Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * proc.c
4 : * routines to manage per-process shared memory data structure
5 : *
6 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/lmgr/proc.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : /*
16 : * Interface (a):
17 : * ProcSleep(), ProcWakeup(),
18 : *
19 : * Waiting for a lock causes the backend to be put to sleep. Whoever releases
20 : * the lock wakes the process up again (and gives it an error code so it knows
21 : * whether it was awoken on an error condition).
22 : *
23 : * Interface (b):
24 : *
25 : * ProcReleaseLocks -- frees the locks associated with current transaction
26 : *
27 : * ProcKill -- destroys the shared memory state (and locks)
28 : * associated with the process.
29 : */
30 : #include "postgres.h"
31 :
32 : #include <signal.h>
33 : #include <unistd.h>
34 : #include <sys/time.h>
35 :
36 : #include "access/transam.h"
37 : #include "access/twophase.h"
38 : #include "access/xlogutils.h"
39 : #include "miscadmin.h"
40 : #include "pgstat.h"
41 : #include "postmaster/autovacuum.h"
42 : #include "replication/slotsync.h"
43 : #include "replication/syncrep.h"
44 : #include "storage/condition_variable.h"
45 : #include "storage/ipc.h"
46 : #include "storage/lmgr.h"
47 : #include "storage/pmsignal.h"
48 : #include "storage/proc.h"
49 : #include "storage/procarray.h"
50 : #include "storage/procsignal.h"
51 : #include "storage/spin.h"
52 : #include "storage/standby.h"
53 : #include "utils/timeout.h"
54 : #include "utils/timestamp.h"
55 :
56 : /* GUC variables */
57 : int DeadlockTimeout = 1000;
58 : int StatementTimeout = 0;
59 : int LockTimeout = 0;
60 : int IdleInTransactionSessionTimeout = 0;
61 : int TransactionTimeout = 0;
62 : int IdleSessionTimeout = 0;
63 : bool log_lock_waits = false;
64 :
65 : /* Pointer to this process's PGPROC struct, if any */
66 : PGPROC *MyProc = NULL;
67 :
68 : /*
69 : * This spinlock protects the freelist of recycled PGPROC structures.
70 : * We cannot use an LWLock because the LWLock manager depends on already
71 : * having a PGPROC and a wait semaphore! But these structures are touched
72 : * relatively infrequently (only at backend startup or shutdown) and not for
73 : * very long, so a spinlock is okay.
74 : */
75 : NON_EXEC_STATIC slock_t *ProcStructLock = NULL;
76 :
77 : /* Pointers to shared-memory structures */
78 : PROC_HDR *ProcGlobal = NULL;
79 : NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
80 : PGPROC *PreparedXactProcs = NULL;
81 :
82 : /* If we are waiting for a lock, this points to the associated LOCALLOCK */
83 : static LOCALLOCK *lockAwaited = NULL;
84 :
85 : static DeadLockState deadlock_state = DS_NOT_YET_CHECKED;
86 :
87 : /* Is a deadlock check pending? */
88 : static volatile sig_atomic_t got_deadlock_timeout;
89 :
90 : static void RemoveProcFromArray(int code, Datum arg);
91 : static void ProcKill(int code, Datum arg);
92 : static void AuxiliaryProcKill(int code, Datum arg);
93 : static void CheckDeadLock(void);
94 :
95 :
96 : /*
97 : * Report shared-memory space needed by InitProcGlobal.
98 : */
99 : Size
100 3298 : ProcGlobalShmemSize(void)
101 : {
102 3298 : Size size = 0;
103 : Size TotalProcs =
104 3298 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
105 :
106 : /* ProcGlobal */
107 3298 : size = add_size(size, sizeof(PROC_HDR));
108 3298 : size = add_size(size, mul_size(TotalProcs, sizeof(PGPROC)));
109 3298 : size = add_size(size, sizeof(slock_t));
110 :
111 3298 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->xids)));
112 3298 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->subxidStates)));
113 3298 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->statusFlags)));
114 :
115 3298 : return size;
116 : }
117 :
118 : /*
119 : * Report number of semaphores needed by InitProcGlobal.
120 : */
121 : int
122 3298 : ProcGlobalSemas(void)
123 : {
124 : /*
125 : * We need a sema per backend (including autovacuum), plus one for each
126 : * auxiliary process.
127 : */
128 3298 : return MaxBackends + NUM_AUXILIARY_PROCS;
129 : }
130 :
131 : /*
132 : * InitProcGlobal -
133 : * Initialize the global process table during postmaster or standalone
134 : * backend startup.
135 : *
136 : * We also create all the per-process semaphores we will need to support
137 : * the requested number of backends. We used to allocate semaphores
138 : * only when backends were actually started up, but that is bad because
139 : * it lets Postgres fail under load --- a lot of Unix systems are
140 : * (mis)configured with small limits on the number of semaphores, and
141 : * running out when trying to start another backend is a common failure.
142 : * So, now we grab enough semaphores to support the desired max number
143 : * of backends immediately at initialization --- if the sysadmin has set
144 : * MaxConnections, max_worker_processes, max_wal_senders, or
145 : * autovacuum_max_workers higher than his kernel will support, he'll
146 : * find out sooner rather than later.
147 : *
148 : * Another reason for creating semaphores here is that the semaphore
149 : * implementation typically requires us to create semaphores in the
150 : * postmaster, not in backends.
151 : *
152 : * Note: this is NOT called by individual backends under a postmaster,
153 : * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
154 : * pointers must be propagated specially for EXEC_BACKEND operation.
155 : */
156 : void
157 1768 : InitProcGlobal(void)
158 : {
159 : PGPROC *procs;
160 : int i,
161 : j;
162 : bool found;
163 1768 : uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
164 :
165 : /* Create the ProcGlobal shared structure */
166 1768 : ProcGlobal = (PROC_HDR *)
167 1768 : ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
168 : Assert(!found);
169 :
170 : /*
171 : * Initialize the data structures.
172 : */
173 1768 : ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
174 1768 : dlist_init(&ProcGlobal->freeProcs);
175 1768 : dlist_init(&ProcGlobal->autovacFreeProcs);
176 1768 : dlist_init(&ProcGlobal->bgworkerFreeProcs);
177 1768 : dlist_init(&ProcGlobal->walsenderFreeProcs);
178 1768 : ProcGlobal->startupBufferPinWaitBufId = -1;
179 1768 : ProcGlobal->walwriterLatch = NULL;
180 1768 : ProcGlobal->checkpointerLatch = NULL;
181 1768 : pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PROC_NUMBER);
182 1768 : pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PROC_NUMBER);
183 :
184 : /*
185 : * Create and initialize all the PGPROC structures we'll need. There are
186 : * five separate consumers: (1) normal backends, (2) autovacuum workers
187 : * and the autovacuum launcher, (3) background workers, (4) auxiliary
188 : * processes, and (5) prepared transactions. Each PGPROC structure is
189 : * dedicated to exactly one of these purposes, and they do not move
190 : * between groups.
191 : */
192 1768 : procs = (PGPROC *) ShmemAlloc(TotalProcs * sizeof(PGPROC));
193 1768 : MemSet(procs, 0, TotalProcs * sizeof(PGPROC));
194 1768 : ProcGlobal->allProcs = procs;
195 : /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
196 1768 : ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
197 :
198 : /*
199 : * Allocate arrays mirroring PGPROC fields in a dense manner. See
200 : * PROC_HDR.
201 : *
202 : * XXX: It might make sense to increase padding for these arrays, given
203 : * how hotly they are accessed.
204 : */
205 3536 : ProcGlobal->xids =
206 1768 : (TransactionId *) ShmemAlloc(TotalProcs * sizeof(*ProcGlobal->xids));
207 75872 : MemSet(ProcGlobal->xids, 0, TotalProcs * sizeof(*ProcGlobal->xids));
208 1768 : ProcGlobal->subxidStates = (XidCacheStatus *) ShmemAlloc(TotalProcs * sizeof(*ProcGlobal->subxidStates));
209 14636 : MemSet(ProcGlobal->subxidStates, 0, TotalProcs * sizeof(*ProcGlobal->subxidStates));
210 1768 : ProcGlobal->statusFlags = (uint8 *) ShmemAlloc(TotalProcs * sizeof(*ProcGlobal->statusFlags));
211 8188 : MemSet(ProcGlobal->statusFlags, 0, TotalProcs * sizeof(*ProcGlobal->statusFlags));
212 :
213 160206 : for (i = 0; i < TotalProcs; i++)
214 : {
215 158438 : PGPROC *proc = &procs[i];
216 :
217 : /* Common initialization for all PGPROCs, regardless of type. */
218 :
219 : /*
220 : * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
221 : * dummy PGPROCs don't need these though - they're never associated
222 : * with a real process
223 : */
224 158438 : if (i < MaxBackends + NUM_AUXILIARY_PROCS)
225 : {
226 156868 : proc->sem = PGSemaphoreCreate();
227 156868 : InitSharedLatch(&(proc->procLatch));
228 156868 : LWLockInitialize(&(proc->fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
229 : }
230 :
231 : /*
232 : * Newly created PGPROCs for normal backends, autovacuum and bgworkers
233 : * must be queued up on the appropriate free list. Because there can
234 : * only ever be a small, fixed number of auxiliary processes, no free
235 : * list is used in that case; InitAuxiliaryProcess() instead uses a
236 : * linear search. PGPROCs for prepared transactions are added to a
237 : * free list by TwoPhaseShmemInit().
