1 /*-------------------------------------------------------------------------
2 *
3 * proc.c
4 * routines to manage per-process shared memory data structure
5 *
6 * Portions Copyright (c) 1996-2020, 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/ipc.h"
49 #include "storage/lmgr.h"
50 #include "storage/pmsignal.h"
51 #include "storage/proc.h"
52 #include "storage/procarray.h"
53 #include "storage/procsignal.h"
54 #include "storage/spin.h"
55 #include "storage/standby.h"
56 #include "utils/timeout.h"
57 #include "utils/timestamp.h"
58
59 /* GUC variables */
60 int DeadlockTimeout = 1000;
61 int StatementTimeout = 0;
62 int LockTimeout = 0;
63 int IdleInTransactionSessionTimeout = 0;
64 bool log_lock_waits = false;
65
66 /* Pointer to this process's PGPROC and PGXACT structs, if any */
67 PGPROC *MyProc = NULL;
68 PGXACT *MyPgXact = NULL;
69
70 /*
71 * This spinlock protects the freelist of recycled PGPROC structures.
72 * We cannot use an LWLock because the LWLock manager depends on already
73 * having a PGPROC and a wait semaphore! But these structures are touched
74 * relatively infrequently (only at backend startup or shutdown) and not for
75 * very long, so a spinlock is okay.
76 */
77 NON_EXEC_STATIC slock_t *ProcStructLock = NULL;
78
79 /* Pointers to shared-memory structures */
80 PROC_HDR *ProcGlobal = NULL;
81 NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
82 PGPROC *PreparedXactProcs = NULL;
83
84 /* If we are waiting for a lock, this points to the associated LOCALLOCK */
85 static LOCALLOCK *lockAwaited = NULL;
86
87 static DeadLockState deadlock_state = DS_NOT_YET_CHECKED;
88
89 /* Is a deadlock check pending? */
90 static volatile sig_atomic_t got_deadlock_timeout;
91
92 static void RemoveProcFromArray(int code, Datum arg);
93 static void ProcKill(int code, Datum arg);
94 static void AuxiliaryProcKill(int code, Datum arg);
95 static void CheckDeadLock(void);
96
97
98 /*
99 * Report shared-memory space needed by InitProcGlobal.
100 */
101 Size
ProcGlobalShmemSize(void)102 ProcGlobalShmemSize(void)
103 {
104 Size size = 0;
105
106 /* ProcGlobal */
107 size = add_size(size, sizeof(PROC_HDR));
108 /* MyProcs, including autovacuum workers and launcher */
109 size = add_size(size, mul_size(MaxBackends, sizeof(PGPROC)));
110 /* AuxiliaryProcs */
111 size = add_size(size, mul_size(NUM_AUXILIARY_PROCS, sizeof(PGPROC)));
112 /* Prepared xacts */
113 size = add_size(size, mul_size(max_prepared_xacts, sizeof(PGPROC)));
114 /* ProcStructLock */
115 size = add_size(size, sizeof(slock_t));
116
117 size = add_size(size, mul_size(MaxBackends, sizeof(PGXACT)));
118 size = add_size(size, mul_size(NUM_AUXILIARY_PROCS, sizeof(PGXACT)));
119 size = add_size(size, mul_size(max_prepared_xacts, sizeof(PGXACT)));
120
121 return size;
122 }
123
124 /*
125 * Report number of semaphores needed by InitProcGlobal.
126 */
127 int
ProcGlobalSemas(void)128 ProcGlobalSemas(void)
129 {
130 /*
131 * We need a sema per backend (including autovacuum), plus one for each
132 * auxiliary process.
133 */
134 return MaxBackends + NUM_AUXILIARY_PROCS;
135 }
136
137 /*
138 * InitProcGlobal -
139 * Initialize the global process table during postmaster or standalone
140 * backend startup.
141 *
142 * We also create all the per-process semaphores we will need to support
143 * the requested number of backends. We used to allocate semaphores
144 * only when backends were actually started up, but that is bad because
145 * it lets Postgres fail under load --- a lot of Unix systems are
146 * (mis)configured with small limits on the number of semaphores, and
147 * running out when trying to start another backend is a common failure.
