1/*-------------------------------------------------------------------------
2 *
3 * proc.c
4 * routines to manage per-process shared memory data structure
5 *
6 * Portions Copyright (c) 1996-2019, 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 * ProcQueueAlloc() -- create a shm queue for sleeping processes
19 * ProcQueueInit() -- create a queue without allocing memory
20 *
21 * Waiting for a lock causes the backend to be put to sleep. Whoever releases
22 * the lock wakes the process up again (and gives it an error code so it knows
23 * whether it was awoken on an error condition).
24 *
25 * Interface (b):
26 *
27 * ProcReleaseLocks -- frees the locks associated with current transaction
28 *
29 * ProcKill -- destroys the shared memory state (and locks)
30 * associated with the process.
31 */
32#include "postgres.h"
33
34#include <signal.h>
35#include <unistd.h>
36#include <sys/time.h>
37
38#include "access/transam.h"
39#include "access/twophase.h"
40#include "access/xact.h"
41#include "miscadmin.h"
42#include "pgstat.h"
43#include "postmaster/autovacuum.h"
44#include "replication/slot.h"
45#include "replication/syncrep.h"
46#include "replication/walsender.h"
47#include "storage/condition_variable.h"
48#include "storage/standby.h"
49#include "storage/ipc.h"
50#include "storage/lmgr.h"
51#include "storage/pmsignal.h"
52#include "storage/proc.h"
53#include "storage/procarray.h"
54#include "storage/procsignal.h"
55#include "storage/spin.h"
56#include "utils/timeout.h"
57#include "utils/timestamp.h"
58
59
60/* GUC variables */
61int DeadlockTimeout = 1000;
62int StatementTimeout = 0;
63int LockTimeout = 0;
64int IdleInTransactionSessionTimeout = 0;
65bool log_lock_waits = false;
66
67/* Pointer to this process's PGPROC and PGXACT structs, if any */
68PGPROC *MyProc = NULL;
69PGXACT *MyPgXact = NULL;
70
71/*
72 * This spinlock protects the freelist of recycled PGPROC structures.
73 * We cannot use an LWLock because the LWLock manager depends on already
74 * having a PGPROC and a wait semaphore! But these structures are touched
75 * relatively infrequently (only at backend startup or shutdown) and not for
76 * very long, so a spinlock is okay.
77 */
78NON_EXEC_STATIC slock_t *ProcStructLock = NULL;
79
80/* Pointers to shared-memory structures */
81PROC_HDR *ProcGlobal = NULL;
82NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
83PGPROC *PreparedXactProcs = NULL;
84
85/* If we are waiting for a lock, this points to the associated LOCALLOCK */
86static LOCALLOCK *lockAwaited = NULL;
87
88static DeadLockState deadlock_state = DS_NOT_YET_CHECKED;
89
90/* Is a deadlock check pending? */
91static volatile sig_atomic_t got_deadlock_timeout;
92
93static void RemoveProcFromArray(int code, Datum arg);
94static void ProcKill(int code, Datum arg);
95static void AuxiliaryProcKill(int code, Datum arg);
96static void CheckDeadLock(void);
97
98
99/*
100 * Report shared-memory space needed by InitProcGlobal.
101 */
102Size
103ProcGlobalShmemSize(void)
104{
105 Size size = 0;
106
107 /* ProcGlobal */
108 size = add_size(size, sizeof(PROC_HDR));
109 /* MyProcs, including autovacuum workers and launcher */
110 size = add_size(size, mul_size(MaxBackends, sizeof(PGPROC)));
111 /* AuxiliaryProcs */
112 size = add_size(size, mul_size(NUM_AUXILIARY_PROCS, sizeof(PGPROC)));
113 /* Prepared xacts */
114 size = add_size(size, mul_size(max_prepared_xacts, sizeof(PGPROC)));
115 /* ProcStructLock */
116 size = add_size(size, sizeof(slock_t));
117
118 size = add_size(size, mul_size(MaxBackends, sizeof(PGXACT)));
119 size = add_size(size, mul_size(NUM_AUXILIARY_PROCS, sizeof(PGXACT)));
120 size = add_size(size, mul_size(max_prepared_xacts, sizeof(PGXACT)));
121
122 return size;
123}
124
125/*
126 * Report number of semaphores needed by InitProcGlobal.
127 */
128int
129ProcGlobalSemas(void)
130{
131 /*
132 * We need a sema per backend (including autovacuum), plus one for each
133 * auxiliary process.
134 */
135 return MaxBackends + NUM_AUXILIARY_PROCS;
136}
137
138/*
139 * InitProcGlobal -
140 * Initialize the global process table during postmaster or standalone
141 * backend startup.
142 *
143 * We also create all the per-process semaphores we will need to support
144 * the requested number of backends. We used to allocate semaphores
145 * only when backends were actually started up, but that is bad because
146 * it lets Postgres fail under load --- a lot of Unix systems are
147 * (mis)configured with small limits on the number of semaphores, and
148 * running out when trying to start another backend is a common failure.
149 * So, now we grab enough semaphores to support the desired max number
150 * of backends immediately at initialization --- if the sysadmin has set
151 * MaxConnections, max_worker_processes, max_wal_senders, or
152 * autovacuum_max_workers higher than his kernel will support, he'll
153 * find out sooner rather than later.
154 *
155 * Another reason for creating semaphores here is that the semaphore
156 * implementation typically requires us to create semaphores in the
157 * postmaster, not in backends.
158 *
159 * Note: this is NOT called by individual backends under a postmaster,
160 * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
161 * pointers must be propagated specially for EXEC_BACKEND operation.
162 */
163void
164InitProcGlobal(void)
165{
166 PGPROC *procs;
167 PGXACT *pgxacts;
168 int i,
169 j;
170 bool found;
171 uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
172
173 /* Create the ProcGlobal shared structure */
174 ProcGlobal = (PROC_HDR *)
175 ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
176 Assert(!found);
177
178 /*
179 * Initialize the data structures.
180 */
181 ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
182 ProcGlobal->freeProcs = NULL;
183 ProcGlobal->autovacFreeProcs = NULL;
184 ProcGlobal->bgworkerFreeProcs = NULL;
185 ProcGlobal->walsenderFreeProcs = NULL;
186 ProcGlobal->startupProc = NULL;
187 ProcGlobal->startupProcPid = 0;
188 ProcGlobal->startupBufferPinWaitBufId = -1;
189 ProcGlobal->walwriterLatch = NULL;
190 ProcGlobal->checkpointerLatch = NULL;
191 pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PGPROCNO);
192 pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PGPROCNO);
193
194 /*
195 * Create and initialize all the PGPROC structures we'll need. There are
196 * five separate consumers: (1) normal backends, (2) autovacuum workers
197 * and the autovacuum launcher, (3) background workers, (4) auxiliary
198 * processes, and (5) prepared transactions. Each PGPROC structure is
199 * dedicated to exactly one of these purposes, and they do not move
200 * between groups.
201 */
202 procs = (PGPROC *) ShmemAlloc(TotalProcs * sizeof(PGPROC));
203 MemSet(procs, 0, TotalProcs * sizeof(PGPROC));
204 ProcGlobal->allProcs = procs;
205 /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
206 ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
207
208 /*
209 * Also allocate a separate array of PGXACT structures. This is separate
210 * from the main PGPROC array so that the most heavily accessed data is
211 * stored contiguously in memory in as few cache lines as possible. This
212 * provides significant performance benefits, especially on a
213 * multiprocessor system. There is one PGXACT structure for every PGPROC
214 * structure.
215 */
216 pgxacts = (PGXACT *) ShmemAlloc(TotalProcs * sizeof(PGXACT));
217 MemSet(pgxacts, 0, TotalProcs * sizeof(PGXACT));
218 ProcGlobal->allPgXact = pgxacts;
219
220 for (i = 0; i < TotalProcs; i++)
221 {
222 /* Common initialization for all PGPROCs, regardless of type. */
223
224 /*
225 * Set up per-PGPROC semaphore, latch, and backendLock. Prepared xact
226 * dummy PGPROCs don't need these though - they're never associated
227 * with a real process
228 */
229 if (i < MaxBackends + NUM_AUXILIARY_PROCS)
230 {
231 procs[i].sem = PGSemaphoreCreate();
232 InitSharedLatch(&(procs[i].procLatch));
233 LWLockInitialize(&(procs[i].backendLock), LWTRANCHE_PROC);
234 }
235 procs[i].pgprocno = i;
236
237 /*
238 * Newly created PGPROCs for normal backends, autovacuum and bgworkers
239 * must be queued up on the appropriate free list. Because there can
240 * only ever be a small, fixed number of auxiliary processes, no free
241 * list is used in that case; InitAuxiliaryProcess() instead uses a
242 * linear search. PGPROCs for prepared transactions are added to a
243 * free list by TwoPhaseShmemInit().