238 : */
239 158438 : if (i < MaxConnections)
240 : {
241 : /* PGPROC for normal backend, add to freeProcs list */
242 113140 : dlist_push_tail(&ProcGlobal->freeProcs, &proc->links);
243 113140 : proc->procgloballist = &ProcGlobal->freeProcs;
244 : }
245 45298 : else if (i < MaxConnections + autovacuum_max_workers + 1)
246 : {
247 : /* PGPROC for AV launcher/worker, add to autovacFreeProcs list */
248 7072 : dlist_push_tail(&ProcGlobal->autovacFreeProcs, &proc->links);
249 7072 : proc->procgloballist = &ProcGlobal->autovacFreeProcs;
250 : }
251 38226 : else if (i < MaxConnections + autovacuum_max_workers + 1 + max_worker_processes)
252 : {
253 : /* PGPROC for bgworker, add to bgworkerFreeProcs list */
254 14168 : dlist_push_tail(&ProcGlobal->bgworkerFreeProcs, &proc->links);
255 14168 : proc->procgloballist = &ProcGlobal->bgworkerFreeProcs;
256 : }
257 24058 : else if (i < MaxBackends)
258 : {
259 : /* PGPROC for walsender, add to walsenderFreeProcs list */
260 11880 : dlist_push_tail(&ProcGlobal->walsenderFreeProcs, &proc->links);
261 11880 : proc->procgloballist = &ProcGlobal->walsenderFreeProcs;
262 : }
263 :
264 : /* Initialize myProcLocks[] shared memory queues. */
265 2693446 : for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
266 2535008 : dlist_init(&(proc->myProcLocks[j]));
267 :
268 : /* Initialize lockGroupMembers list. */
269 158438 : dlist_init(&proc->lockGroupMembers);
270 :
271 : /*
272 : * Initialize the atomic variables, otherwise, it won't be safe to
273 : * access them for backends that aren't currently in use.
274 : */
275 158438 : pg_atomic_init_u32(&(proc->procArrayGroupNext), INVALID_PROC_NUMBER);
276 158438 : pg_atomic_init_u32(&(proc->clogGroupNext), INVALID_PROC_NUMBER);
277 158438 : pg_atomic_init_u64(&(proc->waitStart), 0);
278 : }
279 :
280 : /*
281 : * Save pointers to the blocks of PGPROC structures reserved for auxiliary
282 : * processes and prepared transactions.
283 : */
284 1768 : AuxiliaryProcs = &procs[MaxBackends];
285 1768 : PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
286 :
287 : /* Create ProcStructLock spinlock, too */
288 1768 : ProcStructLock = (slock_t *) ShmemAlloc(sizeof(slock_t));
289 1768 : SpinLockInit(ProcStructLock);
290 1768 : }
291 :
292 : /*
293 : * InitProcess -- initialize a per-process PGPROC entry for this backend
294 : */
295 : void
296 25728 : InitProcess(void)
297 : {
298 : dlist_head *procgloballist;
299 :
300 : /*
301 : * ProcGlobal should be set up already (if we are a backend, we inherit
302 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
303 : */
304 25728 : if (ProcGlobal == NULL)
305 0 : elog(PANIC, "proc header uninitialized");
306 :
307 25728 : if (MyProc != NULL)
308 0 : elog(ERROR, "you already exist");
309 :
310 : /* Decide which list should supply our PGPROC. */
311 25728 : if (AmAutoVacuumLauncherProcess() || AmAutoVacuumWorkerProcess())
312 676 : procgloballist = &ProcGlobal->autovacFreeProcs;
313 25052 : else if (AmBackgroundWorkerProcess())
314 4084 : procgloballist = &ProcGlobal->bgworkerFreeProcs;
315 20968 : else if (AmWalSenderProcess())
316 1942 : procgloballist = &ProcGlobal->walsenderFreeProcs;
317 : else
318 19026 : procgloballist = &ProcGlobal->freeProcs;
319 :
320 : /*
321 : * Try to get a proc struct from the appropriate free list. If this
322 : * fails, we must be out of PGPROC structures (not to mention semaphores).
323 : *
324 : * While we are holding the ProcStructLock, also copy the current shared
325 : * estimate of spins_per_delay to local storage.
326 : */
327 25728 : SpinLockAcquire(ProcStructLock);
328 :
329 25728 : set_spins_per_delay(ProcGlobal->spins_per_delay);
330 :
331 25728 : if (!dlist_is_empty(procgloballist))
332 : {
333 25724 : MyProc = (PGPROC *) dlist_pop_head_node(procgloballist);
334 25724 : SpinLockRelease(ProcStructLock);
335 : }
336 : else
337 : {
338 : /*
339 : * If we reach here, all the PGPROCs are in use. This is one of the
340 : * possible places to detect "too many backends", so give the standard
341 : * error message. XXX do we need to give a different failure message
342 : * in the autovacuum case?
343 : */
344 4 : SpinLockRelease(ProcStructLock);
345 4 : if (AmWalSenderProcess())
346 4 : ereport(FATAL,
347 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
348 : errmsg("number of requested standby connections exceeds max_wal_senders (currently %d)",
349 : max_wal_senders)));
350 0 : ereport(FATAL,
351 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
352 : errmsg("sorry, too many clients already")));
353 : }
354 25724 : MyProcNumber = GetNumberFromPGProc(MyProc);
355 :
356 : /*
357 : * Cross-check that the PGPROC is of the type we expect; if this were not
358 : * the case, it would get returned to the wrong list.
359 : */
360 : Assert(MyProc->procgloballist == procgloballist);
361 :
362 : /*
363 : * Now that we have a PGPROC, mark ourselves as an active postmaster
364 : * child; this is so that the postmaster can detect it if we exit without
365 : * cleaning up. (XXX autovac launcher currently doesn't participate in
366 : * this; it probably should.)
367 : *
368 : * Slot sync worker also does not participate in it, see comments atop
369 : * 'struct bkend' in postmaster.c.
370 : */
371 25724 : if (IsUnderPostmaster && !AmAutoVacuumLauncherProcess() &&
372 24894 : !AmLogicalSlotSyncWorkerProcess())
373 24886 : MarkPostmasterChildActive();
374 :
375 : /*
376 : * Initialize all fields of MyProc, except for those previously
377 : * initialized by InitProcGlobal.
378 : */
379 25724 : dlist_node_init(&MyProc->links);
380 25724 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
381 25724 : MyProc->fpVXIDLock = false;
382 25724 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
383 25724 : MyProc->xid = InvalidTransactionId;
384 25724 : MyProc->xmin = InvalidTransactionId;
385 25724 : MyProc->pid = MyProcPid;
386 25724 : MyProc->vxid.procNumber = MyProcNumber;
387 25724 : MyProc->vxid.lxid = InvalidLocalTransactionId;
388 : /* databaseId and roleId will be filled in later */
389 25724 : MyProc->databaseId = InvalidOid;
390 25724 : MyProc->roleId = InvalidOid;
391 25724 : MyProc->tempNamespaceId = InvalidOid;
392 25724 : MyProc->isBackgroundWorker = AmBackgroundWorkerProcess();
393 25724 : MyProc->delayChkptFlags = 0;
394 25724 : MyProc->statusFlags = 0;
395 : /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
396 25724 : if (AmAutoVacuumWorkerProcess())
397 24 : MyProc->statusFlags |= PROC_IS_AUTOVACUUM;
398 25724 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
399 25724 : MyProc->lwWaitMode = 0;
400 25724 : MyProc->waitLock = NULL;
401 25724 : MyProc->waitProcLock = NULL;
402 25724 : pg_atomic_write_u64(&MyProc->waitStart, 0);
403 : #ifdef USE_ASSERT_CHECKING
404 : {
405 : int i;
406 :
407 : /* Last process should have released all locks. */
408 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
409 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
410 : }
411 : #endif
412 25724 : MyProc->recoveryConflictPending = false;
413 :
414 : /* Initialize fields for sync rep */
415 25724 : MyProc->waitLSN = 0;
416 25724 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
417 25724 : dlist_node_init(&MyProc->syncRepLinks);
418 :
419 : /* Initialize fields for group XID clearing. */
420 25724 : MyProc->procArrayGroupMember = false;
421 25724 : MyProc->procArrayGroupMemberXid = InvalidTransactionId;
422 : Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PROC_NUMBER);
423 :
424 : /* Check that group locking fields are in a proper initial state. */
425 : Assert(MyProc->lockGroupLeader == NULL);
426 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
427 :
428 : /* Initialize wait event information. */
429 25724 : MyProc->wait_event_info = 0;
430 :
431 : /* Initialize fields for group transaction status update. */
432 25724 : MyProc->clogGroupMember = false;
433 25724 : MyProc->clogGroupMemberXid = InvalidTransactionId;
434 25724 : MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
435 25724 : MyProc->clogGroupMemberPage = -1;
436 25724 : MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
437 : Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PROC_NUMBER);
438 :
439 : /*
440 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
441 : * on it. That allows us to repoint the process latch, which so far
442 : * points to process local one, to the shared one.
443 : */
444 25724 : OwnLatch(&MyProc->procLatch);
445 25724 : SwitchToSharedLatch();
446 :
447 : /* now that we have a proc, report wait events to shared memory */
448 25724 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
449 :
450 : /*
451 : * We might be reusing a semaphore that belonged to a failed process. So
452 : * be careful and reinitialize its value here. (This is not strictly
453 : * necessary anymore, but seems like a good idea for cleanliness.)
454 : */
455 25724 : PGSemaphoreReset(MyProc->sem);
456 :
457 : /*
458 : * Arrange to clean up at backend exit.