148 * So, now we grab enough semaphores to support the desired max number
149 * of backends immediately at initialization --- if the sysadmin has set
150 * MaxConnections, max_worker_processes, max_wal_senders, or
151 * autovacuum_max_workers higher than his kernel will support, he'll
152 * find out sooner rather than later.
153 *
154 * Another reason for creating semaphores here is that the semaphore
155 * implementation typically requires us to create semaphores in the
156 * postmaster, not in backends.
157 *
158 * Note: this is NOT called by individual backends under a postmaster,
159 * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
160 * pointers must be propagated specially for EXEC_BACKEND operation.
161 */
162 void
InitProcGlobal(void)163 InitProcGlobal(void)
164 {
165 PGPROC *procs;
166 PGXACT *pgxacts;
167 int i,
168 j;
169 bool found;
170 uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
171
172 /* Create the ProcGlobal shared structure */
173 ProcGlobal = (PROC_HDR *)
174 ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
175 Assert(!found);
176
177 /*
178 * Initialize the data structures.
179 */
180 ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
181 ProcGlobal->freeProcs = NULL;
182 ProcGlobal->autovacFreeProcs = NULL;
183 ProcGlobal->bgworkerFreeProcs = NULL;
184 ProcGlobal->walsenderFreeProcs = NULL;
185 ProcGlobal->startupProc = NULL;
186 ProcGlobal->startupProcPid = 0;
187 ProcGlobal->startupBufferPinWaitBufId = -1;
188 ProcGlobal->walwriterLatch = NULL;
189 ProcGlobal->checkpointerLatch = NULL;
190 pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PGPROCNO);
191 pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PGPROCNO);
192
193 /*
194 * Create and initialize all the PGPROC structures we'll need. There are
195 * five separate consumers: (1) normal backends, (2) autovacuum workers
196 * and the autovacuum launcher, (3) background workers, (4) auxiliary
197 * processes, and (5) prepared transactions. Each PGPROC structure is
198 * dedicated to exactly one of these purposes, and they do not move
199 * between groups.
200 */
201 procs = (PGPROC *) ShmemAlloc(TotalProcs * sizeof(PGPROC));
202 MemSet(procs, 0, TotalProcs * sizeof(PGPROC));
203 ProcGlobal->allProcs = procs;
204 /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
205 ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
206
207 /*
208 * Also allocate a separate array of PGXACT structures. This is separate
209 * from the main PGPROC array so that the most heavily accessed data is
210 * stored contiguously in memory in as few cache lines as possible. This
211 * provides significant performance benefits, especially on a
212 * multiprocessor system. There is one PGXACT structure for every PGPROC
213 * structure.
214 */
215 pgxacts = (PGXACT *) ShmemAlloc(TotalProcs * sizeof(PGXACT));
216 MemSet(pgxacts, 0, TotalProcs * sizeof(PGXACT));
217 ProcGlobal->allPgXact = pgxacts;
218
219 for (i = 0; i < TotalProcs; i++)
220 {
221 /* Common initialization for all PGPROCs, regardless of type. */
222
223 /*
224 * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
225 * dummy PGPROCs don't need these though - they're never associated
226 * with a real process
227 */
228 if (i < MaxBackends + NUM_AUXILIARY_PROCS)
229 {
230 procs[i].sem = PGSemaphoreCreate();
231 InitSharedLatch(&(procs[i].procLatch));
232 LWLockInitialize(&(procs[i].fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
233 }
234 procs[i].pgprocno = i;
235
236 /*
237 * Newly created PGPROCs for normal backends, autovacuum and bgworkers
238 * must be queued up on the appropriate free list. Because there can
239 * only ever be a small, fixed number of auxiliary processes, no free
240 * list is used in that case; InitAuxiliaryProcess() instead uses a
241 * linear search. PGPROCs for prepared transactions are added to a
242 * free list by TwoPhaseShmemInit().