244 */
245 if (i < MaxConnections)
246 {
247 /* PGPROC for normal backend, add to freeProcs list */
248 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->freeProcs;
249 ProcGlobal->freeProcs = &procs[i];
250 procs[i].procgloballist = &ProcGlobal->freeProcs;
251 }
252 else if (i < MaxConnections + autovacuum_max_workers + 1)
253 {
254 /* PGPROC for AV launcher/worker, add to autovacFreeProcs list */
255 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->autovacFreeProcs;
256 ProcGlobal->autovacFreeProcs = &procs[i];
257 procs[i].procgloballist = &ProcGlobal->autovacFreeProcs;
258 }
259 else if (i < MaxConnections + autovacuum_max_workers + 1 + max_worker_processes)
260 {
261 /* PGPROC for bgworker, add to bgworkerFreeProcs list */
262 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->bgworkerFreeProcs;
263 ProcGlobal->bgworkerFreeProcs = &procs[i];
264 procs[i].procgloballist = &ProcGlobal->bgworkerFreeProcs;
265 }
266 else if (i < MaxBackends)
267 {
268 /* PGPROC for walsender, add to walsenderFreeProcs list */
269 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->walsenderFreeProcs;
270 ProcGlobal->walsenderFreeProcs = &procs[i];
271 procs[i].procgloballist = &ProcGlobal->walsenderFreeProcs;
272 }
273
274 /* Initialize myProcLocks[] shared memory queues. */
275 for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
276 SHMQueueInit(&(procs[i].myProcLocks[j]));
277
278 /* Initialize lockGroupMembers list. */
279 dlist_init(&procs[i].lockGroupMembers);
280
281 /*
282 * Initialize the atomic variables, otherwise, it won't be safe to
283 * access them for backends that aren't currently in use.
284 */
285 pg_atomic_init_u32(&(procs[i].procArrayGroupNext), INVALID_PGPROCNO);
286 pg_atomic_init_u32(&(procs[i].clogGroupNext), INVALID_PGPROCNO);
287 }
288
289 /*
290 * Save pointers to the blocks of PGPROC structures reserved for auxiliary
291 * processes and prepared transactions.
292 */
293 AuxiliaryProcs = &procs[MaxBackends];
294 PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
295
296 /* Create ProcStructLock spinlock, too */
297 ProcStructLock = (slock_t *) ShmemAlloc(sizeof(slock_t));
298 SpinLockInit(ProcStructLock);
299}
300
301/*
302 * InitProcess -- initialize a per-process data structure for this backend
303 */
304void
305InitProcess(void)
306{
307 PGPROC *volatile *procgloballist;
308
309 /*
310 * ProcGlobal should be set up already (if we are a backend, we inherit
311 * this by fork() or EXEC_BACKEND mechanism from the postmaster).
312 */
313 if (ProcGlobal == NULL)
314 elog(PANIC, "proc header uninitialized");
315
316 if (MyProc != NULL)
317 elog(ERROR, "you already exist");
318
319 /* Decide which list should supply our PGPROC. */
320 if (IsAnyAutoVacuumProcess())
321 procgloballist = &ProcGlobal->autovacFreeProcs;
322 else if (IsBackgroundWorker)
323 procgloballist = &ProcGlobal->bgworkerFreeProcs;
324 else if (am_walsender)
325 procgloballist = &ProcGlobal->walsenderFreeProcs;
326 else
327 procgloballist = &ProcGlobal->freeProcs;
328
329 /*
330 * Try to get a proc struct from the appropriate free list. If this
331 * fails, we must be out of PGPROC structures (not to mention semaphores).
332 *
333 * While we are holding the ProcStructLock, also copy the current shared
334 * estimate of spins_per_delay to local storage.
335 */
336 SpinLockAcquire(ProcStructLock);
337
338 set_spins_per_delay(ProcGlobal->spins_per_delay);
339
340 MyProc = *procgloballist;
341
342 if (MyProc != NULL)
343 {
344 *procgloballist = (PGPROC *) MyProc->links.next;
345 SpinLockRelease(ProcStructLock);
346 }
347 else
348 {
349 /*
350 * If we reach here, all the PGPROCs are in use. This is one of the
351 * possible places to detect "too many backends", so give the standard
352 * error message. XXX do we need to give a different failure message
353 * in the autovacuum case?
354 */
355 SpinLockRelease(ProcStructLock);
356 if (am_walsender)
357 ereport(FATAL,
358 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
359 errmsg("number of requested standby connections exceeds max_wal_senders (currently %d)",
360 max_wal_senders)));
361 ereport(FATAL,
362 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
363 errmsg("sorry, too many clients already")));
364 }
365 MyPgXact = &ProcGlobal->allPgXact[MyProc->pgprocno];
366
367 /*
368 * Cross-check that the PGPROC is of the type we expect; if this were not
369 * the case, it would get returned to the wrong list.
370 */
371 Assert(MyProc->procgloballist == procgloballist);
372
373 /*
374 * Now that we have a PGPROC, mark ourselves as an active postmaster
375 * child; this is so that the postmaster can detect it if we exit without
376 * cleaning up. (XXX autovac launcher currently doesn't participate in
377 * this; it probably should.)
378 */
379 if (IsUnderPostmaster && !IsAutoVacuumLauncherProcess())
380 MarkPostmasterChildActive();
381
382 /*
383 * Initialize all fields of MyProc, except for those previously
384 * initialized by InitProcGlobal.
385 */
386 SHMQueueElemInit(&(MyProc->links));
387 MyProc->waitStatus = STATUS_OK;
388 MyProc->lxid = InvalidLocalTransactionId;
389 MyProc->fpVXIDLock = false;
390 MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
391 MyPgXact->xid = InvalidTransactionId;
392 MyPgXact->xmin = InvalidTransactionId;
393 MyProc->pid = MyProcPid;
394 /* backendId, databaseId and roleId will be filled in later */
395 MyProc->backendId = InvalidBackendId;
396 MyProc->databaseId = InvalidOid;
397 MyProc->roleId = InvalidOid;
398 MyProc->tempNamespaceId = InvalidOid;
399 MyProc->isBackgroundWorker = IsBackgroundWorker;
400 MyPgXact->delayChkpt = false;
401 MyPgXact->vacuumFlags = 0;
402 /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
403 if (IsAutoVacuumWorkerProcess())
404 MyPgXact->vacuumFlags |= PROC_IS_AUTOVACUUM;
405 MyProc->lwWaiting = false;
406 MyProc->lwWaitMode = 0;
407 MyProc->waitLock = NULL;
408 MyProc->waitProcLock = NULL;
409#ifdef USE_ASSERT_CHECKING
410 {
411 int i;
412
413 /* Last process should have released all locks. */
414 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
415 Assert(SHMQueueEmpty(&(MyProc->myProcLocks[i])));
416 }
417#endif
418 MyProc->recoveryConflictPending = false;
419
420 /* Initialize fields for sync rep */
421 MyProc->waitLSN = 0;
422 MyProc->syncRepState = SYNC_REP_NOT_WAITING;
423 SHMQueueElemInit(&(MyProc->syncRepLinks));
424
425 /* Initialize fields for group XID clearing. */
426 MyProc->procArrayGroupMember = false;
427 MyProc->procArrayGroupMemberXid = InvalidTransactionId;
428 Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PGPROCNO);
429
430 /* Check that group locking fields are in a proper initial state. */
431 Assert(MyProc->lockGroupLeader == NULL);
432 Assert(dlist_is_empty(&MyProc->lockGroupMembers));
433
434 /* Initialize wait event information. */
435 MyProc->wait_event_info = 0;
436
437 /* Initialize fields for group transaction status update. */
438 MyProc->clogGroupMember = false;
439 MyProc->clogGroupMemberXid = InvalidTransactionId;
440 MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
441 MyProc->clogGroupMemberPage = -1;
442 MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
443 Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PGPROCNO);
444
445 /*
446 * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
447 * on it. That allows us to repoint the process latch, which so far
448 * points to process local one, to the shared one.