459 : */
460 25724 : on_shmem_exit(ProcKill, 0);
461 :
462 : /*
463 : * Now that we have a PGPROC, we could try to acquire locks, so initialize
464 : * local state needed for LWLocks, and the deadlock checker.
465 : */
466 25724 : InitLWLockAccess();
467 25724 : InitDeadLockChecking();
468 :
469 : #ifdef EXEC_BACKEND
470 :
471 : /*
472 : * Initialize backend-local pointers to all the shared data structures.
473 : * (We couldn't do this until now because it needs LWLocks.)
474 : */
475 : if (IsUnderPostmaster)
476 : AttachSharedMemoryStructs();
477 : #endif
478 25724 : }
479 :
480 : /*
481 : * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
482 : *
483 : * This is separate from InitProcess because we can't acquire LWLocks until
484 : * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
485 : * work until after we've done AttachSharedMemoryStructs.
486 : */
487 : void
488 25706 : InitProcessPhase2(void)
489 : {
490 : Assert(MyProc != NULL);
491 :
492 : /*
493 : * Add our PGPROC to the PGPROC array in shared memory.
494 : */
495 25706 : ProcArrayAdd(MyProc);
496 :
497 : /*
498 : * Arrange to clean that up at backend exit.
499 : */
500 25706 : on_shmem_exit(RemoveProcFromArray, 0);
501 25706 : }
502 :
503 : /*
504 : * InitAuxiliaryProcess -- create a PGPROC entry for an auxiliary process
505 : *
506 : * This is called by bgwriter and similar processes so that they will have a
507 : * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
508 : * and sema that are assigned are one of the extra ones created during
509 : * InitProcGlobal.
510 : *
511 : * Auxiliary processes are presently not expected to wait for real (lockmgr)
512 : * locks, so we need not set up the deadlock checker. They are never added
513 : * to the ProcArray or the sinval messaging mechanism, either. They also
514 : * don't get a VXID assigned, since this is only useful when we actually
515 : * hold lockmgr locks.
516 : *
517 : * Startup process however uses locks but never waits for them in the
518 : * normal backend sense. Startup process also takes part in sinval messaging
519 : * as a sendOnly process, so never reads messages from sinval queue. So
520 : * Startup process does have a VXID and does show up in pg_locks.
521 : */
522 : void
523 4152 : InitAuxiliaryProcess(void)
524 : {
525 : PGPROC *auxproc;
526 : int proctype;
527 :
528 : /*
529 : * ProcGlobal should be set up already (if we are a backend, we inherit
530 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
531 : */
532 4152 : if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
533 0 : elog(PANIC, "proc header uninitialized");
534 :
535 4152 : if (MyProc != NULL)
536 0 : elog(ERROR, "you already exist");
537 :
538 : /*
539 : * We use the ProcStructLock to protect assignment and releasing of
540 : * AuxiliaryProcs entries.
541 : *
542 : * While we are holding the ProcStructLock, also copy the current shared
543 : * estimate of spins_per_delay to local storage.
544 : */
545 4152 : SpinLockAcquire(ProcStructLock);
546 :
547 4152 : set_spins_per_delay(ProcGlobal->spins_per_delay);
548 :
549 : /*
550 : * Find a free auxproc ... *big* trouble if there isn't one ...
551 : */
552 10358 : for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
553 : {
554 10358 : auxproc = &AuxiliaryProcs[proctype];
555 10358 : if (auxproc->pid == 0)
556 4152 : break;
557 : }
558 4152 : if (proctype >= NUM_AUXILIARY_PROCS)
559 : {
560 0 : SpinLockRelease(ProcStructLock);
561 0 : elog(FATAL, "all AuxiliaryProcs are in use");
562 : }
563 :
564 : /* Mark auxiliary proc as in use by me */
565 : /* use volatile pointer to prevent code rearrangement */
566 4152 : ((volatile PGPROC *) auxproc)->pid = MyProcPid;
567 :
568 4152 : SpinLockRelease(ProcStructLock);
569 :
570 4152 : MyProc = auxproc;
571 4152 : MyProcNumber = GetNumberFromPGProc(MyProc);
572 :
573 : /*
574 : * Initialize all fields of MyProc, except for those previously
575 : * initialized by InitProcGlobal.
576 : */
577 4152 : dlist_node_init(&MyProc->links);
578 4152 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
579 4152 : MyProc->fpVXIDLock = false;
580 4152 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
581 4152 : MyProc->xid = InvalidTransactionId;
582 4152 : MyProc->xmin = InvalidTransactionId;
583 4152 : MyProc->vxid.procNumber = INVALID_PROC_NUMBER;
584 4152 : MyProc->vxid.lxid = InvalidLocalTransactionId;
585 4152 : MyProc->databaseId = InvalidOid;
586 4152 : MyProc->roleId = InvalidOid;
587 4152 : MyProc->tempNamespaceId = InvalidOid;
588 4152 : MyProc->isBackgroundWorker = AmBackgroundWorkerProcess();
589 4152 : MyProc->delayChkptFlags = 0;
590 4152 : MyProc->statusFlags = 0;
591 4152 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
592 4152 : MyProc->lwWaitMode = 0;
593 4152 : MyProc->waitLock = NULL;
594 4152 : MyProc->waitProcLock = NULL;
595 4152 : pg_atomic_write_u64(&MyProc->waitStart, 0);
596 : #ifdef USE_ASSERT_CHECKING
597 : {
598 : int i;
599 :
600 : /* Last process should have released all locks. */
601 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
602 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
603 : }
604 : #endif
605 :
606 : /*
607 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
608 : * on it. That allows us to repoint the process latch, which so far
609 : * points to process local one, to the shared one.
610 : */
611 4152 : OwnLatch(&MyProc->procLatch);
612 4152 : SwitchToSharedLatch();
613 :
614 : /* now that we have a proc, report wait events to shared memory */
615 4152 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
616 :
617 : /* Check that group locking fields are in a proper initial state. */
618 : Assert(MyProc->lockGroupLeader == NULL);
619 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
620 :
621 : /*
622 : * We might be reusing a semaphore that belonged to a failed process. So
623 : * be careful and reinitialize its value here. (This is not strictly
624 : * necessary anymore, but seems like a good idea for cleanliness.)
625 : */
626 4152 : PGSemaphoreReset(MyProc->sem);
627 :
628 : /*
629 : * Arrange to clean up at process exit.
630 : */
631 4152 : on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
632 :
633 : /*
634 : * Now that we have a PGPROC, we could try to acquire lightweight locks.
635 : * Initialize local state needed for them. (Heavyweight locks cannot be
636 : * acquired in aux processes.)
637 : */
638 4152 : InitLWLockAccess();
639 :
640 : #ifdef EXEC_BACKEND
641 :
642 : /*
643 : * Initialize backend-local pointers to all the shared data structures.
644 : * (We couldn't do this until now because it needs LWLocks.)
645 : */
646 : if (IsUnderPostmaster)
647 : AttachSharedMemoryStructs();
648 : #endif
649 4152 : }
650 :
651 : /*
652 : * Used from bufmgr to share the value of the buffer that Startup waits on,
653 : * or to reset the value to "not waiting" (-1). This allows processing
654 : * of recovery conflicts for buffer pins. Set is made before backends look
655 : * at this value, so locking not required, especially since the set is
656 : * an atomic integer set operation.
657 : */
658 : void
659 40 : SetStartupBufferPinWaitBufId(int bufid)
660 : {
661 : /* use volatile pointer to prevent code rearrangement */
662 40 : volatile PROC_HDR *procglobal = ProcGlobal;
663 :
664 40 : procglobal->startupBufferPinWaitBufId = bufid;
665 40 : }
666 :
667 : /*
668 : * Used by backends when they receive a request to check for buffer pin waits.
669 : */
670 : int
671 10 : GetStartupBufferPinWaitBufId(void)
672 : {
673 : /* use volatile pointer to prevent code rearrangement */
674 10 : volatile PROC_HDR *procglobal = ProcGlobal;
675 :
676 10 : return procglobal->startupBufferPinWaitBufId;
677 : }
678 :
679 : /*
680 : * Check whether there are at least N free PGPROC objects. If false is
681 : * returned, *nfree will be set to the number of free PGPROC objects.
682 : * Otherwise, *nfree will be set to n.
683 : *
684 : * Note: this is designed on the assumption that N will generally be small.
685 : */
686 : bool
687 420 : HaveNFreeProcs(int n, int *nfree)
688 : {
689 : dlist_iter iter;
690 :
691 : Assert(n > 0);
692 : Assert(nfree);
693 :
694 420 : SpinLockAcquire(ProcStructLock);
695 :
696 420 : *nfree = 0;
697 1260 : dlist_foreach(iter, &ProcGlobal->freeProcs)
698 : {
699 1260 : (*nfree)++;
700 1260 : if (*nfree == n)
701 420 : break;
702 : }
703 :
704 420 : SpinLockRelease(ProcStructLock);
705 :
706 420 : return (*nfree == n);
707 : }
708 :
709 : /*
710 : * Check if the current process is awaiting a lock.
711 : */
712 : bool
713 16 : IsWaitingForLock(void)
714 : {
715 16 : if (lockAwaited == NULL)
716 12 : return false;
717 :
718 4 : return true;
719 : }
720 :
721 : /*
722 : * Cancel any pending wait for lock, when aborting a transaction, and revert
723 : * any strong lock count acquisition for a lock being acquired.
724 : *
725 : * (Normally, this would only happen if we accept a cancel/die
726 : * interrupt while waiting; but an ereport(ERROR) before or during the lock
727 : * wait is within the realm of possibility, too.)