243 */
244 if (i < MaxConnections)
245 {
246 /* PGPROC for normal backend, add to freeProcs list */
247 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->freeProcs;
248 ProcGlobal->freeProcs = &procs[i];
249 procs[i].procgloballist = &ProcGlobal->freeProcs;
250 }
251 else if (i < MaxConnections + autovacuum_max_workers + 1)
252 {
253 /* PGPROC for AV launcher/worker, add to autovacFreeProcs list */
254 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->autovacFreeProcs;
255 ProcGlobal->autovacFreeProcs = &procs[i];
256 procs[i].procgloballist = &ProcGlobal->autovacFreeProcs;
257 }
258 else if (i < MaxConnections + autovacuum_max_workers + 1 + max_worker_processes)
259 {
260 /* PGPROC for bgworker, add to bgworkerFreeProcs list */
261 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->bgworkerFreeProcs;
262 ProcGlobal->bgworkerFreeProcs = &procs[i];
263 procs[i].procgloballist = &ProcGlobal->bgworkerFreeProcs;
264 }
265 else if (i < MaxBackends)
266 {
267 /* PGPROC for walsender, add to walsenderFreeProcs list */
268 procs[i].links.next = (SHM_QUEUE *) ProcGlobal->walsenderFreeProcs;
269 ProcGlobal->walsenderFreeProcs = &procs[i];
270 procs[i].procgloballist = &ProcGlobal->walsenderFreeProcs;
271 }
272
273 /* Initialize myProcLocks[] shared memory queues. */
274 for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
275 SHMQueueInit(&(procs[i].myProcLocks[j]));
276
277 /* Initialize lockGroupMembers list. */
278 dlist_init(&procs[i].lockGroupMembers);
279
280 /*
281 * Initialize the atomic variables, otherwise, it won't be safe to
282 * access them for backends that aren't currently in use.
283 */
284 pg_atomic_init_u32(&(procs[i].procArrayGroupNext), INVALID_PGPROCNO);
285 pg_atomic_init_u32(&(procs[i].clogGroupNext), INVALID_PGPROCNO);
286 }
287
288 /*
289 * Save pointers to the blocks of PGPROC structures reserved for auxiliary
290 * processes and prepared transactions.
291 */
292 AuxiliaryProcs = &procs[MaxBackends];
293 PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
294
295 /* Create ProcStructLock spinlock, too */
296 ProcStructLock = (slock_t *) ShmemAlloc(sizeof(slock_t));
297 SpinLockInit(ProcStructLock);
298 }
299
300 /*
301 * InitProcess -- initialize a per-process data structure for this backend
302 */
303 void
InitProcess(void)304 InitProcess(void)
305 {
306 PGPROC *volatile *procgloballist;
307
308 /*
309 * ProcGlobal should be set up already (if we are a backend, we inherit
310 * this by fork() or EXEC_BACKEND mechanism from the postmaster).
311 */
312 if (ProcGlobal == NULL)
313 elog(PANIC, "proc header uninitialized");
314
315 if (MyProc != NULL)
316 elog(ERROR, "you already exist");
317
318 /* Decide which list should supply our PGPROC. */
319 if (IsAnyAutoVacuumProcess())
320 procgloballist = &ProcGlobal->autovacFreeProcs;
321 else if (IsBackgroundWorker)
322 procgloballist = &ProcGlobal->bgworkerFreeProcs;
323 else if (am_walsender)
324 procgloballist = &ProcGlobal->walsenderFreeProcs;
325 else
326 procgloballist = &ProcGlobal->freeProcs;
327
328 /*
329 * Try to get a proc struct from the appropriate free list. If this
330 * fails, we must be out of PGPROC structures (not to mention semaphores).
331 *
332 * While we are holding the ProcStructLock, also copy the current shared
333 * estimate of spins_per_delay to local storage.
334 */
335 SpinLockAcquire(ProcStructLock);
336
337 set_spins_per_delay(ProcGlobal->spins_per_delay);
338
339 MyProc = *procgloballist;
340
341 if (MyProc != NULL)
342 {
343 *procgloballist = (PGPROC *) MyProc->links.next;
344 SpinLockRelease(ProcStructLock);
345 }
346 else
347 {
348 /*
349 * If we reach here, all the PGPROCs are in use. This is one of the
350 * possible places to detect "too many backends", so give the standard
351 * error message. XXX do we need to give a different failure message
352 * in the autovacuum case?