449 */
450 OwnLatch(&MyProc->procLatch);
451 SwitchToSharedLatch();
452
453 /*
454 * We might be reusing a semaphore that belonged to a failed process. So
455 * be careful and reinitialize its value here. (This is not strictly
456 * necessary anymore, but seems like a good idea for cleanliness.)
457 */
458 PGSemaphoreReset(MyProc->sem);
459
460 /*
461 * Arrange to clean up at backend exit.
462 */
463 on_shmem_exit(ProcKill, 0);
464
465 /*
466 * Now that we have a PGPROC, we could try to acquire locks, so initialize
467 * local state needed for LWLocks, and the deadlock checker.
468 */
469 InitLWLockAccess();
470 InitDeadLockChecking();
471}
472
473/*
474 * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
475 *
476 * This is separate from InitProcess because we can't acquire LWLocks until
477 * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
478 * work until after we've done CreateSharedMemoryAndSemaphores.
479 */
480void
481InitProcessPhase2(void)
482{
483 Assert(MyProc != NULL);
484
485 /*
486 * Add our PGPROC to the PGPROC array in shared memory.
487 */
488 ProcArrayAdd(MyProc);
489
490 /*
491 * Arrange to clean that up at backend exit.
492 */
493 on_shmem_exit(RemoveProcFromArray, 0);
494}
495
496/*
497 * InitAuxiliaryProcess -- create a per-auxiliary-process data structure
498 *
499 * This is called by bgwriter and similar processes so that they will have a
500 * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
501 * and sema that are assigned are one of the extra ones created during
502 * InitProcGlobal.
503 *
504 * Auxiliary processes are presently not expected to wait for real (lockmgr)
505 * locks, so we need not set up the deadlock checker. They are never added
506 * to the ProcArray or the sinval messaging mechanism, either. They also
507 * don't get a VXID assigned, since this is only useful when we actually
508 * hold lockmgr locks.
509 *
510 * Startup process however uses locks but never waits for them in the
511 * normal backend sense. Startup process also takes part in sinval messaging
512 * as a sendOnly process, so never reads messages from sinval queue. So
513 * Startup process does have a VXID and does show up in pg_locks.
514 */
515void
516InitAuxiliaryProcess(void)
517{
518 PGPROC *auxproc;
519 int proctype;
520
521 /*
522 * ProcGlobal should be set up already (if we are a backend, we inherit
523 * this by fork() or EXEC_BACKEND mechanism from the postmaster).
524 */
525 if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
526 elog(PANIC, "proc header uninitialized");
527
528 if (MyProc != NULL)
529 elog(ERROR, "you already exist");
530
531 /*
532 * We use the ProcStructLock to protect assignment and releasing of
533 * AuxiliaryProcs entries.
534 *
535 * While we are holding the ProcStructLock, also copy the current shared
536 * estimate of spins_per_delay to local storage.
537 */
538 SpinLockAcquire(ProcStructLock);
539
540 set_spins_per_delay(ProcGlobal->spins_per_delay);
541
542 /*
543 * Find a free auxproc ... *big* trouble if there isn't one ...
544 */
545 for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
546 {
547 auxproc = &AuxiliaryProcs[proctype];
548 if (auxproc->pid == 0)
549 break;
550 }
551 if (proctype >= NUM_AUXILIARY_PROCS)
552 {
553 SpinLockRelease(ProcStructLock);
554 elog(FATAL, "all AuxiliaryProcs are in use");
555 }
556
557 /* Mark auxiliary proc as in use by me */
558 /* use volatile pointer to prevent code rearrangement */
559 ((volatile PGPROC *) auxproc)->pid = MyProcPid;
560
561 MyProc = auxproc;
562 MyPgXact = &ProcGlobal->allPgXact[auxproc->pgprocno];
563
564 SpinLockRelease(ProcStructLock);
565
566 /*
567 * Initialize all fields of MyProc, except for those previously
568 * initialized by InitProcGlobal.
569 */
570 SHMQueueElemInit(&(MyProc->links));
571 MyProc->waitStatus = STATUS_OK;
572 MyProc->lxid = InvalidLocalTransactionId;
573 MyProc->fpVXIDLock = false;
574 MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
575 MyPgXact->xid = InvalidTransactionId;
576 MyPgXact->xmin = InvalidTransactionId;
577 MyProc->backendId = InvalidBackendId;
578 MyProc->databaseId = InvalidOid;
579 MyProc->roleId = InvalidOid;
580 MyProc->tempNamespaceId = InvalidOid;
581 MyProc->isBackgroundWorker = IsBackgroundWorker;
582 MyPgXact->delayChkpt = false;
583 MyPgXact->vacuumFlags = 0;
584 MyProc->lwWaiting = false;
585 MyProc->lwWaitMode = 0;
586 MyProc->waitLock = NULL;
587 MyProc->waitProcLock = NULL;
588#ifdef USE_ASSERT_CHECKING
589 {
590 int i;
591
592 /* Last process should have released all locks. */
593 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
594 Assert(SHMQueueEmpty(&(MyProc->myProcLocks[i])));
595 }
596#endif
597
598 /*
599 * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
600 * on it. That allows us to repoint the process latch, which so far
601 * points to process local one, to the shared one.
602 */
603 OwnLatch(&MyProc->procLatch);
604 SwitchToSharedLatch();
605
606 /* Check that group locking fields are in a proper initial state. */
607 Assert(MyProc->lockGroupLeader == NULL);
608 Assert(dlist_is_empty(&MyProc->lockGroupMembers));
609
610 /*
611 * We might be reusing a semaphore that belonged to a failed process. So
612 * be careful and reinitialize its value here. (This is not strictly
613 * necessary anymore, but seems like a good idea for cleanliness.)
614 */
615 PGSemaphoreReset(MyProc->sem);
616
617 /*
618 * Arrange to clean up at process exit.
619 */
620 on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
621}
622
623/*
624 * Record the PID and PGPROC structures for the Startup process, for use in
625 * ProcSendSignal(). See comments there for further explanation.
626 */
627void
628PublishStartupProcessInformation(void)
629{
630 SpinLockAcquire(ProcStructLock);
631
632 ProcGlobal->startupProc = MyProc;
633 ProcGlobal->startupProcPid = MyProcPid;
634
635 SpinLockRelease(ProcStructLock);
636}
637
638/*
639 * Used from bufgr to share the value of the buffer that Startup waits on,
640 * or to reset the value to "not waiting" (-1). This allows processing
641 * of recovery conflicts for buffer pins. Set is made before backends look
642 * at this value, so locking not required, especially since the set is
643 * an atomic integer set operation.
644 */
645void
646SetStartupBufferPinWaitBufId(int bufid)
647{
648 /* use volatile pointer to prevent code rearrangement */
649 volatile PROC_HDR *procglobal = ProcGlobal;
650
651 procglobal->startupBufferPinWaitBufId = bufid;
652}
653
654/*
655 * Used by backends when they receive a request to check for buffer pin waits.
656 */
657int
658GetStartupBufferPinWaitBufId(void)
659{
660 /* use volatile pointer to prevent code rearrangement */
661 volatile PROC_HDR *procglobal = ProcGlobal;
662
663 return procglobal->startupBufferPinWaitBufId;
664}
665
666/*
667 * Check whether there are at least N free PGPROC objects.
668 *
669 * Note: this is designed on the assumption that N will generally be small.
670 */
671bool
672HaveNFreeProcs(int n)
673{
674 PGPROC *proc;
675
676 SpinLockAcquire(ProcStructLock);
677
678 proc = ProcGlobal->freeProcs;
679
680 while (n > 0 && proc != NULL)
681 {
682 proc = (PGPROC *) proc->links.next;
683 n--;
684 }
685
686 SpinLockRelease(ProcStructLock);
687
688 return (n <= 0);
689}
690
691/*
692 * Check if the current process is awaiting a lock.