728 : */
729 : void
730 621516 : LockErrorCleanup(void)
731 : {
732 : LWLock *partitionLock;
733 : DisableTimeoutParams timeouts[2];
734 :
735 621516 : HOLD_INTERRUPTS();
736 :
737 621516 : AbortStrongLockAcquire();
738 :
739 : /* Nothing to do if we weren't waiting for a lock */
740 621516 : if (lockAwaited == NULL)
741 : {
742 621434 : RESUME_INTERRUPTS();
743 621434 : return;
744 : }
745 :
746 : /*
747 : * Turn off the deadlock and lock timeout timers, if they are still
748 : * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
749 : * indicator flag, since this function is executed before
750 : * ProcessInterrupts when responding to SIGINT; else we'd lose the
751 : * knowledge that the SIGINT came from a lock timeout and not an external
752 : * source.
753 : */
754 82 : timeouts[0].id = DEADLOCK_TIMEOUT;
755 82 : timeouts[0].keep_indicator = false;
756 82 : timeouts[1].id = LOCK_TIMEOUT;
757 82 : timeouts[1].keep_indicator = true;
758 82 : disable_timeouts(timeouts, 2);
759 :
760 : /* Unlink myself from the wait queue, if on it (might not be anymore!) */
761 82 : partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
762 82 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
763 :
764 82 : if (!dlist_node_is_detached(&MyProc->links))
765 : {
766 : /* We could not have been granted the lock yet */
767 82 : RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
768 : }
769 : else
770 : {
771 : /*
772 : * Somebody kicked us off the lock queue already. Perhaps they
773 : * granted us the lock, or perhaps they detected a deadlock. If they
774 : * did grant us the lock, we'd better remember it in our local lock
775 : * table.
776 : */
777 0 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
778 0 : GrantAwaitedLock();
779 : }
780 :
781 82 : lockAwaited = NULL;
782 :
783 82 : LWLockRelease(partitionLock);
784 :
785 82 : RESUME_INTERRUPTS();
786 : }
787 :
788 :
789 : /*
790 : * ProcReleaseLocks() -- release locks associated with current transaction
791 : * at main transaction commit or abort
792 : *
793 : * At main transaction commit, we release standard locks except session locks.
794 : * At main transaction abort, we release all locks including session locks.
795 : *
796 : * Advisory locks are released only if they are transaction-level;
797 : * session-level holds remain, whether this is a commit or not.
798 : *
799 : * At subtransaction commit, we don't release any locks (so this func is not
800 : * needed at all); we will defer the releasing to the parent transaction.
801 : * At subtransaction abort, we release all locks held by the subtransaction;
802 : * this is implemented by retail releasing of the locks under control of
803 : * the ResourceOwner mechanism.
804 : */
805 : void
806 566216 : ProcReleaseLocks(bool isCommit)
807 : {
808 566216 : if (!MyProc)
809 0 : return;
810 : /* If waiting, get off wait queue (should only be needed after error) */
811 566216 : LockErrorCleanup();
812 : /* Release standard locks, including session-level if aborting */
813 566216 : LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
814 : /* Release transaction-level advisory locks */
815 566216 : LockReleaseAll(USER_LOCKMETHOD, false);
816 : }
817 :
818 :
819 : /*
820 : * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
821 : */
822 : static void
823 25706 : RemoveProcFromArray(int code, Datum arg)
824 : {
825 : Assert(MyProc != NULL);
826 25706 : ProcArrayRemove(MyProc, InvalidTransactionId);
827 25706 : }
828 :
829 : /*
830 : * ProcKill() -- Destroy the per-proc data structure for
831 : * this process. Release any of its held LW locks.
832 : */
833 : static void
834 25724 : ProcKill(int code, Datum arg)
835 : {
836 : PGPROC *proc;
837 : dlist_head *procgloballist;
838 :
839 : Assert(MyProc != NULL);
840 :
841 : /* not safe if forked by system(), etc. */
842 25724 : if (MyProc->pid != (int) getpid())
843 0 : elog(PANIC, "ProcKill() called in child process");
844 :
845 : /* Make sure we're out of the sync rep lists */
846 25724 : SyncRepCleanupAtProcExit();
847 :
848 : #ifdef USE_ASSERT_CHECKING
849 : {
850 : int i;
851 :
852 : /* Last process should have released all locks. */
853 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
854 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
855 : }
856 : #endif
857 :
858 : /*
859 : * Release any LW locks I am holding. There really shouldn't be any, but
860 : * it's cheap to check again before we cut the knees off the LWLock
861 : * facility by releasing our PGPROC ...
862 : */
863 25724 : LWLockReleaseAll();
864 :
865 : /* Cancel any pending condition variable sleep, too */
866 25724 : ConditionVariableCancelSleep();
867 :
868 : /*
869 : * Detach from any lock group of which we are a member. If the leader
870 : * exits before all other group members, its PGPROC will remain allocated
871 : * until the last group process exits; that process must return the
872 : * leader's PGPROC to the appropriate list.
873 : */
874 25724 : if (MyProc->lockGroupLeader != NULL)
875 : {
876 2758 : PGPROC *leader = MyProc->lockGroupLeader;
877 2758 : LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
878 :
879 2758 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
880 : Assert(!dlist_is_empty(&leader->lockGroupMembers));
881 2758 : dlist_delete(&MyProc->lockGroupLink);
882 2758 : if (dlist_is_empty(&leader->lockGroupMembers))
883 : {
884 116 : leader->lockGroupLeader = NULL;
885 116 : if (leader != MyProc)
886 : {
887 0 : procgloballist = leader->procgloballist;
888 :
889 : /* Leader exited first; return its PGPROC. */
890 0 : SpinLockAcquire(ProcStructLock);
891 0 : dlist_push_head(procgloballist, &leader->links);
892 0 : SpinLockRelease(ProcStructLock);
893 : }
894 : }
895 2642 : else if (leader != MyProc)
896 2642 : MyProc->lockGroupLeader = NULL;
897 2758 : LWLockRelease(leader_lwlock);
898 : }
899 :
900 : /*
901 : * Reset MyLatch to the process local one. This is so that signal
902 : * handlers et al can continue using the latch after the shared latch
903 : * isn't ours anymore.
904 : *
905 : * Similarly, stop reporting wait events to MyProc->wait_event_info.
906 : *
907 : * After that clear MyProc and disown the shared latch.
908 : */
909 25724 : SwitchBackToLocalLatch();
910 25724 : pgstat_reset_wait_event_storage();
911 :
912 25724 : proc = MyProc;
913 25724 : MyProc = NULL;
914 25724 : MyProcNumber = INVALID_PROC_NUMBER;
915 25724 : DisownLatch(&proc->procLatch);
916 :
917 : /* Mark the proc no longer in use */
918 25724 : proc->pid = 0;
919 25724 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
920 25724 : proc->vxid.lxid = InvalidTransactionId;
921 :
922 25724 : procgloballist = proc->procgloballist;
923 25724 : SpinLockAcquire(ProcStructLock);
924 :
925 : /*
926 : * If we're still a member of a locking group, that means we're a leader
927 : * which has somehow exited before its children. The last remaining child
928 : * will release our PGPROC. Otherwise, release it now.
929 : */
930 25724 : if (proc->lockGroupLeader == NULL)
931 : {
932 : /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
933 : Assert(dlist_is_empty(&proc->lockGroupMembers));
934 :
935 : /* Return PGPROC structure (and semaphore) to appropriate freelist */
936 25724 : dlist_push_tail(procgloballist, &proc->links);
937 : }
938 :
939 : /* Update shared estimate of spins_per_delay */
940 25724 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
941 :
942 25724 : SpinLockRelease(ProcStructLock);
943 :
944 : /*
945 : * This process is no longer present in shared memory in any meaningful
946 : * way, so tell the postmaster we've cleaned up acceptably well. (XXX
947 : * autovac launcher should be included here someday)
948 : *
949 : * Slot sync worker is also not a postmaster child, so skip this shared
950 : * memory related processing here.
951 : */
952 25724 : if (IsUnderPostmaster && !AmAutoVacuumLauncherProcess() &&
953 24894 : !AmLogicalSlotSyncWorkerProcess())
954 24886 : MarkPostmasterChildInactive();
955 :
956 : /* wake autovac launcher if needed -- see comments in FreeWorkerInfo */
957 25724 : if (AutovacuumLauncherPid != 0)
958 24 : kill(AutovacuumLauncherPid, SIGUSR2);
959 25724 : }
960 :
961 : /*
962 : * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
963 : * processes (bgwriter, etc). The PGPROC and sema are not released, only
964 : * marked as not-in-use.
965 : */
966 : static void
967 4152 : AuxiliaryProcKill(int code, Datum arg)
968 : {
969 4152 : int proctype = DatumGetInt32(arg);
970 : PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
971 : PGPROC *proc;
972 :
973 : Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
974 :
975 : /* not safe if forked by system(), etc. */
976 4152 : if (MyProc->pid != (int) getpid())
977 0 : elog(PANIC, "AuxiliaryProcKill() called in child process");
978 :
979 4152 : auxproc = &AuxiliaryProcs[proctype];
980 :
981 : Assert(MyProc == auxproc);
982 :
983 : /* Release any LW locks I am holding (see notes above) */
984 4152 : LWLockReleaseAll();
985 :
986 : /* Cancel any pending condition variable sleep, too */
987 4152 : ConditionVariableCancelSleep();
988 :
989 : /* look at the equivalent ProcKill() code for comments */
990 4152 : SwitchBackToLocalLatch();
991 4152 : pgstat_reset_wait_event_storage();
992 :
993 4152 : proc = MyProc;
994 4152 : MyProc = NULL;
995 4152 : MyProcNumber = INVALID_PROC_NUMBER;
996 4152 : DisownLatch(&proc->procLatch);
997 :
998 4152 : SpinLockAcquire(ProcStructLock);
999 :
1000 : /* Mark auxiliary proc no longer in use */
1001 4152 : proc->pid = 0;
1002 4152 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1003 4152 : proc->vxid.lxid = InvalidTransactionId;
1004 :
1005 : /* Update shared estimate of spins_per_delay */
1006 4152 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1007 :
1008 4152 : SpinLockRelease(ProcStructLock);
1009 4152 : }
1010 :
1011 : /*
1012 : * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
1013 : * given its PID
1014 : *
1015 : * Returns NULL if not found.