353 */
354 SpinLockRelease(ProcStructLock);
355 if (am_walsender)
356 ereport(FATAL,
357 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
358 errmsg("number of requested standby connections exceeds max_wal_senders (currently %d)",
359 max_wal_senders)));
360 ereport(FATAL,
361 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
362 errmsg("sorry, too many clients already")));
363 }
364 MyPgXact = &ProcGlobal->allPgXact[MyProc->pgprocno];
365
366 /*
367 * Cross-check that the PGPROC is of the type we expect; if this were not
368 * the case, it would get returned to the wrong list.
369 */
370 Assert(MyProc->procgloballist == procgloballist);
371
372 /*
373 * Now that we have a PGPROC, mark ourselves as an active postmaster
374 * child; this is so that the postmaster can detect it if we exit without
375 * cleaning up. (XXX autovac launcher currently doesn't participate in
376 * this; it probably should.)
377 */
378 if (IsUnderPostmaster && !IsAutoVacuumLauncherProcess())
379 MarkPostmasterChildActive();
380
381 /*
382 * Initialize all fields of MyProc, except for those previously
383 * initialized by InitProcGlobal.
384 */
385 SHMQueueElemInit(&(MyProc->links));
386 MyProc->waitStatus = STATUS_OK;
387 MyProc->lxid = InvalidLocalTransactionId;
388 MyProc->fpVXIDLock = false;
389 MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
390 MyPgXact->xid = InvalidTransactionId;
391 MyPgXact->xmin = InvalidTransactionId;
392 MyProc->pid = MyProcPid;
393 /* backendId, databaseId and roleId will be filled in later */
394 MyProc->backendId = InvalidBackendId;
395 MyProc->databaseId = InvalidOid;
396 MyProc->roleId = InvalidOid;
397 MyProc->tempNamespaceId = InvalidOid;
398 MyProc->isBackgroundWorker = IsBackgroundWorker;
399 MyProc->delayChkpt = false;
400 MyPgXact->vacuumFlags = 0;
401 /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
402 if (IsAutoVacuumWorkerProcess())
403 MyPgXact->vacuumFlags |= PROC_IS_AUTOVACUUM;
404 MyProc->lwWaiting = false;
405 MyProc->lwWaitMode = 0;
406 MyProc->waitLock = NULL;
407 MyProc->waitProcLock = NULL;
408 #ifdef USE_ASSERT_CHECKING
409 {
410 int i;
411
412 /* Last process should have released all locks. */
413 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
414 Assert(SHMQueueEmpty(&(MyProc->myProcLocks[i])));
415 }
416 #endif
417 MyProc->recoveryConflictPending = false;
418
419 /* Initialize fields for sync rep */
420 MyProc->waitLSN = 0;
421 MyProc->syncRepState = SYNC_REP_NOT_WAITING;
422 SHMQueueElemInit(&(MyProc->syncRepLinks));
423
424 /* Initialize fields for group XID clearing. */
425 MyProc->procArrayGroupMember = false;
426 MyProc->procArrayGroupMemberXid = InvalidTransactionId;
427 Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PGPROCNO);
428
429 /* Check that group locking fields are in a proper initial state. */
430 Assert(MyProc->lockGroupLeader == NULL);
431 Assert(dlist_is_empty(&MyProc->lockGroupMembers));
432
433 /* Initialize wait event information. */
434 MyProc->wait_event_info = 0;
435
436 /* Initialize fields for group transaction status update. */
437 MyProc->clogGroupMember = false;
438 MyProc->clogGroupMemberXid = InvalidTransactionId;
439 MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
440 MyProc->clogGroupMemberPage = -1;
441 MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
442 Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PGPROCNO);
443
444 /*
445 * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
446 * on it. That allows us to repoint the process latch, which so far
447 * points to process local one, to the shared one.
448 */
449 OwnLatch(&MyProc->procLatch);
450 SwitchToSharedLatch();
451
452 /*
453 * We might be reusing a semaphore that belonged to a failed process. So
454 * be careful and reinitialize its value here. (This is not strictly
455 * necessary anymore, but seems like a good idea for cleanliness.)
