: mcxt.c:374
#define HASH_ELEM
Definition: hsearch.h:87
void PGSemaphoreUnlock(PGSemaphore sema)
Definition: posix_sema.c:323
MemoryContext hcxt
Definition: hsearch.h:78
void RequestNamedLWLockTranche(const char *tranche_name, int num_lwlocks)
Definition: lwlock.c:633
static bool lock_named_request_allowed
Definition: lwlock.c:158
bool LWLockHeldByMe(LWLock *l)
Definition: lwlock.c:1837
#define LW_FLAG_HAS_WAITERS
Definition: lwlock.c:99
static bool pg_atomic_compare_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 *expected, uint32 newval)
Definition: atomics.h:322
PGPROC * MyProc
Definition: proc.c:67
proclist_head waiters
Definition: lwlock.h:36
static uint32 pg_atomic_sub_fetch_u32(volatile pg_atomic_uint32 *ptr, int32 sub_)
Definition: atomics.h:412
void LWLockUpdateVar(LWLock *lock, uint64 *valptr, uint64 val)
Definition: lwlock.c:1667
char tranche_name[NAMEDATALEN]
Definition: lwlock.c:148
struct LWLockHandle LWLockHandle
bool lwWaiting
Definition: proc.h:127
#define MAX_SIMUL_LWLOCKS
Definition: lwlock.c:133
Size entrysize
Definition: hsearch.h:73
int errhidestmt(bool hide_stmt)
Definition: elog.c:1068
#define MemSet(start, val, len)
Definition: c.h:863
static void LWLockWaitListUnlock(LWLock *lock)
Definition: lwlock.c:858
#define proclist_foreach_modify(iter, lhead, link_member)
Definition: proclist.h:195
#define LW_SHARED_MASK
Definition: lwlock.c:108
uint8 lwWaitMode
Definition: proc.h:128
void * ShmemAlloc(Size size)
Definition: shmem.c:157
pg_atomic_uint32 state
Definition: lwlock.h:35
#define LW_LOCK_MASK
Definition: lwlock.c:106
void * hash_search(HTAB *hashp, const void *keyPtr, HASHACTION action, bool *foundPtr)
Definition: dynahash.c:902
#define LOG
Definition: elog.h:26
NamedLWLockTranche * NamedLWLockTrancheArray
Definition: lwlock.c:156
#define PANIC
Definition: elog.h:53
#define proclist_delete(list, procno, link_member)
Definition: proclist.h:176
Size LWLockShmemSize(void)
Definition: lwlock.c:349
#define LW_VAL_SHARED
Definition: lwlock.c:104
void LWLockReleaseClearVar(LWLock *lock, uint64 *valptr, uint64 val)
Definition: lwlock.c:1794
static void RegisterLWLockTranches(void)
Definition: lwlock.c:491
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1721
#define RESUME_INTERRUPTS()
Definition: miscadmin.h:117
static int LWLockTranchesAllocated
Definition: lwlock.c:115
#define NAMEDATALEN
#define SpinLockAcquire(lock)
Definition: spin.h:62
Definition: dynahash.c:208
#define NUM_FIXED_LWLOCKS
Definition: lwlock.h:129
unsigned short uint16
Definition: c.h:257
static LWLockHandle held_lwlocks[MAX_SIMUL_LWLOCKS]
Definition: lwlock.c:143
#define LWLOCK_MINIMAL_SIZE
Definition: lwlock.h:74
#define ERROR
Definition: elog.h:43
#define MAX_BACKENDS
Definition: postmaster.h:75
LWLockMode mode
Definition: lwlock.c:139
#define proclist_push_head(list, procno, link_member)
Definition: proclist.h:178
int NamedLWLockTrancheRequests
Definition: lwlock.c:154
static struct @121 value
#define ALLOCSET_DEFAULT_SIZES
Definition: memutils.h:165
void finish_spin_delay(SpinDelayStatus *status)
Definition: s_lock.c:175
void on_shmem_exit(pg_on_exit_callback function, Datum arg)
Definition: ipc.c:348
bool LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1289
bool IsUnderPostmaster
Definition: globals.c:101
bool LWLockWaitForVar(LWLock *lock, uint64 *valptr, uint64 oldval, uint64 *newval)
Definition: lwlock.c:1528
static void LWLockReportWaitEnd(void)
Definition: lwlock.c:700
static void LWLockWakeup(LWLock *lock)
Definition: lwlock.c:871
static uint32 pg_atomic_fetch_and_u32(volatile pg_atomic_uint32 *ptr, uint32 and_)
Definition: atomics.h:369
#define NUM_BUFFER_PARTITIONS
Definition: lwlock.h:113
unsigned int uint32
Definition: c.h:258
static void InitializeLWLocks(void)
Definition: lwlock.c:426
static void pgstat_report_wait_end(void)
Definition: pgstat.h:1244
uint16 tranche
Definition: lwlock.h:34
static void LWLockDequeueSelf(LWLock *lock)
Definition: lwlock.c:1022
#define ereport(elevel, rest)
Definition: elog.h:122
#define AssertArg(condition)
Definition: c.h:683
char * trancheName
Definition: lwlock.h:97
MemoryContext TopMemoryContext
Definition: mcxt.c:43
void LWLockInitialize(LWLock *lock, int tranche_id)
Definition: lwlock.c:673
static void LWLockReportWaitStart(LWLock *lock)
Definition: lwlock.c:691
#define SpinLockRelease(lock)
Definition: spin.h:64
#define HASH_BLOBS
Definition: hsearch.h:88
slock_t * ShmemLock
Definition: shmem.c:84
Size mul_size(Size s1, Size s2)
Definition: shmem.c:492
MemoryContext AllocSetContextCreate(MemoryContext parent, const char *name, Size minContextSize, Size initBlockSize, Size maxBlockSize)
Definition: aset.c:322
HTAB * hash_create(const char *tabname, long nelem, HASHCTL *info, int flags)
Definition: dynahash.c:316
uintptr_t Datum
Definition: postgres.h:372
Size add_size(Size s1, Size s2)
Definition: shmem.c:475
static int NamedLWLockTrancheRequestsAllocated
Definition: lwlock.c:153
Size keysize
Definition: hsearch.h:72
bool LWLockAcquireOrWait(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1344
#define StaticAssertExpr(condition, errmessage)
Definition: c.h:765
void * MemoryContextAllocZero(MemoryContext context, Size size)
Definition: mcxt.c:741
static void LWLockQueueSelf(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:978
static int num_held_lwlocks
Definition: lwlock.c:142
LWLock lock
Definition: lwlock.h:79
static uint32 pg_atomic_fetch_add_u32(volatile pg_atomic_uint32 *ptr, int32 add_)
Definition: atomics.h:339
int errmsg_internal(const char *fmt,...)