1016 : */
1017 : PGPROC *
1018 3824 : AuxiliaryPidGetProc(int pid)
1019 : {
1020 3824 : PGPROC *result = NULL;
1021 : int index;
1022 :
1023 3824 : if (pid == 0) /* never match dummy PGPROCs */
1024 0 : return NULL;
1025 :
1026 7790 : for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
1027 : {
1028 7790 : PGPROC *proc = &AuxiliaryProcs[index];
1029 :
1030 7790 : if (proc->pid == pid)
1031 : {
1032 3824 : result = proc;
1033 3824 : break;
1034 : }
1035 : }
1036 3824 : return result;
1037 : }
1038 :
1039 :
1040 : /*
1041 : * ProcSleep -- put a process to sleep on the specified lock
1042 : *
1043 : * Caller must have set MyProc->heldLocks to reflect locks already held
1044 : * on the lockable object by this process (under all XIDs).
1045 : *
1046 : * It's not actually guaranteed that we need to wait when this function is
1047 : * called, because it could be that when we try to find a position at which
1048 : * to insert ourself into the wait queue, we discover that we must be inserted
1049 : * ahead of everyone who wants a lock that conflict with ours. In that case,
1050 : * we get the lock immediately. Because of this, it's sensible for this function
1051 : * to have a dontWait argument, despite the name.
1052 : *
1053 : * The lock table's partition lock must be held at entry, and will be held
1054 : * at exit.
1055 : *
1056 : * Result: PROC_WAIT_STATUS_OK if we acquired the lock, PROC_WAIT_STATUS_ERROR
1057 : * if not (if dontWait = true, this is a deadlock; if dontWait = false, we
1058 : * would have had to wait).
1059 : *
1060 : * ASSUME: that no one will fiddle with the queue until after
1061 : * we release the partition lock.
1062 : *
1063 : * NOTES: The process queue is now a priority queue for locking.
1064 : */
1065 : ProcWaitStatus
1066 3478 : ProcSleep(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
1067 : {
1068 3478 : LOCKMODE lockmode = locallock->tag.mode;
1069 3478 : LOCK *lock = locallock->lock;
1070 3478 : PROCLOCK *proclock = locallock->proclock;
1071 3478 : uint32 hashcode = locallock->hashcode;
1072 3478 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1073 3478 : dclist_head *waitQueue = &lock->waitProcs;
1074 3478 : PGPROC *insert_before = NULL;
1075 3478 : LOCKMASK myHeldLocks = MyProc->heldLocks;
1076 3478 : TimestampTz standbyWaitStart = 0;
1077 3478 : bool early_deadlock = false;
1078 3478 : bool allow_autovacuum_cancel = true;
1079 3478 : bool logged_recovery_conflict = false;
1080 : ProcWaitStatus myWaitStatus;
1081 3478 : PGPROC *leader = MyProc->lockGroupLeader;
1082 :
1083 : /*
1084 : * If group locking is in use, locks held by members of my locking group
1085 : * need to be included in myHeldLocks. This is not required for relation
1086 : * extension lock which conflict among group members. However, including
1087 : * them in myHeldLocks will give group members the priority to get those
1088 : * locks as compared to other backends which are also trying to acquire
1089 : * those locks. OTOH, we can avoid giving priority to group members for
1090 : * that kind of locks, but there doesn't appear to be a clear advantage of
1091 : * the same.
1092 : */
1093 3478 : if (leader != NULL)
1094 : {
1095 : dlist_iter iter;
1096 :
1097 60 : dlist_foreach(iter, &lock->procLocks)
1098 : {
1099 : PROCLOCK *otherproclock;
1100 :
1101 46 : otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
1102 :
1103 46 : if (otherproclock->groupLeader == leader)
1104 22 : myHeldLocks |= otherproclock->holdMask;
1105 : }
1106 : }
1107 :
1108 : /*
1109 : * Determine where to add myself in the wait queue.
1110 : *
1111 : * Normally I should go at the end of the queue. However, if I already
1112 : * hold locks that conflict with the request of any previous waiter, put
1113 : * myself in the queue just in front of the first such waiter. This is not
1114 : * a necessary step, since deadlock detection would move me to before that
1115 : * waiter anyway; but it's relatively cheap to detect such a conflict
1116 : * immediately, and avoid delaying till deadlock timeout.
1117 : *
1118 : * Special case: if I find I should go in front of some waiter, check to
1119 : * see if I conflict with already-held locks or the requests before that
1120 : * waiter. If not, then just grant myself the requested lock immediately.
1121 : * This is the same as the test for immediate grant in LockAcquire, except
1122 : * we are only considering the part of the wait queue before my insertion
1123 : * point.
1124 : */
1125 3478 : if (myHeldLocks != 0 && !dclist_is_empty(waitQueue))
1126 : {
1127 8 : LOCKMASK aheadRequests = 0;
1128 : dlist_iter iter;
1129 :
1130 8 : dclist_foreach(iter, waitQueue)
1131 : {
1132 8 : PGPROC *proc = dlist_container(PGPROC, links, iter.cur);
1133 :
1134 : /*
1135 : * If we're part of the same locking group as this waiter, its
1136 : * locks neither conflict with ours nor contribute to
1137 : * aheadRequests.
1138 : */
1139 8 : if (leader != NULL && leader == proc->lockGroupLeader)
1140 0 : continue;
1141 :
1142 : /* Must he wait for me? */
1143 8 : if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1144 : {
1145 : /* Must I wait for him ? */
1146 8 : if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1147 : {
1148 : /*
1149 : * Yes, so we have a deadlock. Easiest way to clean up
1150 : * correctly is to call RemoveFromWaitQueue(), but we
1151 : * can't do that until we are *on* the wait queue. So, set
1152 : * a flag to check below, and break out of loop. Also,
1153 : * record deadlock info for later message.
1154 : */
1155 2 : RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1156 2 : early_deadlock = true;
1157 2 : break;
1158 : }
1159 : /* I must go before this waiter. Check special case. */
1160 6 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1161 6 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1162 : proclock))
1163 : {
1164 : /* Skip the wait and just grant myself the lock. */
1165 6 : GrantLock(lock, proclock, lockmode);
1166 6 : GrantAwaitedLock();
1167 6 : return PROC_WAIT_STATUS_OK;
1168 : }
1169 :
1170 : /* Put myself into wait queue before conflicting process */
1171 0 : insert_before = proc;
1172 0 : break;
1173 : }
1174 : /* Nope, so advance to next waiter */
1175 0 : aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1176 : }
1177 : }
1178 :
1179 : /*
1180 : * At this point we know that we'd really need to sleep. If we've been
1181 : * commanded not to do that, bail out.
1182 : */
1183 3472 : if (dontWait)
1184 1328 : return PROC_WAIT_STATUS_ERROR;
1185 :
1186 : /*
1187 : * Insert self into queue, at the position determined above.
1188 : */
1189 2144 : if (insert_before)
1190 0 : dclist_insert_before(waitQueue, &insert_before->links, &MyProc->links);
1191 : else
1192 2144 : dclist_push_tail(waitQueue, &MyProc->links);
1193 :
1194 2144 : lock->waitMask |= LOCKBIT_ON(lockmode);
1195 :
1196 : /* Set up wait information in PGPROC object, too */
1197 2144 : MyProc->waitLock = lock;
1198 2144 : MyProc->waitProcLock = proclock;
1199 2144 : MyProc->waitLockMode = lockmode;
1200 :
1201 2144 : MyProc->waitStatus = PROC_WAIT_STATUS_WAITING;
1202 :
1203 : /*
1204 : * If we detected deadlock, give up without waiting. This must agree with
1205 : * CheckDeadLock's recovery code.
1206 : */
1207 2144 : if (early_deadlock)
1208 : {
1209 2 : RemoveFromWaitQueue(MyProc, hashcode);
1210 2 : return PROC_WAIT_STATUS_ERROR;
1211 : }
1212 :
1213 : /* mark that we are waiting for a lock */
1214 2142 : lockAwaited = locallock;
1215 :
1216 : /*
1217 : * Release the lock table's partition lock.
1218 : *
1219 : * NOTE: this may also cause us to exit critical-section state, possibly
1220 : * allowing a cancel/die interrupt to be accepted. This is OK because we
1221 : * have recorded the fact that we are waiting for a lock, and so
1222 : * LockErrorCleanup will clean up if cancel/die happens.
1223 : */
1224 2142 : LWLockRelease(partitionLock);
1225 :
1226 : /*
1227 : * Also, now that we will successfully clean up after an ereport, it's
1228 : * safe to check to see if there's a buffer pin deadlock against the
1229 : * Startup process. Of course, that's only necessary if we're doing Hot
1230 : * Standby and are not the Startup process ourselves.