693 */
694bool
695IsWaitingForLock(void)
696{
697 if (lockAwaited == NULL)
698 return false;
699
700 return true;
701}
702
703/*
704 * Cancel any pending wait for lock, when aborting a transaction, and revert
705 * any strong lock count acquisition for a lock being acquired.
706 *
707 * (Normally, this would only happen if we accept a cancel/die
708 * interrupt while waiting; but an ereport(ERROR) before or during the lock
709 * wait is within the realm of possibility, too.)
710 */
711void
712LockErrorCleanup(void)
713{
714 LWLock *partitionLock;
715 DisableTimeoutParams timeouts[2];
716
717 HOLD_INTERRUPTS();
718
719 AbortStrongLockAcquire();
720
721 /* Nothing to do if we weren't waiting for a lock */
722 if (lockAwaited == NULL)
723 {
724 RESUME_INTERRUPTS();
725 return;
726 }
727
728 /*
729 * Turn off the deadlock and lock timeout timers, if they are still
730 * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
731 * indicator flag, since this function is executed before
732 * ProcessInterrupts when responding to SIGINT; else we'd lose the
733 * knowledge that the SIGINT came from a lock timeout and not an external
734 * source.
735 */
736 timeouts[0].id = DEADLOCK_TIMEOUT;
737 timeouts[0].keep_indicator = false;
738 timeouts[1].id = LOCK_TIMEOUT;
739 timeouts[1].keep_indicator = true;
740 disable_timeouts(timeouts, 2);
741
742 /* Unlink myself from the wait queue, if on it (might not be anymore!) */
743 partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
744 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
745
746 if (MyProc->links.next != NULL)
747 {
748 /* We could not have been granted the lock yet */
749 RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
750 }
751 else
752 {
753 /*
754 * Somebody kicked us off the lock queue already. Perhaps they
755 * granted us the lock, or perhaps they detected a deadlock. If they
756 * did grant us the lock, we'd better remember it in our local lock
757 * table.
758 */
759 if (MyProc->waitStatus == STATUS_OK)
760 GrantAwaitedLock();
761 }
762
763 lockAwaited = NULL;
764
765 LWLockRelease(partitionLock);
766
767 RESUME_INTERRUPTS();
768}
769
770
771/*
772 * ProcReleaseLocks() -- release locks associated with current transaction
773 * at main transaction commit or abort
774 *
775 * At main transaction commit, we release standard locks except session locks.
776 * At main transaction abort, we release all locks including session locks.
777 *
778 * Advisory locks are released only if they are transaction-level;
779 * session-level holds remain, whether this is a commit or not.
780 *
781 * At subtransaction commit, we don't release any locks (so this func is not
782 * needed at all); we will defer the releasing to the parent transaction.
783 * At subtransaction abort, we release all locks held by the subtransaction;
784 * this is implemented by retail releasing of the locks under control of
785 * the ResourceOwner mechanism.
786 */
787void
788ProcReleaseLocks(bool isCommit)
789{
790 if (!MyProc)
791 return;
792 /* If waiting, get off wait queue (should only be needed after error) */
793 LockErrorCleanup();
794 /* Release standard locks, including session-level if aborting */
795 LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
796 /* Release transaction-level advisory locks */
797 LockReleaseAll(USER_LOCKMETHOD, false);
798}
799
800
801/*
802 * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
803 */
804static void
805RemoveProcFromArray(int code, Datum arg)
806{
807 Assert(MyProc != NULL);
808 ProcArrayRemove(MyProc, InvalidTransactionId);
809}
810
811/*
812 * ProcKill() -- Destroy the per-proc data structure for
813 * this process. Release any of its held LW locks.
814 */
815static void
816ProcKill(int code, Datum arg)
817{
818 PGPROC *proc;
819 PGPROC *volatile *procgloballist;
820
821 Assert(MyProc != NULL);
822
823 /* Make sure we're out of the sync rep lists */
824 SyncRepCleanupAtProcExit();
825
826#ifdef USE_ASSERT_CHECKING
827 {
828 int i;
829
830 /* Last process should have released all locks. */
831 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
832 Assert(SHMQueueEmpty(&(MyProc->myProcLocks[i])));
833 }
834#endif
835
836 /*
837 * Release any LW locks I am holding. There really shouldn't be any, but
838 * it's cheap to check again before we cut the knees off the LWLock
839 * facility by releasing our PGPROC ...
840 */
841 LWLockReleaseAll();
842
843 /* Cancel any pending condition variable sleep, too */
844 ConditionVariableCancelSleep();
845
846 /* Make sure active replication slots are released */
847 if (MyReplicationSlot != NULL)
848 ReplicationSlotRelease();
849
850 /* Also cleanup all the temporary slots. */
851 ReplicationSlotCleanup();
852
853 /*
854 * Detach from any lock group of which we are a member. If the leader
855 * exist before all other group members, it's PGPROC will remain allocated
856 * until the last group process exits; that process must return the
857 * leader's PGPROC to the appropriate list.
858 */
859 if (MyProc->lockGroupLeader != NULL)
860 {
861 PGPROC *leader = MyProc->lockGroupLeader;
862 LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
863
864 LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
865 Assert(!dlist_is_empty(&leader->lockGroupMembers));
866 dlist_delete(&MyProc->lockGroupLink);
867 if (dlist_is_empty(&leader->lockGroupMembers))
868 {
869 leader->lockGroupLeader = NULL;
870 if (leader != MyProc)
871 {
872 procgloballist = leader->procgloballist;
873
874 /* Leader exited first; return its PGPROC. */
875 SpinLockAcquire(ProcStructLock);
876 leader->links.next = (SHM_QUEUE *) *procgloballist;
877 *procgloballist = leader;
878 SpinLockRelease(ProcStructLock);
879 }
880 }
881 else if (leader != MyProc)
882 MyProc->lockGroupLeader = NULL;
883 LWLockRelease(leader_lwlock);
884 }
885
886 /*
887 * Reset MyLatch to the process local one. This is so that signal
888 * handlers et al can continue using the latch after the shared latch
889 * isn't ours anymore. After that clear MyProc and disown the shared
890 * latch.
891 */
892 SwitchBackToLocalLatch();
893 proc = MyProc;
894 MyProc = NULL;
895 DisownLatch(&proc->procLatch);
896
897 procgloballist = proc->procgloballist;
898 SpinLockAcquire(ProcStructLock);
899
900 /*
901 * If we're still a member of a locking group, that means we're a leader
902 * which has somehow exited before its children. The last remaining child
903 * will release our PGPROC. Otherwise, release it now.
904 */
905 if (proc->lockGroupLeader == NULL)
906 {
907 /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
908 Assert(dlist_is_empty(&proc->lockGroupMembers));
909
910 /* Return PGPROC structure (and semaphore) to appropriate freelist */
911 proc->links.next = (SHM_QUEUE *) *procgloballist;
912 *procgloballist = proc;
913 }
914
915 /* Update shared estimate of spins_per_delay */
916 ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
917
918 SpinLockRelease(ProcStructLock);
919
920 /*
921 * This process is no longer present in shared memory in any meaningful
922 * way, so tell the postmaster we've cleaned up acceptably well. (XXX
923 * autovac launcher should be included here someday)
924 */
925 if (IsUnderPostmaster && !IsAutoVacuumLauncherProcess())
926 MarkPostmasterChildInactive();
927
928 /* wake autovac launcher if needed -- see comments in FreeWorkerInfo */
929 if (AutovacuumLauncherPid != 0)
930 kill(AutovacuumLauncherPid, SIGUSR2);
931}
932
933/*
934 * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
935 * processes (bgwriter, etc). The PGPROC and sema are not released, only
936 * marked as not-in-use.
937 */
938static void
939AuxiliaryProcKill(int code, Datum arg)
940{
941 int proctype = DatumGetInt32(arg);
942 PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
943 PGPROC *proc;
944
945 Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
946
947 auxproc = &AuxiliaryProcs[proctype];
948
949 Assert(MyProc == auxproc);
950
951 /* Release any LW locks I am holding (see notes above) */
952 LWLockReleaseAll();
953
954 /* Cancel any pending condition variable sleep, too */
955 ConditionVariableCancelSleep();
956
957 /*
958 * Reset MyLatch to the process local one. This is so that signal
959 * handlers et al can continue using the latch after the shared latch
960 * isn't ours anymore. After that clear MyProc and disown the shared
961 * latch.