456 */
457 PGSemaphoreReset(MyProc->sem);
458
459 /*
460 * Arrange to clean up at backend exit.
461 */
462 on_shmem_exit(ProcKill, 0);
463
464 /*
465 * Now that we have a PGPROC, we could try to acquire locks, so initialize
466 * local state needed for LWLocks, and the deadlock checker.
467 */
468 InitLWLockAccess();
469 InitDeadLockChecking();
470 }
471
472 /*
473 * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
474 *
475 * This is separate from InitProcess because we can't acquire LWLocks until
476 * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
477 * work until after we've done CreateSharedMemoryAndSemaphores.
478 */
479 void
InitProcessPhase2(void)480 InitProcessPhase2(void)
481 {
482 Assert(MyProc != NULL);
483
484 /*
485 * Add our PGPROC to the PGPROC array in shared memory.
486 */
487 ProcArrayAdd(MyProc);
488
489 /*
490 * Arrange to clean that up at backend exit.
491 */
492 on_shmem_exit(RemoveProcFromArray, 0);
493 }
494
495 /*
496 * InitAuxiliaryProcess -- create a per-auxiliary-process data structure
497 *
498 * This is called by bgwriter and similar processes so that they will have a
499 * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
500 * and sema that are assigned are one of the extra ones created during
501 * InitProcGlobal.
502 *
503 * Auxiliary processes are presently not expected to wait for real (lockmgr)
504 * locks, so we need not set up the deadlock checker. They are never added
505 * to the ProcArray or the sinval messaging mechanism, either. They also
506 * don't get a VXID assigned, since this is only useful when we actually
507 * hold lockmgr locks.
508 *
509 * Startup process however uses locks but never waits for them in the
510 * normal backend sense. Startup process also takes part in sinval messaging
511 * as a sendOnly process, so never reads messages from sinval queue. So
512 * Startup process does have a VXID and does show up in pg_locks.
513 */
514 void
InitAuxiliaryProcess(void)515 InitAuxiliaryProcess(void)
516 {
517 PGPROC *auxproc;
518 int proctype;
519
520 /*
521 * ProcGlobal should be set up already (if we are a backend, we inherit
522 * this by fork() or EXEC_BACKEND mechanism from the postmaster).
523 */
524 if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
525 elog(PANIC, "proc header uninitialized");
526
527 if (MyProc != NULL)
528 elog(ERROR, "you already exist");
529
530 /*
531 * We use the ProcStructLock to protect assignment and releasing of
532 * AuxiliaryProcs entries.
533 *
534 * While we are holding the ProcStructLock, also copy the current shared
535 * estimate of spins_per_delay to local storage.
536 */
537 SpinLockAcquire(ProcStructLock);
538
539 set_spins_per_delay(ProcGlobal->spins_per_delay);
540
541 /*
542 * Find a free auxproc ... *big* trouble if there isn't one ...
543 */
544 for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
545 {
546 auxproc = &AuxiliaryProcs[proctype];
547 if (auxproc->pid == 0)
548 break;
549 }
550 if (proctype >= NUM_AUXILIARY_PROCS)
551 {
552 SpinLockRelease(ProcStructLock);
553 elog(FATAL, "all AuxiliaryProcs are in use");
554 }
555
556 /* Mark auxiliary proc as in use by me */
557 /* use volatile pointer to prevent code rearrangement */
558 ((volatile PGPROC *) auxproc)->pid = MyProcPid;
559
560 MyProc = auxproc;
561 MyPgXact = &ProcGlobal->allPgXact[auxproc->pgprocno];
562
563 SpinLockRelease(ProcStructLock);
564
565 /*
566 * Initialize all fields of MyProc, except for those previously
567 * initialized by InitProcGlobal.