1231 : */
1232 2142 : if (RecoveryInProgress() && !InRecovery)
1233 2 : CheckRecoveryConflictDeadlock();
1234 :
1235 : /* Reset deadlock_state before enabling the timeout handler */
1236 2142 : deadlock_state = DS_NOT_YET_CHECKED;
1237 2142 : got_deadlock_timeout = false;
1238 :
1239 : /*
1240 : * Set timer so we can wake up after awhile and check for a deadlock. If a
1241 : * deadlock is detected, the handler sets MyProc->waitStatus =
1242 : * PROC_WAIT_STATUS_ERROR, allowing us to know that we must report failure
1243 : * rather than success.
1244 : *
1245 : * By delaying the check until we've waited for a bit, we can avoid
1246 : * running the rather expensive deadlock-check code in most cases.
1247 : *
1248 : * If LockTimeout is set, also enable the timeout for that. We can save a
1249 : * few cycles by enabling both timeout sources in one call.
1250 : *
1251 : * If InHotStandby we set lock waits slightly later for clarity with other
1252 : * code.
1253 : */
1254 2142 : if (!InHotStandby)
1255 : {
1256 2140 : if (LockTimeout > 0)
1257 : {
1258 : EnableTimeoutParams timeouts[2];
1259 :
1260 226 : timeouts[0].id = DEADLOCK_TIMEOUT;
1261 226 : timeouts[0].type = TMPARAM_AFTER;
1262 226 : timeouts[0].delay_ms = DeadlockTimeout;
1263 226 : timeouts[1].id = LOCK_TIMEOUT;
1264 226 : timeouts[1].type = TMPARAM_AFTER;
1265 226 : timeouts[1].delay_ms = LockTimeout;
1266 226 : enable_timeouts(timeouts, 2);
1267 : }
1268 : else
1269 1914 : enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1270 :
1271 : /*
1272 : * Use the current time obtained for the deadlock timeout timer as
1273 : * waitStart (i.e., the time when this process started waiting for the
1274 : * lock). Since getting the current time newly can cause overhead, we
1275 : * reuse the already-obtained time to avoid that overhead.
1276 : *
1277 : * Note that waitStart is updated without holding the lock table's
1278 : * partition lock, to avoid the overhead by additional lock
1279 : * acquisition. This can cause "waitstart" in pg_locks to become NULL
1280 : * for a very short period of time after the wait started even though
1281 : * "granted" is false. This is OK in practice because we can assume
1282 : * that users are likely to look at "waitstart" when waiting for the
1283 : * lock for a long time.
1284 : */
1285 2140 : pg_atomic_write_u64(&MyProc->waitStart,
1286 2140 : get_timeout_start_time(DEADLOCK_TIMEOUT));
1287 : }
1288 2 : else if (log_recovery_conflict_waits)
1289 : {
1290 : /*
1291 : * Set the wait start timestamp if logging is enabled and in hot
1292 : * standby.
1293 : */
1294 2 : standbyWaitStart = GetCurrentTimestamp();
1295 : }
1296 :
1297 : /*
1298 : * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1299 : * will not wait. But a set latch does not necessarily mean that the lock
1300 : * is free now, as there are many other sources for latch sets than
1301 : * somebody releasing the lock.
1302 : *
1303 : * We process interrupts whenever the latch has been set, so cancel/die
1304 : * interrupts are processed quickly. This means we must not mind losing
1305 : * control to a cancel/die interrupt here. We don't, because we have no
1306 : * shared-state-change work to do after being granted the lock (the
1307 : * grantor did it all). We do have to worry about canceling the deadlock
1308 : * timeout and updating the locallock table, but if we lose control to an
1309 : * error, LockErrorCleanup will fix that up.
1310 : */
1311 : do
1312 : {
1313 5296 : if (InHotStandby)
1314 : {
1315 8 : bool maybe_log_conflict =
1316 8 : (standbyWaitStart != 0 && !logged_recovery_conflict);
1317 :
1318 : /* Set a timer and wait for that or for the lock to be granted */
1319 8 : ResolveRecoveryConflictWithLock(locallock->tag.lock,
1320 : maybe_log_conflict);
1321 :
1322 : /*
1323 : * Emit the log message if the startup process is waiting longer
1324 : * than deadlock_timeout for recovery conflict on lock.
1325 : */
1326 8 : if (maybe_log_conflict)
1327 : {
1328 4 : TimestampTz now = GetCurrentTimestamp();
1329 :
1330 4 : if (TimestampDifferenceExceeds(standbyWaitStart, now,
1331 : DeadlockTimeout))
1332 : {
1333 : VirtualTransactionId *vxids;
1334 : int cnt;
1335 :
1336 2 : vxids = GetLockConflicts(&locallock->tag.lock,
1337 : AccessExclusiveLock, &cnt);
1338 :
1339 : /*
1340 : * Log the recovery conflict and the list of PIDs of
1341 : * backends holding the conflicting lock. Note that we do
1342 : * logging even if there are no such backends right now
1343 : * because the startup process here has already waited
1344 : * longer than deadlock_timeout.
1345 : */
1346 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1347 : standbyWaitStart, now,
1348 2 : cnt > 0 ? vxids : NULL, true);
1349 2 : logged_recovery_conflict = true;
1350 : }
1351 : }
1352 : }
1353 : else
1354 : {
1355 5288 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1356 5288 : PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1357 5288 : ResetLatch(MyLatch);
1358 : /* check for deadlocks first, as that's probably log-worthy */
1359 5288 : if (got_deadlock_timeout)
1360 : {
1361 44 : CheckDeadLock();
1362 44 : got_deadlock_timeout = false;
1363 : }
1364 5288 : CHECK_FOR_INTERRUPTS();
1365 : }
1366 :
1367 : /*
1368 : * waitStatus could change from PROC_WAIT_STATUS_WAITING to something
1369 : * else asynchronously. Read it just once per loop to prevent
1370 : * surprising behavior (such as missing log messages).
1371 : */
1372 5214 : myWaitStatus = *((volatile ProcWaitStatus *) &MyProc->waitStatus);
1373 :
1374 : /*
1375 : * If we are not deadlocked, but are waiting on an autovacuum-induced
1376 : * task, send a signal to interrupt it.
1377 : */
1378 5214 : if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1379 : {
1380 0 : PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1381 : uint8 statusFlags;
1382 : uint8 lockmethod_copy;
1383 : LOCKTAG locktag_copy;
1384 :
1385 : /*
1386 : * Grab info we need, then release lock immediately. Note this
1387 : * coding means that there is a tiny chance that the process
1388 : * terminates its current transaction and starts a different one
1389 : * before we have a change to send the signal; the worst possible
1390 : * consequence is that a for-wraparound vacuum is canceled. But
1391 : * that could happen in any case unless we were to do kill() with
1392 : * the lock held, which is much more undesirable.
1393 : */
1394 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1395 0 : statusFlags = ProcGlobal->statusFlags[autovac->pgxactoff];
1396 0 : lockmethod_copy = lock->tag.locktag_lockmethodid;
1397 0 : locktag_copy = lock->tag;
1398 0 : LWLockRelease(ProcArrayLock);
1399 :
1400 : /*
1401 : * Only do it if the worker is not working to protect against Xid
1402 : * wraparound.
1403 : */
1404 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
1405 0 : !(statusFlags & PROC_VACUUM_FOR_WRAPAROUND))
1406 : {
1407 0 : int pid = autovac->pid;
1408 :
1409 : /* report the case, if configured to do so */
1410 0 : if (message_level_is_interesting(DEBUG1))
1411 : {
1412 : StringInfoData locktagbuf;
1413 : StringInfoData logbuf; /* errdetail for server log */
1414 :
1415 0 : initStringInfo(&locktagbuf);
1416 0 : initStringInfo(&logbuf);
1417 0 : DescribeLockTag(&locktagbuf, &locktag_copy);
1418 0 : appendStringInfo(&logbuf,
1419 : "Process %d waits for %s on %s.",
1420 : MyProcPid,
1421 : GetLockmodeName(lockmethod_copy, lockmode),
1422 : locktagbuf.data);
1423 :
1424 0 : ereport(DEBUG1,
1425 : (errmsg_internal("sending cancel to blocking autovacuum PID %d",
1426 : pid),
1427 : errdetail_log("%s", logbuf.data)));
1428 :
1429 0 : pfree(locktagbuf.data);
1430 0 : pfree(logbuf.data);
1431 : }
1432 :
1433 : /* send the autovacuum worker Back to Old Kent Road */
1434 0 : if (kill(pid, SIGINT) < 0)
1435 : {
1436 : /*
1437 : * There's a race condition here: once we release the
1438 : * ProcArrayLock, it's possible for the autovac worker to
1439 : * close up shop and exit before we can do the kill().
1440 : * Therefore, we do not whinge about no-such-process.
1441 : * Other errors such as EPERM could conceivably happen if
1442 : * the kernel recycles the PID fast enough, but such cases
1443 : * seem improbable enough that it's probably best to issue
1444 : * a warning if we see some other errno.
1445 : */
1446 0 : if (errno != ESRCH)
1447 0 : ereport(WARNING,
1448 : (errmsg("could not send signal to process %d: %m",
1449 : pid)));
1450 : }
1451 : }
1452 :
1453 : /* prevent signal from being sent again more than once */
1454 0 : allow_autovacuum_cancel = false;
1455 : }
1456 :
1457 : /*
1458 : * If awoken after the deadlock check interrupt has run, and
1459 : * log_lock_waits is on, then report about the wait.