962 */
963 SwitchBackToLocalLatch();
964 proc = MyProc;
965 MyProc = NULL;
966 DisownLatch(&proc->procLatch);
967
968 SpinLockAcquire(ProcStructLock);
969
970 /* Mark auxiliary proc no longer in use */
971 proc->pid = 0;
972
973 /* Update shared estimate of spins_per_delay */
974 ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
975
976 SpinLockRelease(ProcStructLock);
977}
978
979/*
980 * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
981 * given its PID
982 *
983 * Returns NULL if not found.
984 */
985PGPROC *
986AuxiliaryPidGetProc(int pid)
987{
988 PGPROC *result = NULL;
989 int index;
990
991 if (pid == 0) /* never match dummy PGPROCs */
992 return NULL;
993
994 for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
995 {
996 PGPROC *proc = &AuxiliaryProcs[index];
997
998 if (proc->pid == pid)
999 {
1000 result = proc;
1001 break;
1002 }
1003 }
1004 return result;
1005}
1006
1007/*
1008 * ProcQueue package: routines for putting processes to sleep
1009 * and waking them up
1010 */
1011
1012/*
1013 * ProcQueueAlloc -- alloc/attach to a shared memory process queue
1014 *
1015 * Returns: a pointer to the queue
1016 * Side Effects: Initializes the queue if it wasn't there before
1017 */
1018#ifdef NOT_USED
1019PROC_QUEUE *
1020ProcQueueAlloc(const char *name)
1021{
1022 PROC_QUEUE *queue;
1023 bool found;
1024
1025 queue = (PROC_QUEUE *)
1026 ShmemInitStruct(name, sizeof(PROC_QUEUE), &found);
1027
1028 if (!found)
1029 ProcQueueInit(queue);
1030
1031 return queue;
1032}
1033#endif
1034
1035/*
1036 * ProcQueueInit -- initialize a shared memory process queue
1037 */
1038void
1039ProcQueueInit(PROC_QUEUE *queue)
1040{
1041 SHMQueueInit(&(queue->links));
1042 queue->size = 0;
1043}
1044
1045
1046/*
1047 * ProcSleep -- put a process to sleep on the specified lock
1048 *
1049 * Caller must have set MyProc->heldLocks to reflect locks already held
1050 * on the lockable object by this process (under all XIDs).
1051 *
1052 * The lock table's partition lock must be held at entry, and will be held
1053 * at exit.
1054 *
1055 * Result: STATUS_OK if we acquired the lock, STATUS_ERROR if not (deadlock).
1056 *
1057 * ASSUME: that no one will fiddle with the queue until after
1058 * we release the partition lock.
1059 *
1060 * NOTES: The process queue is now a priority queue for locking.
1061 */
1062int
1063ProcSleep(LOCALLOCK *locallock, LockMethod lockMethodTable)
1064{
1065 LOCKMODE lockmode = locallock->tag.mode;
1066 LOCK *lock = locallock->lock;
1067 PROCLOCK *proclock = locallock->proclock;
1068 uint32 hashcode = locallock->hashcode;
1069 LWLock *partitionLock = LockHashPartitionLock(hashcode);
1070 PROC_QUEUE *waitQueue = &(lock->waitProcs);
1071 LOCKMASK myHeldLocks = MyProc->heldLocks;
1072 bool early_deadlock = false;
1073 bool allow_autovacuum_cancel = true;
1074 int myWaitStatus;
1075 PGPROC *proc;
1076 PGPROC *leader = MyProc->lockGroupLeader;
1077 int i;
1078
1079 /*
1080 * If group locking is in use, locks held by members of my locking group
1081 * need to be included in myHeldLocks.
1082 */
1083 if (leader != NULL)
1084 {
1085 SHM_QUEUE *procLocks = &(lock->procLocks);
1086 PROCLOCK *otherproclock;
1087
1088 otherproclock = (PROCLOCK *)
1089 SHMQueueNext(procLocks, procLocks, offsetof(PROCLOCK, lockLink));
1090 while (otherproclock != NULL)
1091 {
1092 if (otherproclock->groupLeader == leader)
1093 myHeldLocks |= otherproclock->holdMask;
1094 otherproclock = (PROCLOCK *)
1095 SHMQueueNext(procLocks, &otherproclock->lockLink,
1096 offsetof(PROCLOCK, lockLink));
1097 }
1098 }
1099
1100 /*
1101 * Determine where to add myself in the wait queue.
1102 *
1103 * Normally I should go at the end of the queue. However, if I already
1104 * hold locks that conflict with the request of any previous waiter, put
1105 * myself in the queue just in front of the first such waiter. This is not
1106 * a necessary step, since deadlock detection would move me to before that
1107 * waiter anyway; but it's relatively cheap to detect such a conflict
1108 * immediately, and avoid delaying till deadlock timeout.
1109 *
1110 * Special case: if I find I should go in front of some waiter, check to
1111 * see if I conflict with already-held locks or the requests before that
1112 * waiter. If not, then just grant myself the requested lock immediately.
1113 * This is the same as the test for immediate grant in LockAcquire, except
1114 * we are only considering the part of the wait queue before my insertion
1115 * point.
1116 */
1117 if (myHeldLocks != 0)
1118 {
1119 LOCKMASK aheadRequests = 0;
1120
1121 proc = (PGPROC *) waitQueue->links.next;
1122 for (i = 0; i < waitQueue->size; i++)
1123 {
1124 /*
1125 * If we're part of the same locking group as this waiter, its
1126 * locks neither conflict with ours nor contribute to
1127 * aheadRequests.
1128 */
1129 if (leader != NULL && leader == proc->lockGroupLeader)
1130 {
1131 proc = (PGPROC *) proc->links.next;
1132 continue;
1133 }
1134 /* Must he wait for me? */
1135 if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1136 {
1137 /* Must I wait for him ? */
1138 if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1139 {
1140 /*
1141 * Yes, so we have a deadlock. Easiest way to clean up
1142 * correctly is to call RemoveFromWaitQueue(), but we
1143 * can't do that until we are *on* the wait queue. So, set
1144 * a flag to check below, and break out of loop. Also,
1145 * record deadlock info for later message.
1146 */
1147 RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1148 early_deadlock = true;
1149 break;
1150 }
1151 /* I must go before this waiter. Check special case. */
1152 if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1153 LockCheckConflicts(lockMethodTable,
1154 lockmode,
1155 lock,
1156 proclock) == STATUS_OK)
1157 {
1158 /* Skip the wait and just grant myself the lock. */
1159 GrantLock(lock, proclock, lockmode);
1160 GrantAwaitedLock();
1161 return STATUS_OK;
1162 }
1163 /* Break out of loop to put myself before him */
1164 break;
1165 }
1166 /* Nope, so advance to next waiter */
1167 aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1168 proc = (PGPROC *) proc->links.next;
1169 }
1170
1171 /*
1172 * If we fall out of loop normally, proc points to waitQueue head, so
1173 * we will insert at tail of queue as desired.
1174 */
1175 }
1176 else
1177 {
1178 /* I hold no locks, so I can't push in front of anyone. */
1179 proc = (PGPROC *) &(waitQueue->links);
1180 }
1181
1182 /*
1183 * Insert self into queue, ahead of the given proc (or at tail of queue).
1184 */
1185 SHMQueueInsertBefore(&(proc->links), &(MyProc->links));
1186 waitQueue->size++;
1187
1188 lock->waitMask |= LOCKBIT_ON(lockmode);
1189
1190 /* Set up wait information in PGPROC object, too */
1191 MyProc->waitLock = lock;
1192 MyProc->waitProcLock = proclock;
1193 MyProc->waitLockMode = lockmode;
1194
1195 MyProc->waitStatus = STATUS_WAITING;
1196
1197 /*
1198 * If we detected deadlock, give up without waiting. This must agree with
1199 * CheckDeadLock's recovery code.
1200 */
1201 if (early_deadlock)
1202 {
1203 RemoveFromWaitQueue(MyProc, hashcode);
1204 return STATUS_ERROR;
1205 }
1206
1207 /* mark that we are waiting for a lock */
1208 lockAwaited = locallock;
1209
1210 /*
1211 * Release the lock table's partition lock.