568 */
569 SHMQueueElemInit(&(MyProc->links));
570 MyProc->waitStatus = STATUS_OK;
571 MyProc->lxid = InvalidLocalTransactionId;
572 MyProc->fpVXIDLock = false;
573 MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
574 MyPgXact->xid = InvalidTransactionId;
575 MyPgXact->xmin = InvalidTransactionId;
576 MyProc->backendId = InvalidBackendId;
577 MyProc->databaseId = InvalidOid;
578 MyProc->roleId = InvalidOid;
579 MyProc->tempNamespaceId = InvalidOid;
580 MyProc->isBackgroundWorker = IsBackgroundWorker;
581 MyProc->delayChkpt = false;
582 MyPgXact->vacuumFlags = 0;
583 MyProc->lwWaiting = false;
584 MyProc->lwWaitMode = 0;
585 MyProc->waitLock = NULL;
586 MyProc->waitProcLock = NULL;
587 #ifdef USE_ASSERT_CHECKING
588 {
589 int i;
590
591 /* Last process should have released all locks. */
592 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
593 Assert(SHMQueueEmpty(&(MyProc->myProcLocks[i])));
594 }
595 #endif
596
597 /*
598 * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
599 * on it. That allows us to repoint the process latch, which so far
600 * points to process local one, to the shared one.
601 */
602 OwnLatch(&MyProc->procLatch);
603 SwitchToSharedLatch();
604
605 /* Check that group locking fields are in a proper initial state. */
606 Assert(MyProc->lockGroupLeader == NULL);
607 Assert(dlist_is_empty(&MyProc->lockGroupMembers));
608
609 /*
610 * We might be reusing a semaphore that belonged to a failed process. So
611 * be careful and reinitialize its value here. (This is not strictly
612 * necessary anymore, but seems like a good idea for cleanliness.)
613 */
614 PGSemaphoreReset(MyProc->sem);
615
616 /*
617 * Arrange to clean up at process exit.
618 */
619 on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
620 }
621
622 /*
623 * Record the PID and PGPROC structures for the Startup process, for use in
624 * ProcSendSignal(). See comments there for further explanation.
625 */
626 void
PublishStartupProcessInformation(void)627 PublishStartupProcessInformation(void)
628 {
629 SpinLockAcquire(ProcStructLock);
630
631 ProcGlobal->startupProc = MyProc;
632 ProcGlobal->startupProcPid = MyProcPid;
633
634 SpinLockRelease(ProcStructLock);
635 }
636
637 /*
638 * Used from bufmgr to share the value of the buffer that Startup waits on,
639 * or to reset the value to "not waiting" (-1). This allows processing
640 * of recovery conflicts for buffer pins. Set is made before backends look
641 * at this value, so locking not required, especially since the set is
642 * an atomic integer set operation.
643 */
644 void
SetStartupBufferPinWaitBufId(int bufid)645 SetStartupBufferPinWaitBufId(int bufid)
646 {
647 /* use volatile pointer to prevent code rearrangement */
648 volatile PROC_HDR *procglobal = ProcGlobal;
649
650 procglobal->startupBufferPinWaitBufId = bufid;
651 }
652
653 /*
654 * Used by backends when they receive a request to check for buffer pin waits.
655 */
656 int
GetStartupBufferPinWaitBufId(void)657 GetStartupBufferPinWaitBufId(void)
658 {
659 /* use volatile pointer to prevent code rearrangement */
660 volatile PROC_HDR *procglobal = ProcGlobal;
661
662 return procglobal->startupBufferPinWaitBufId;
663 }
664
665 /*
666 * Check whether there are at least N free PGPROC objects.
667 *
668 * Note: this is designed on the assumption that N will generally be small.
669 */
670 bool
HaveNFreeProcs(int n)671 HaveNFreeProcs(int n)
672 {
673 PGPROC *proc;
674
675 SpinLockAcquire(ProcStructLock);
676
677 proc = ProcGlobal->freeProcs;
678
679 while (n > 0 && proc != NULL)
680 {
681 proc = (PGPROC *) proc->links.next;
682 n--;
683 }
684
685 SpinLockRelease(ProcStructLock);
686
687 return (n <= 0);
688 }
689
690 /*
691 * Check if the current process is awaiting a lock.