1460 : */
1461 5214 : if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1462 : {
1463 : StringInfoData buf,
1464 : lock_waiters_sbuf,
1465 : lock_holders_sbuf;
1466 : const char *modename;
1467 : long secs;
1468 : int usecs;
1469 : long msecs;
1470 : dlist_iter proc_iter;
1471 : PROCLOCK *curproclock;
1472 28 : bool first_holder = true,
1473 28 : first_waiter = true;
1474 28 : int lockHoldersNum = 0;
1475 :
1476 28 : initStringInfo(&buf);
1477 28 : initStringInfo(&lock_waiters_sbuf);
1478 28 : initStringInfo(&lock_holders_sbuf);
1479 :
1480 28 : DescribeLockTag(&buf, &locallock->tag.lock);
1481 28 : modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1482 : lockmode);
1483 28 : TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1484 : GetCurrentTimestamp(),
1485 : &secs, &usecs);
1486 28 : msecs = secs * 1000 + usecs / 1000;
1487 28 : usecs = usecs % 1000;
1488 :
1489 : /*
1490 : * we loop over the lock's procLocks to gather a list of all
1491 : * holders and waiters. Thus we will be able to provide more
1492 : * detailed information for lock debugging purposes.
1493 : *
1494 : * lock->procLocks contains all processes which hold or wait for
1495 : * this lock.
1496 : */
1497 :
1498 28 : LWLockAcquire(partitionLock, LW_SHARED);
1499 :
1500 98 : dlist_foreach(proc_iter, &lock->procLocks)
1501 : {
1502 70 : curproclock =
1503 70 : dlist_container(PROCLOCK, lockLink, proc_iter.cur);
1504 :
1505 : /*
1506 : * we are a waiter if myProc->waitProcLock == curproclock; we
1507 : * are a holder if it is NULL or something different
1508 : */
1509 70 : if (curproclock->tag.myProc->waitProcLock == curproclock)
1510 : {
1511 34 : if (first_waiter)
1512 : {
1513 16 : appendStringInfo(&lock_waiters_sbuf, "%d",
1514 16 : curproclock->tag.myProc->pid);
1515 16 : first_waiter = false;
1516 : }
1517 : else
1518 18 : appendStringInfo(&lock_waiters_sbuf, ", %d",
1519 18 : curproclock->tag.myProc->pid);
1520 : }
1521 : else
1522 : {
1523 36 : if (first_holder)
1524 : {
1525 28 : appendStringInfo(&lock_holders_sbuf, "%d",
1526 28 : curproclock->tag.myProc->pid);
1527 28 : first_holder = false;
1528 : }
1529 : else
1530 8 : appendStringInfo(&lock_holders_sbuf, ", %d",
1531 8 : curproclock->tag.myProc->pid);
1532 :
1533 36 : lockHoldersNum++;
1534 : }
1535 : }
1536 :
1537 28 : LWLockRelease(partitionLock);
1538 :
1539 28 : if (deadlock_state == DS_SOFT_DEADLOCK)
1540 6 : ereport(LOG,
1541 : (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1542 : MyProcPid, modename, buf.data, msecs, usecs),
1543 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1544 : "Processes holding the lock: %s. Wait queue: %s.",
1545 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1546 22 : else if (deadlock_state == DS_HARD_DEADLOCK)
1547 : {
1548 : /*
1549 : * This message is a bit redundant with the error that will be
1550 : * reported subsequently, but in some cases the error report
1551 : * might not make it to the log (eg, if it's caught by an
1552 : * exception handler), and we want to ensure all long-wait
1553 : * events get logged.
1554 : */
1555 2 : ereport(LOG,
1556 : (errmsg("process %d detected deadlock while waiting for %s on %s after %ld.%03d ms",
1557 : MyProcPid, modename, buf.data, msecs, usecs),
1558 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1559 : "Processes holding the lock: %s. Wait queue: %s.",
1560 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1561 : }
1562 :
1563 28 : if (myWaitStatus == PROC_WAIT_STATUS_WAITING)
1564 14 : ereport(LOG,
1565 : (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1566 : MyProcPid, modename, buf.data, msecs, usecs),
1567 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1568 : "Processes holding the lock: %s. Wait queue: %s.",
1569 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1570 14 : else if (myWaitStatus == PROC_WAIT_STATUS_OK)
1571 12 : ereport(LOG,
1572 : (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1573 : MyProcPid, modename, buf.data, msecs, usecs)));
1574 : else
1575 : {
1576 : Assert(myWaitStatus == PROC_WAIT_STATUS_ERROR);
1577 :
1578 : /*
1579 : * Currently, the deadlock checker always kicks its own
1580 : * process, which means that we'll only see
1581 : * PROC_WAIT_STATUS_ERROR when deadlock_state ==
1582 : * DS_HARD_DEADLOCK, and there's no need to print redundant
1583 : * messages. But for completeness and future-proofing, print
1584 : * a message if it looks like someone else kicked us off the
1585 : * lock.
1586 : */
1587 2 : if (deadlock_state != DS_HARD_DEADLOCK)
1588 0 : ereport(LOG,
1589 : (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1590 : MyProcPid, modename, buf.data, msecs, usecs),
1591 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1592 : "Processes holding the lock: %s. Wait queue: %s.",
1593 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1594 : }
1595 :
1596 : /*
1597 : * At this point we might still need to wait for the lock. Reset
1598 : * state so we don't print the above messages again.
1599 : */
1600 28 : deadlock_state = DS_NO_DEADLOCK;
1601 :
1602 28 : pfree(buf.data);
1603 28 : pfree(lock_holders_sbuf.data);
1604 28 : pfree(lock_waiters_sbuf.data);
1605 : }
1606 5214 : } while (myWaitStatus == PROC_WAIT_STATUS_WAITING);
1607 :
1608 : /*
1609 : * Disable the timers, if they are still running. As in LockErrorCleanup,
1610 : * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1611 : * already caused QueryCancelPending to become set, we want the cancel to
1612 : * be reported as a lock timeout, not a user cancel.
1613 : */
1614 2060 : if (!InHotStandby)
1615 : {
1616 2058 : if (LockTimeout > 0)
1617 : {
1618 : DisableTimeoutParams timeouts[2];
1619 :
1620 214 : timeouts[0].id = DEADLOCK_TIMEOUT;
1621 214 : timeouts[0].keep_indicator = false;
1622 214 : timeouts[1].id = LOCK_TIMEOUT;
1623 214 : timeouts[1].keep_indicator = true;
1624 214 : disable_timeouts(timeouts, 2);
1625 : }
1626 : else
1627 1844 : disable_timeout(DEADLOCK_TIMEOUT, false);
1628 : }
1629 :
1630 : /*
1631 : * Emit the log message if recovery conflict on lock was resolved but the
1632 : * startup process waited longer than deadlock_timeout for it.
1633 : */
1634 2060 : if (InHotStandby && logged_recovery_conflict)
1635 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1636 : standbyWaitStart, GetCurrentTimestamp(),
1637 : NULL, false);
1638 :
1639 : /*
1640 : * Re-acquire the lock table's partition lock. We have to do this to hold
1641 : * off cancel/die interrupts before we can mess with lockAwaited (else we
1642 : * might have a missed or duplicated locallock update).
1643 : */
1644 2060 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
1645 :
1646 : /*
1647 : * We no longer want LockErrorCleanup to do anything.
1648 : */
1649 2060 : lockAwaited = NULL;
1650 :
1651 : /*
1652 : * If we got the lock, be sure to remember it in the locallock table.
1653 : */
1654 2060 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
1655 2052 : GrantAwaitedLock();
1656 :
1657 : /*
1658 : * We don't have to do anything else, because the awaker did all the
1659 : * necessary update of the lock table and MyProc.
1660 : */
1661 2060 : return MyProc->waitStatus;
1662 : }
1663 :
1664 :
1665 : /*
1666 : * ProcWakeup -- wake up a process by setting its latch.
1667 : *
1668 : * Also remove the process from the wait queue and set its links invalid.
1669 : *
1670 : * The appropriate lock partition lock must be held by caller.
1671 : *
1672 : * XXX: presently, this code is only used for the "success" case, and only
1673 : * works correctly for that case. To clean up in failure case, would need
1674 : * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1675 : * Hence, in practice the waitStatus parameter must be PROC_WAIT_STATUS_OK.
1676 : */
1677 : void
1678 2060 : ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus)
1679 : {
1680 2060 : if (dlist_node_is_detached(&proc->links))
1681 0 : return;
1682 :
1683 : Assert(proc->waitStatus == PROC_WAIT_STATUS_WAITING);
1684 :
1685 : /* Remove process from wait queue */
1686 2060 : dclist_delete_from_thoroughly(&proc->waitLock->waitProcs, &proc->links);
1687 :
1688 : /* Clean up process' state and pass it the ok/fail signal */
1689 2060 : proc->waitLock = NULL;
1690 2060 : proc->waitProcLock = NULL;
1691 2060 : proc->waitStatus = waitStatus;
1692 2060 : pg_atomic_write_u64(&MyProc->waitStart, 0);
1693 :
1694 : /* And awaken it */
1695 2060 : SetLatch(&proc->procLatch);
1696 : }
1697 :
1698 : /*
1699 : * ProcLockWakeup -- routine for waking up processes when a lock is
1700 : * released (or a prior waiter is aborted). Scan all waiters
1701 : * for lock, waken any that are no longer blocked.
1702 : *
1703 : * The appropriate lock partition lock must be held by caller.