1212 *
1213 * NOTE: this may also cause us to exit critical-section state, possibly
1214 * allowing a cancel/die interrupt to be accepted. This is OK because we
1215 * have recorded the fact that we are waiting for a lock, and so
1216 * LockErrorCleanup will clean up if cancel/die happens.
1217 */
1218 LWLockRelease(partitionLock);
1219
1220 /*
1221 * Also, now that we will successfully clean up after an ereport, it's
1222 * safe to check to see if there's a buffer pin deadlock against the
1223 * Startup process. Of course, that's only necessary if we're doing Hot
1224 * Standby and are not the Startup process ourselves.
1225 */
1226 if (RecoveryInProgress() && !InRecovery)
1227 CheckRecoveryConflictDeadlock();
1228
1229 /* Reset deadlock_state before enabling the timeout handler */
1230 deadlock_state = DS_NOT_YET_CHECKED;
1231 got_deadlock_timeout = false;
1232
1233 /*
1234 * Set timer so we can wake up after awhile and check for a deadlock. If a
1235 * deadlock is detected, the handler sets MyProc->waitStatus =
1236 * STATUS_ERROR, allowing us to know that we must report failure rather
1237 * than success.
1238 *
1239 * By delaying the check until we've waited for a bit, we can avoid
1240 * running the rather expensive deadlock-check code in most cases.
1241 *
1242 * If LockTimeout is set, also enable the timeout for that. We can save a
1243 * few cycles by enabling both timeout sources in one call.
1244 *
1245 * If InHotStandby we set lock waits slightly later for clarity with other
1246 * code.
1247 */
1248 if (!InHotStandby)
1249 {
1250 if (LockTimeout > 0)
1251 {
1252 EnableTimeoutParams timeouts[2];
1253
1254 timeouts[0].id = DEADLOCK_TIMEOUT;
1255 timeouts[0].type = TMPARAM_AFTER;
1256 timeouts[0].delay_ms = DeadlockTimeout;
1257 timeouts[1].id = LOCK_TIMEOUT;
1258 timeouts[1].type = TMPARAM_AFTER;
1259 timeouts[1].delay_ms = LockTimeout;
1260 enable_timeouts(timeouts, 2);
1261 }
1262 else
1263 enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1264 }
1265
1266 /*
1267 * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1268 * will not wait. But a set latch does not necessarily mean that the lock
1269 * is free now, as there are many other sources for latch sets than
1270 * somebody releasing the lock.
1271 *
1272 * We process interrupts whenever the latch has been set, so cancel/die
1273 * interrupts are processed quickly. This means we must not mind losing
1274 * control to a cancel/die interrupt here. We don't, because we have no
1275 * shared-state-change work to do after being granted the lock (the
1276 * grantor did it all). We do have to worry about canceling the deadlock
1277 * timeout and updating the locallock table, but if we lose control to an
1278 * error, LockErrorCleanup will fix that up.
1279 */
1280 do
1281 {
1282 if (InHotStandby)
1283 {
1284 /* Set a timer and wait for that or for the Lock to be granted */
1285 ResolveRecoveryConflictWithLock(locallock->tag.lock);
1286 }
1287 else
1288 {
1289 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1290 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1291 ResetLatch(MyLatch);
1292 /* check for deadlocks first, as that's probably log-worthy */
1293 if (got_deadlock_timeout)
1294 {
1295 CheckDeadLock();
1296 got_deadlock_timeout = false;
1297 }
1298 CHECK_FOR_INTERRUPTS();
1299 }
1300
1301 /*
1302 * waitStatus could change from STATUS_WAITING to something else
1303 * asynchronously. Read it just once per loop to prevent surprising
1304 * behavior (such as missing log messages).
1305 */
1306 myWaitStatus = *((volatile int *) &MyProc->waitStatus);
1307
1308 /*
1309 * If we are not deadlocked, but are waiting on an autovacuum-induced
1310 * task, send a signal to interrupt it.
1311 */
1312 if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1313 {
1314 PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1315 PGXACT *autovac_pgxact = &ProcGlobal->allPgXact[autovac->pgprocno];
1316
1317 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1318
1319 /*
1320 * Only do it if the worker is not working to protect against Xid
1321 * wraparound.
1322 */
1323 if ((autovac_pgxact->vacuumFlags & PROC_IS_AUTOVACUUM) &&
1324 !(autovac_pgxact->vacuumFlags & PROC_VACUUM_FOR_WRAPAROUND))
1325 {
1326 int pid = autovac->pid;
1327 StringInfoData locktagbuf;
1328 StringInfoData logbuf; /* errdetail for server log */
1329
1330 initStringInfo(&locktagbuf);
1331 initStringInfo(&logbuf);
1332 DescribeLockTag(&locktagbuf, &lock->tag);
1333 appendStringInfo(&logbuf,
1334 _("Process %d waits for %s on %s."),
1335 MyProcPid,
1336 GetLockmodeName(lock->tag.locktag_lockmethodid,
1337 lockmode),
1338 locktagbuf.data);
1339
1340 /* release lock as quickly as possible */
1341 LWLockRelease(ProcArrayLock);
1342
1343 /* send the autovacuum worker Back to Old Kent Road */
1344 ereport(DEBUG1,
1345 (errmsg("sending cancel to blocking autovacuum PID %d",
1346 pid),
1347 errdetail_log("%s", logbuf.data)));
1348
1349 if (kill(pid, SIGINT) < 0)
1350 {
1351 /*
1352 * There's a race condition here: once we release the
1353 * ProcArrayLock, it's possible for the autovac worker to
1354 * close up shop and exit before we can do the kill().
1355 * Therefore, we do not whinge about no-such-process.
1356 * Other errors such as EPERM could conceivably happen if
1357 * the kernel recycles the PID fast enough, but such cases
1358 * seem improbable enough that it's probably best to issue
1359 * a warning if we see some other errno.
1360 */
1361 if (errno != ESRCH)
1362 ereport(WARNING,
1363 (errmsg("could not send signal to process %d: %m",
1364 pid)));
1365 }
1366
1367 pfree(logbuf.data);
1368 pfree(locktagbuf.data);
1369 }
1370 else
1371 LWLockRelease(ProcArrayLock);
1372
1373 /* prevent signal from being resent more than once */
1374 allow_autovacuum_cancel = false;
1375 }
1376
1377 /*
1378 * If awoken after the deadlock check interrupt has run, and
1379 * log_lock_waits is on, then report about the wait.
1380 */
1381 if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1382 {
1383 StringInfoData buf,
1384 lock_waiters_sbuf,
1385 lock_holders_sbuf;
1386 const char *modename;
1387 long secs;
1388 int usecs;
1389 long msecs;
1390 SHM_QUEUE *procLocks;
1391 PROCLOCK *proclock;
1392 bool first_holder = true,
1393 first_waiter = true;
1394 int lockHoldersNum = 0;
1395
1396 initStringInfo(&buf);
1397 initStringInfo(&lock_waiters_sbuf);
1398 initStringInfo(&lock_holders_sbuf);
1399
1400 DescribeLockTag(&buf, &locallock->tag.lock);
1401 modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1402 lockmode);
1403 TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1404 GetCurrentTimestamp(),
1405 &secs, &usecs);
1406 msecs = secs * 1000 + usecs / 1000;
1407 usecs = usecs % 1000;
1408
1409 /*
1410 * we loop over the lock's procLocks to gather a list of all
1411 * holders and waiters. Thus we will be able to provide more
1412 * detailed information for lock debugging purposes.
1413 *
1414 * lock->procLocks contains all processes which hold or wait for
1415 * this lock.