692 */
693 bool
IsWaitingForLock(void)694 IsWaitingForLock(void)
695 {
696 if (lockAwaited == NULL)
697 return false;
698
699 return true;
700 }
701
702 /*
703 * Cancel any pending wait for lock, when aborting a transaction, and revert
704 * any strong lock count acquisition for a lock being acquired.
705 *
706 * (Normally, this would only happen if we accept a cancel/die
707 * interrupt while waiting; but an ereport(ERROR) before or during the lock
708 * wait is within the realm of possibility, too.)
709 */
710 void
LockErrorCleanup(void)711 LockErrorCleanup(void)
712 {
713 LWLock *partitionLock;
714 DisableTimeoutParams timeouts[2];
715
716 HOLD_INTERRUPTS();
717
718 AbortStrongLockAcquire();
719
720 /* Nothing to do if we weren't waiting for a lock */
721 if (lockAwaited == NULL)
722 {
723 RESUME_INTERRUPTS();
724 return;
725 }
726
727 /*
728 * Turn off the deadlock and lock timeout timers, if they are still
729 * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
730 * indicator flag, since this function is executed before
731 * ProcessInterrupts when responding to SIGINT; else we'd lose the
732 * knowledge that the SIGINT came from a lock timeout and not an external
733 * source.
734 */
735 timeouts[0].id = DEADLOCK_TIMEOUT;
736 timeouts[0].keep_indicator = false;
737 timeouts[1].id = LOCK_TIMEOUT;
738 timeouts[1].keep_indicator = true;
739 disable_timeouts(timeouts, 2);
740
741 /* Unlink myself from the wait queue, if on it (might not be anymore!) */
742 partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
743 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
744
745 if (MyProc->links.next != NULL)
746 {
747 /* We could not have been granted the lock yet */
748 RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
749 }
750 else
751 {
752 /*
753 * Somebody kicked us off the lock queue already. Perhaps they
754 * granted us the lock, or perhaps they detected a deadlock. If they
755 * did grant us the lock, we'd better remember it in our local lock
756 * table.
757 */
758 if (MyProc->waitStatus == STATUS_OK)
759 GrantAwaitedLock();
760 }
761
762 lockAwaited = NULL;
763
764 LWLockRelease(partitionLock);
765
766 RESUME_INTERRUPTS();
767 }
768
769
770 /*
771 * ProcReleaseLocks() -- release locks associated with current transaction
772 * at main transaction commit or abort
773 *
774 * At main transaction commit, we release standard locks except session locks.
775 * At main transaction abort, we release all locks including session locks.
776 *
777 * Advisory locks are released only if they are transaction-level;
778 * session-level holds remain, whether this is a commit or not.
779 *
780 * At subtransaction commit, we don't release any locks (so this func is not
781 * needed at all); we will defer the releasing to the parent transaction.
782 * At subtransaction abort, we release all locks held by the subtransaction;
783 * this is implemented by retail releasing of the locks under control of
784 * the ResourceOwner mechanism.
785 */
786 void
ProcReleaseLocks(bool isCommit)787 ProcReleaseLocks(bool isCommit)
788 {
789 if (!MyProc)
790 return;
791 /* If waiting, get off wait queue (should only be needed after error) */
792 LockErrorCleanup();
793 /* Release standard locks, including session-level if aborting */
794 LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
795 /* Release transaction-level advisory locks */
796 LockReleaseAll(USER_LOCKMETHOD, false);
797 }
798
799
800 /*
801 * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
802 */
803 static void
RemoveProcFromArray(int code,Datum arg)804 RemoveProcFromArray(int code, Datum arg)
805 {
806 Assert(MyProc != NULL);
807 ProcArrayRemove(MyProc, InvalidTransactionId);
808 }
809
810 /*
811 * ProcKill() -- Destroy the per-proc data structure for
812 * this process. Release any of its held LW locks.