Definition: elog.c:827
void CreateLWLocks(void)
Definition: lwlock.c:381
static char ** LWLockTrancheArray
Definition: lwlock.c:114
#define Assert(condition)
Definition: c.h:681
#define StrNCpy(dst, src, len)
Definition: c.h:836
LWLockPadded * GetNamedLWLockTranche(const char *tranche_name)
Definition: lwlock.c:547
Definition: regguts.h:298
size_t Size
Definition: c.h:350
static void pgstat_report_wait_start(uint32 wait_event_info)
Definition: pgstat.h:1220
LWLock * lock
Definition: lwlock.c:138
#define newval
#define PRINT_LWDEBUG(a, b, c)
Definition: lwlock.c:227
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1117
static bool proclist_is_empty(proclist_head *list)
Definition: proclist.h:38
void LWLockRegisterTranche(int tranche_id, char *tranche_name)
Definition: lwlock.c:598
#define LW_FLAG_RELEASE_OK
Definition: lwlock.c:100
char * MainLWLockNames[]
void * hash_seq_search(HASH_SEQ_STATUS *status)
Definition: dynahash.c:1385
static int NumLWLocksByNamedTranches(void)
Definition: lwlock.c:334
void * repalloc(void *pointer, Size size)
Definition: mcxt.c:962
void hash_seq_init(HASH_SEQ_STATUS *status, HTAB *hashp)
Definition: dynahash.c:1375
#define proclist_push_tail(list, procno, link_member)
Definition: proclist.h:180
void PGSemaphoreLock(PGSemaphore sema)
Definition: posix_sema.c:303
const char * name
Definition: encode.c:521
NamedLWLockTrancheRequest * NamedLWLockTrancheRequestArray
Definition: lwlock.c:152
static bool LWLockConflictsWithVar(LWLock *lock, uint64 *valptr, uint64 oldval, uint64 *newval, bool *result)
Definition: lwlock.c:1468
static void LWLockWaitListLock(LWLock *lock)
Definition: lwlock.c:806
struct NamedLWLockTrancheRequest NamedLWLockTrancheRequest
int pgprocno
Definition: proc.h:110
#define LW_VAL_EXCLUSIVE
Definition: lwlock.c:103
void * MemoryContextAlloc(MemoryContext context, Size size)
Definition: mcxt.c:706
const char * GetLWLockIdentifier(uint32 classId, uint16 eventId)
Definition: lwlock.c:709
#define HOLD_INTERRUPTS()
Definition: miscadmin.h:115
int i
static void proclist_init(proclist_head *list)
Definition: proclist.h:29
#define pg_write_barrier()
Definition: atomics.h:162
void * arg
static uint32 pg_atomic_fetch_or_u32(volatile pg_atomic_uint32 *ptr, uint32 or_)
Definition: atomics.h:383
static void pg_atomic_init_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
Definition: atomics.h:234
#define elog
Definition: elog.h:219
void LWLockReleaseAll(void)
Definition: lwlock.c:1820
#define LWLOCK_PADDED_SIZE
Definition: lwlock.h:73
PGSemaphore sem
Definition: proc.h:101
#define PG_USED_FOR_ASSERTS_ONLY
Definition: c.h:996
LWLockPadded * MainLWLockArray
Definition: lwlock.c:125
#define PG_WAIT_LWLOCK
Definition: pgstat.h:737
Definition: proc.h:95
long val
Definition: informix.c:689
int errhidecontext(bool hide_ctx)
Definition: elog.c:1087
void InitLWLockAccess(void)
Definition: lwlock.c:531
#define NUM_LOCK_PARTITIONS
Definition: lwlock.h:117
void perform_spin_delay(SpinDelayStatus *status)
Definition: s_lock.c:125
static uint32 pg_atomic_read_u32(volatile pg_atomic_uint32 *ptr)
Definition: atomics.h:252
#define NUM_PREDICATELOCK_PARTITIONS
Definition: lwlock.h:121
#define LOG_LWDEBUG(a, b, c)
Definition: lwlock.c:228