1416 */
1417
1418 LWLockAcquire(partitionLock, LW_SHARED);
1419
1420 procLocks = &(lock->procLocks);
1421 proclock = (PROCLOCK *) SHMQueueNext(procLocks, procLocks,
1422 offsetof(PROCLOCK, lockLink));
1423
1424 while (proclock)
1425 {
1426 /*
1427 * we are a waiter if myProc->waitProcLock == proclock; we are
1428 * a holder if it is NULL or something different
1429 */
1430 if (proclock->tag.myProc->waitProcLock == proclock)
1431 {
1432 if (first_waiter)
1433 {
1434 appendStringInfo(&lock_waiters_sbuf, "%d",
1435 proclock->tag.myProc->pid);
1436 first_waiter = false;
1437 }
1438 else
1439 appendStringInfo(&lock_waiters_sbuf, ", %d",
1440 proclock->tag.myProc->pid);
1441 }
1442 else
1443 {
1444 if (first_holder)
1445 {
1446 appendStringInfo(&lock_holders_sbuf, "%d",
1447 proclock->tag.myProc->pid);
1448 first_holder = false;
1449 }
1450 else
1451 appendStringInfo(&lock_holders_sbuf, ", %d",
1452 proclock->tag.myProc->pid);
1453
1454 lockHoldersNum++;
1455 }
1456
1457 proclock = (PROCLOCK *) SHMQueueNext(procLocks, &proclock->lockLink,
1458 offsetof(PROCLOCK, lockLink));
1459 }
1460
1461 LWLockRelease(partitionLock);
1462
1463 if (deadlock_state == DS_SOFT_DEADLOCK)
1464 ereport(LOG,
1465 (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1466 MyProcPid, modename, buf.data, msecs, usecs),
1467 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1468 "Processes holding the lock: %s. Wait queue: %s.",
1469 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1470 else if (deadlock_state == DS_HARD_DEADLOCK)
1471 {
1472 /*
1473 * This message is a bit redundant with the error that will be
1474 * reported subsequently, but in some cases the error report
1475 * might not make it to the log (eg, if it's caught by an
1476 * exception handler), and we want to ensure all long-wait
1477 * events get logged.
1478 */
1479 ereport(LOG,
1480 (errmsg("process %d detected deadlock while waiting for %s on %s after %ld.%03d ms",
1481 MyProcPid, modename, buf.data, msecs, usecs),
1482 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1483 "Processes holding the lock: %s. Wait queue: %s.",
1484 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1485 }
1486
1487 if (myWaitStatus == STATUS_WAITING)
1488 ereport(LOG,
1489 (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1490 MyProcPid, modename, buf.data, msecs, usecs),
1491 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1492 "Processes holding the lock: %s. Wait queue: %s.",
1493 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1494 else if (myWaitStatus == STATUS_OK)
1495 ereport(LOG,
1496 (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1497 MyProcPid, modename, buf.data, msecs, usecs)));
1498 else
1499 {
1500 Assert(myWaitStatus == STATUS_ERROR);
1501
1502 /*
1503 * Currently, the deadlock checker always kicks its own
1504 * process, which means that we'll only see STATUS_ERROR when
1505 * deadlock_state == DS_HARD_DEADLOCK, and there's no need to
1506 * print redundant messages. But for completeness and
1507 * future-proofing, print a message if it looks like someone
1508 * else kicked us off the lock.
1509 */
1510 if (deadlock_state != DS_HARD_DEADLOCK)
1511 ereport(LOG,
1512 (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1513 MyProcPid, modename, buf.data, msecs, usecs),
1514 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1515 "Processes holding the lock: %s. Wait queue: %s.",
1516 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1517 }
1518
1519 /*
1520 * At this point we might still need to wait for the lock. Reset
1521 * state so we don't print the above messages again.
1522 */
1523 deadlock_state = DS_NO_DEADLOCK;
1524
1525 pfree(buf.data);
1526 pfree(lock_holders_sbuf.data);
1527 pfree(lock_waiters_sbuf.data);
1528 }
1529 } while (myWaitStatus == STATUS_WAITING);
1530
1531 /*
1532 * Disable the timers, if they are still running. As in LockErrorCleanup,
1533 * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1534 * already caused QueryCancelPending to become set, we want the cancel to
1535 * be reported as a lock timeout, not a user cancel.
1536 */
1537 if (!InHotStandby)
1538 {
1539 if (LockTimeout > 0)
1540 {
1541 DisableTimeoutParams timeouts[2];
1542
1543 timeouts[0].id = DEADLOCK_TIMEOUT;
1544 timeouts[0].keep_indicator = false;
1545 timeouts[1].id = LOCK_TIMEOUT;
1546 timeouts[1].keep_indicator = true;
1547 disable_timeouts(timeouts, 2);
1548 }
1549 else
1550 disable_timeout(DEADLOCK_TIMEOUT, false);
1551 }
1552
1553 /*
1554 * Re-acquire the lock table's partition lock. We have to do this to hold
1555 * off cancel/die interrupts before we can mess with lockAwaited (else we
1556 * might have a missed or duplicated locallock update).
1557 */
1558 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
1559
1560 /*
1561 * We no longer want LockErrorCleanup to do anything.
1562 */
1563 lockAwaited = NULL;
1564
1565 /*
1566 * If we got the lock, be sure to remember it in the locallock table.
1567 */
1568 if (MyProc->waitStatus == STATUS_OK)
1569 GrantAwaitedLock();
1570
1571 /*
1572 * We don't have to do anything else, because the awaker did all the
1573 * necessary update of the lock table and MyProc.
1574 */
1575 return MyProc->waitStatus;
1576}
1577
1578
1579/*
1580 * ProcWakeup -- wake up a process by setting its latch.
1581 *
1582 * Also remove the process from the wait queue and set its links invalid.
1583 * RETURN: the next process in the wait queue.
1584 *
1585 * The appropriate lock partition lock must be held by caller.
1586 *
1587 * XXX: presently, this code is only used for the "success" case, and only
1588 * works correctly for that case. To clean up in failure case, would need
1589 * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1590 * Hence, in practice the waitStatus parameter must be STATUS_OK.
1591 */
1592PGPROC *
1593ProcWakeup(PGPROC *proc, int waitStatus)
1594{
1595 PGPROC *retProc;
1596
1597 /* Proc should be sleeping ... */
1598 if (proc->links.prev == NULL ||
1599 proc->links.next == NULL)
1600 return NULL;
1601 Assert(proc->waitStatus == STATUS_WAITING);
1602
1603 /* Save next process before we zap the list link */
1604 retProc = (PGPROC *) proc->links.next;
1605
1606 /* Remove process from wait queue */
1607 SHMQueueDelete(&(proc->links));
1608 (proc->waitLock->waitProcs.size)--;
1609
1610 /* Clean up process' state and pass it the ok/fail signal */
1611 proc->waitLock = NULL;
1612 proc->waitProcLock = NULL;
1613 proc->waitStatus = waitStatus;
1614
1615 /* And awaken it */
1616 SetLatch(&proc->procLatch);
1617
1618 return retProc;
1619}
1620
1621/*
1622 * ProcLockWakeup -- routine for waking up processes when a lock is
1623 * released (or a prior waiter is aborted). Scan all waiters
1624 * for lock, waken any that are no longer blocked.
1625 *
1626 * The appropriate lock partition lock must be held by caller.
1627 */
1628void
1629ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1630{
1631 PROC_QUEUE *waitQueue = &(lock->waitProcs);
1632 int queue_size = waitQueue->size;
1633 PGPROC *proc;
1634 LOCKMASK aheadRequests = 0;
1635
1636 Assert(queue_size >= 0);
1637
1638 if (queue_size == 0)
1639 return;
1640
1641 proc = (PGPROC *) waitQueue->links.next;
1642
1643 while (queue_size-- > 0)
1644 {
1645 LOCKMODE lockmode = proc->waitLockMode;
1646
1647 /*
1648 * Waken if (a) doesn't conflict with requests of earlier waiters, and
1649 * (b) doesn't conflict with already-held locks.
1650 */
1651 if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1652 LockCheckConflicts(lockMethodTable,
1653 lockmode,
1654 lock,
1655 proc->waitProcLock) == STATUS_OK)
1656 {
1657 /* OK to waken */
1658 GrantLock(lock, proc->waitProcLock, lockmode);
1659 proc = ProcWakeup(proc, STATUS_OK);
1660
1661 /*
1662 * ProcWakeup removes proc from the lock's waiting process queue
1663 * and returns the next proc in chain; don't use proc's next-link,
1664 * because it's been cleared.
1665 */
1666 }
1667 else
1668 {
1669 /*
1670 * Cannot wake this guy. Remember his request for later checks.