1704 : */
1705 : void
1706 2088 : ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1707 : {
1708 2088 : dclist_head *waitQueue = &lock->waitProcs;
1709 2088 : LOCKMASK aheadRequests = 0;
1710 : dlist_mutable_iter miter;
1711 :
1712 2088 : if (dclist_is_empty(waitQueue))
1713 90 : return;
1714 :
1715 4206 : dclist_foreach_modify(miter, waitQueue)
1716 : {
1717 2208 : PGPROC *proc = dlist_container(PGPROC, links, miter.cur);
1718 2208 : LOCKMODE lockmode = proc->waitLockMode;
1719 :
1720 : /*
1721 : * Waken if (a) doesn't conflict with requests of earlier waiters, and
1722 : * (b) doesn't conflict with already-held locks.
1723 : */
1724 2208 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1725 2146 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1726 : proc->waitProcLock))
1727 : {
1728 : /* OK to waken */
1729 2060 : GrantLock(lock, proc->waitProcLock, lockmode);
1730 : /* removes proc from the lock's waiting process queue */
1731 2060 : ProcWakeup(proc, PROC_WAIT_STATUS_OK);
1732 : }
1733 : else
1734 : {
1735 : /*
1736 : * Lock conflicts: Don't wake, but remember requested mode for
1737 : * later checks.
1738 : */
1739 148 : aheadRequests |= LOCKBIT_ON(lockmode);
1740 : }
1741 : }
1742 : }
1743 :
1744 : /*
1745 : * CheckDeadLock
1746 : *
1747 : * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1748 : * lock to be released by some other process. Check if there's a deadlock; if
1749 : * not, just return. (But signal ProcSleep to log a message, if
1750 : * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1751 : * the lock's wait queue and signal an error to ProcSleep.
1752 : */
1753 : static void
1754 44 : CheckDeadLock(void)
1755 : {
1756 : int i;
1757 :
1758 : /*
1759 : * Acquire exclusive lock on the entire shared lock data structures. Must
1760 : * grab LWLocks in partition-number order to avoid LWLock deadlock.
1761 : *
1762 : * Note that the deadlock check interrupt had better not be enabled
1763 : * anywhere that this process itself holds lock partition locks, else this
1764 : * will wait forever. Also note that LWLockAcquire creates a critical
1765 : * section, so that this routine cannot be interrupted by cancel/die
1766 : * interrupts.
1767 : */
1768 748 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1769 704 : LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1770 :
1771 : /*
1772 : * Check to see if we've been awoken by anyone in the interim.
1773 : *
1774 : * If we have, we can return and resume our transaction -- happy day.
1775 : * Before we are awoken the process releasing the lock grants it to us so
1776 : * we know that we don't have to wait anymore.
1777 : *
1778 : * We check by looking to see if we've been unlinked from the wait queue.
1779 : * This is safe because we hold the lock partition lock.
1780 : */
1781 44 : if (MyProc->links.prev == NULL ||
1782 44 : MyProc->links.next == NULL)
1783 0 : goto check_done;
1784 :
1785 : #ifdef LOCK_DEBUG
1786 : if (Debug_deadlocks)
1787 : DumpAllLocks();
1788 : #endif
1789 :
1790 : /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1791 44 : deadlock_state = DeadLockCheck(MyProc);
1792 :
1793 44 : if (deadlock_state == DS_HARD_DEADLOCK)
1794 : {
1795 : /*
1796 : * Oops. We have a deadlock.
1797 : *
1798 : * Get this process out of wait state. (Note: we could do this more
1799 : * efficiently by relying on lockAwaited, but use this coding to
1800 : * preserve the flexibility to kill some other transaction than the
1801 : * one detecting the deadlock.)
1802 : *
1803 : * RemoveFromWaitQueue sets MyProc->waitStatus to
1804 : * PROC_WAIT_STATUS_ERROR, so ProcSleep will report an error after we
1805 : * return from the signal handler.
1806 : */
1807 : Assert(MyProc->waitLock != NULL);
1808 8 : RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1809 :
1810 : /*
1811 : * We're done here. Transaction abort caused by the error that
1812 : * ProcSleep will raise will cause any other locks we hold to be
1813 : * released, thus allowing other processes to wake up; we don't need
1814 : * to do that here. NOTE: an exception is that releasing locks we
1815 : * hold doesn't consider the possibility of waiters that were blocked
1816 : * behind us on the lock we just failed to get, and might now be
1817 : * wakable because we're not in front of them anymore. However,
1818 : * RemoveFromWaitQueue took care of waking up any such processes.
1819 : */
1820 : }
1821 :
1822 : /*
1823 : * And release locks. We do this in reverse order for two reasons: (1)
1824 : * Anyone else who needs more than one of the locks will be trying to lock
1825 : * them in increasing order; we don't want to release the other process
1826 : * until it can get all the locks it needs. (2) This avoids O(N^2)
1827 : * behavior inside LWLockRelease.
1828 : */
1829 36 : check_done:
1830 748 : for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1831 704 : LWLockRelease(LockHashPartitionLockByIndex(i));
1832 44 : }
1833 :
1834 : /*
1835 : * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1836 : *
1837 : * NB: Runs inside a signal handler, be careful.
1838 : */
1839 : void
1840 44 : CheckDeadLockAlert(void)
1841 : {
1842 44 : int save_errno = errno;
1843 :
1844 44 : got_deadlock_timeout = true;
1845 :
1846 : /*
1847 : * Have to set the latch again, even if handle_sig_alarm already did. Back
1848 : * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1849 : * ever would be a problem, but setting a set latch again is cheap.
1850 : *
1851 : * Note that, when this function runs inside procsignal_sigusr1_handler(),
1852 : * the handler function sets the latch again after the latch is set here.
1853 : */
1854 44 : SetLatch(MyLatch);
1855 44 : errno = save_errno;
1856 44 : }
1857 :
1858 : /*
1859 : * ProcWaitForSignal - wait for a signal from another backend.
1860 : *
1861 : * As this uses the generic process latch the caller has to be robust against
1862 : * unrelated wakeups: Always check that the desired state has occurred, and
1863 : * wait again if not.
1864 : */
1865 : void
1866 34 : ProcWaitForSignal(uint32 wait_event_info)
1867 : {
1868 34 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1869 : wait_event_info);
1870 34 : ResetLatch(MyLatch);
1871 34 : CHECK_FOR_INTERRUPTS();
1872 34 : }
1873 :
1874 : /*
1875 : * ProcSendSignal - set the latch of a backend identified by ProcNumber
1876 : */
1877 : void
1878 6 : ProcSendSignal(ProcNumber procNumber)
1879 : {
1880 6 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
1881 0 : elog(ERROR, "procNumber out of range");
1882 :
1883 6 : SetLatch(&ProcGlobal->allProcs[procNumber].procLatch);
1884 6 : }
1885 :
1886 : /*
1887 : * BecomeLockGroupLeader - designate process as lock group leader
1888 : *
1889 : * Once this function has returned, other processes can join the lock group
1890 : * by calling BecomeLockGroupMember.
1891 : */
1892 : void
1893 1086 : BecomeLockGroupLeader(void)
1894 : {
1895 : LWLock *leader_lwlock;
1896 :
1897 : /* If we already did it, we don't need to do it again. */
1898 1086 : if (MyProc->lockGroupLeader == MyProc)
1899 970 : return;
1900 :
1901 : /* We had better not be a follower. */
1902 : Assert(MyProc->lockGroupLeader == NULL);
1903 :
1904 : /* Create single-member group, containing only ourselves. */
1905 116 : leader_lwlock = LockHashPartitionLockByProc(MyProc);
1906 116 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1907 116 : MyProc->lockGroupLeader = MyProc;
1908 116 : dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
1909 116 : LWLockRelease(leader_lwlock);
1910 : }
1911 :
1912 : /*
1913 : * BecomeLockGroupMember - designate process as lock group member
1914 : *
1915 : * This is pretty straightforward except for the possibility that the leader
1916 : * whose group we're trying to join might exit before we manage to do so;
1917 : * and the PGPROC might get recycled for an unrelated process. To avoid
1918 : * that, we require the caller to pass the PID of the intended PGPROC as
1919 : * an interlock. Returns true if we successfully join the intended lock
1920 : * group, and false if not.
1921 : */
1922 : bool
1923 2642 : BecomeLockGroupMember(PGPROC *leader, int pid)
1924 : {
1925 : LWLock *leader_lwlock;
1926 2642 : bool ok = false;
1927 :
1928 : /* Group leader can't become member of group */
1929 : Assert(MyProc != leader);
1930 :
1931 : /* Can't already be a member of a group */
1932 : Assert(MyProc->lockGroupLeader == NULL);
1933 :
1934 : /* PID must be valid. */
1935 : Assert(pid != 0);
1936 :
1937 : /*
1938 : * Get lock protecting the group fields. Note LockHashPartitionLockByProc
1939 : * calculates the proc number based on the PGPROC slot without looking at
1940 : * its contents, so we will acquire the correct lock even if the leader
1941 : * PGPROC is in process of being recycled.
1942 : */
1943 2642 : leader_lwlock = LockHashPartitionLockByProc(leader);
1944 2642 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1945 :
1946 : /* Is this the leader we're looking for? */
1947 2642 : if (leader->pid == pid && leader->lockGroupLeader == leader)
1948 : {
1949 : /* OK, join the group */
1950 2642 : ok = true;
1951 2642 : MyProc->lockGroupLeader = leader;
1952 2642 : dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
1953 : }
1954 2642 : LWLockRelease(leader_lwlock);
1955 :
1956 2642 : return ok;
1957 : }
|