1671 */
1672 aheadRequests |= LOCKBIT_ON(lockmode);
1673 proc = (PGPROC *) proc->links.next;
1674 }
1675 }
1676
1677 Assert(waitQueue->size >= 0);
1678}
1679
1680/*
1681 * CheckDeadLock
1682 *
1683 * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1684 * lock to be released by some other process. Check if there's a deadlock; if
1685 * not, just return. (But signal ProcSleep to log a message, if
1686 * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1687 * the lock's wait queue and signal an error to ProcSleep.
1688 */
1689static void
1690CheckDeadLock(void)
1691{
1692 int i;
1693
1694 /*
1695 * Acquire exclusive lock on the entire shared lock data structures. Must
1696 * grab LWLocks in partition-number order to avoid LWLock deadlock.
1697 *
1698 * Note that the deadlock check interrupt had better not be enabled
1699 * anywhere that this process itself holds lock partition locks, else this
1700 * will wait forever. Also note that LWLockAcquire creates a critical
1701 * section, so that this routine cannot be interrupted by cancel/die
1702 * interrupts.
1703 */
1704 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1705 LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1706
1707 /*
1708 * Check to see if we've been awoken by anyone in the interim.
1709 *
1710 * If we have, we can return and resume our transaction -- happy day.
1711 * Before we are awoken the process releasing the lock grants it to us so
1712 * we know that we don't have to wait anymore.
1713 *
1714 * We check by looking to see if we've been unlinked from the wait queue.
1715 * This is safe because we hold the lock partition lock.
1716 */
1717 if (MyProc->links.prev == NULL ||
1718 MyProc->links.next == NULL)
1719 goto check_done;
1720
1721#ifdef LOCK_DEBUG
1722 if (Debug_deadlocks)
1723 DumpAllLocks();
1724#endif
1725
1726 /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1727 deadlock_state = DeadLockCheck(MyProc);
1728
1729 if (deadlock_state == DS_HARD_DEADLOCK)
1730 {
1731 /*
1732 * Oops. We have a deadlock.
1733 *
1734 * Get this process out of wait state. (Note: we could do this more
1735 * efficiently by relying on lockAwaited, but use this coding to
1736 * preserve the flexibility to kill some other transaction than the
1737 * one detecting the deadlock.)
1738 *
1739 * RemoveFromWaitQueue sets MyProc->waitStatus to STATUS_ERROR, so
1740 * ProcSleep will report an error after we return from the signal
1741 * handler.
1742 */
1743 Assert(MyProc->waitLock != NULL);
1744 RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1745
1746 /*
1747 * We're done here. Transaction abort caused by the error that
1748 * ProcSleep will raise will cause any other locks we hold to be
1749 * released, thus allowing other processes to wake up; we don't need
1750 * to do that here. NOTE: an exception is that releasing locks we
1751 * hold doesn't consider the possibility of waiters that were blocked
1752 * behind us on the lock we just failed to get, and might now be
1753 * wakable because we're not in front of them anymore. However,
1754 * RemoveFromWaitQueue took care of waking up any such processes.
1755 */
1756 }
1757
1758 /*
1759 * And release locks. We do this in reverse order for two reasons: (1)
1760 * Anyone else who needs more than one of the locks will be trying to lock
1761 * them in increasing order; we don't want to release the other process
1762 * until it can get all the locks it needs. (2) This avoids O(N^2)
1763 * behavior inside LWLockRelease.
1764 */
1765check_done:
1766 for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1767 LWLockRelease(LockHashPartitionLockByIndex(i));
1768}
1769
1770/*
1771 * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1772 *
1773 * NB: Runs inside a signal handler, be careful.
1774 */
1775void
1776CheckDeadLockAlert(void)
1777{
1778 int save_errno = errno;
1779
1780 got_deadlock_timeout = true;
1781
1782 /*
1783 * Have to set the latch again, even if handle_sig_alarm already did. Back
1784 * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1785 * ever would be a problem, but setting a set latch again is cheap.
1786 */
1787 SetLatch(MyLatch);
1788 errno = save_errno;
1789}
1790
1791/*
1792 * ProcWaitForSignal - wait for a signal from another backend.
1793 *
1794 * As this uses the generic process latch the caller has to be robust against
1795 * unrelated wakeups: Always check that the desired state has occurred, and
1796 * wait again if not.
1797 */
1798void
1799ProcWaitForSignal(uint32 wait_event_info)
1800{
1801 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1802 wait_event_info);
1803 ResetLatch(MyLatch);
1804 CHECK_FOR_INTERRUPTS();
1805}
1806
1807/*
1808 * ProcSendSignal - send a signal to a backend identified by PID
1809 */
1810void
1811ProcSendSignal(int pid)
1812{
1813 PGPROC *proc = NULL;
1814
1815 if (RecoveryInProgress())
1816 {
1817 SpinLockAcquire(ProcStructLock);
1818
1819 /*
1820 * Check to see whether it is the Startup process we wish to signal.
1821 * This call is made by the buffer manager when it wishes to wake up a
1822 * process that has been waiting for a pin in so it can obtain a
1823 * cleanup lock using LockBufferForCleanup(). Startup is not a normal
1824 * backend, so BackendPidGetProc() will not return any pid at all. So
1825 * we remember the information for this special case.
1826 */
1827 if (pid == ProcGlobal->startupProcPid)
1828 proc = ProcGlobal->startupProc;
1829
1830 SpinLockRelease(ProcStructLock);
1831 }
1832
1833 if (proc == NULL)
1834 proc = BackendPidGetProc(pid);
1835
1836 if (proc != NULL)
1837 {
1838 SetLatch(&proc->procLatch);
1839 }
1840}
1841
1842/*
1843 * BecomeLockGroupLeader - designate process as lock group leader
1844 *
1845 * Once this function has returned, other processes can join the lock group
1846 * by calling BecomeLockGroupMember.
1847 */
1848void
1849BecomeLockGroupLeader(void)
1850{
1851 LWLock *leader_lwlock;
1852
1853 /* If we already did it, we don't need to do it again. */
1854 if (MyProc->lockGroupLeader == MyProc)
1855 return;
1856
1857 /* We had better not be a follower. */
1858 Assert(MyProc->lockGroupLeader == NULL);
1859
1860 /* Create single-member group, containing only ourselves. */
1861 leader_lwlock = LockHashPartitionLockByProc(MyProc);
1862 LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1863 MyProc->lockGroupLeader = MyProc;
1864 dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
1865 LWLockRelease(leader_lwlock);
1866}
1867
1868/*
1869 * BecomeLockGroupMember - designate process as lock group member
1870 *
1871 * This is pretty straightforward except for the possibility that the leader
1872 * whose group we're trying to join might exit before we manage to do so;
1873 * and the PGPROC might get recycled for an unrelated process. To avoid
1874 * that, we require the caller to pass the PID of the intended PGPROC as
1875 * an interlock. Returns true if we successfully join the intended lock
1876 * group, and false if not.
1877 */
1878bool
1879BecomeLockGroupMember(PGPROC *leader, int pid)
1880{
1881 LWLock *leader_lwlock;
1882 bool ok = false;
1883
1884 /* Group leader can't become member of group */
1885 Assert(MyProc != leader);
1886
1887 /* Can't already be a member of a group */
1888 Assert(MyProc->lockGroupLeader == NULL);
1889
1890 /* PID must be valid. */
1891 Assert(pid != 0);
1892
1893 /*
1894 * Get lock protecting the group fields. Note LockHashPartitionLockByProc
1895 * accesses leader->pgprocno in a PGPROC that might be free. This is safe
1896 * because all PGPROCs' pgprocno fields are set during shared memory
1897 * initialization and never change thereafter; so we will acquire the
1898 * correct lock even if the leader PGPROC is in process of being recycled.
1899 */
1900 leader_lwlock = LockHashPartitionLockByProc(leader);
1901 LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1902
1903 /* Is this the leader we're looking for? */
1904 if (leader->pid == pid && leader->lockGroupLeader == leader)
1905 {
1906 /* OK, join the group */
1907 ok = true;
1908 MyProc->lockGroupLeader = leader;
1909 dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
1910 }
1911 LWLockRelease(leader_lwlock);
1912
1913 return ok;
1914}
1915