813 */
814 static void
ProcKill(int code,Datum arg)815 ProcKill(int code, Datum arg)
816 {
817 PGPROC *proc;
818 PGPROC *volatile *procgloballist;
819
820 Assert(MyProc != NULL);
821
822 /* Make sure we're out of the sync rep lists */
823 SyncRepCleanupAtProcExit();
824
825 #ifdef USE_ASSERT_CHECKING
826 {
827 int i;
828
829 /* Last process should have released all locks. */
830 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
831 Assert(SHMQueueEmpty(&(MyProc->myProcLocks[i])));
832 }
833 #endif
834
835 /*
836 * Release any LW locks I am holding. There really shouldn't be any, but
837 * it's cheap to check again before we cut the knees off the LWLock
838 * facility by releasing our PGPROC ...
839 */
840 LWLockReleaseAll();
841
842 /* Cancel any pending condition variable sleep, too */
843 ConditionVariableCancelSleep();
844
845 /* Make sure active replication slots are released */
846 if (MyReplicationSlot != NULL)
847 ReplicationSlotRelease();
848
849 /* Also cleanup all the temporary slots. */
850 ReplicationSlotCleanup();
851
852 /*
853 * Detach from any lock group of which we are a member. If the leader
854 * exist before all other group members, its PGPROC will remain allocated
855 * until the last group process exits; that process must return the
856 * leader's PGPROC to the appropriate list.
857 */
858 if (MyProc->lockGroupLeader != NULL)
859 {
860 PGPROC *leader = MyProc->lockGroupLeader;
861 LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
862
863 LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
864 Assert(!dlist_is_empty(&leader->lockGroupMembers));
865 dlist_delete(&MyProc->lockGroupLink);
866 if (dlist_is_empty(&leader->lockGroupMembers))
867 {
868 leader->lockGroupLeader = NULL;
869 if (leader != MyProc)
870 {
871 procgloballist = leader->procgloballist;
872
873 /* Leader exited first; return its PGPROC. */
874 SpinLockAcquire(ProcStructLock);
875 leader->links.next = (SHM_QUEUE *) *procgloballist;
876 *procgloballist = leader;
877 SpinLockRelease(ProcStructLock);
878 }
879 }
880 else if (leader != MyProc)
881 MyProc->lockGroupLeader = NULL;
882 LWLockRelease(leader_lwlock);
883 }
884
885 /*
886 * Reset MyLatch to the process local one. This is so that signal
887 * handlers et al can continue using the latch after the shared latch
888 * isn't ours anymore. After that clear MyProc and disown the shared
889 * latch.
890 */
891 SwitchBackToLocalLatch();
892 proc = MyProc;
893 MyProc = NULL;
894 DisownLatch(&proc->procLatch);
895
896 procgloballist = proc->procgloballist;
897 SpinLockAcquire(ProcStructLock);
898
899 /*
900 * If we're still a member of a locking group, that means we're a leader
901 * which has somehow exited before its children. The last remaining child
902 * will release our PGPROC. Otherwise, release it now.
903 */
904 if (proc->lockGroupLeader == NULL)
905 {
906 /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
907 Assert(dlist_is_empty(&proc->lockGroupMembers));
908
909 /* Return PGPROC structure (and semaphore) to appropriate freelist */
910 proc->links.next = (SHM_QUEUE *) *procgloballist;
911 *procgloballist = proc;
912 }
913
914 /* Update shared estimate of spins_per_delay */
915 ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
916
917 SpinLockRelease(ProcStructLock);
918
919 /*
920 * This process is no longer present in shared memory in any meaningful
921 * way, so tell the postmaster we've cleaned up acceptably well. (XXX
922 * autovac launcher should be included here someday)
923 */
924 if (IsUnderPostmaster && !IsAutoVacuumLauncherProcess())
925 MarkPostmasterChildInactive();
926
927 /* wake autovac launcher if needed -- see comments in FreeWorkerInfo */
928 if (AutovacuumLauncherPid != 0)
929 kill(AutovacuumLauncherPid, SIGUSR2);
930 }
931
932 /*
933 * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
934 * processes (bgwriter, etc). The PGPROC and sema are not released, only
935 * marked as not-in-use.
936 */
937 static void
AuxiliaryProcKill(int code,Datum arg)938 AuxiliaryProcKill(int code, Datum arg)
939 {
940 int proctype = DatumGetInt32(arg);
941 PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
942 PGPROC *proc;
943
944 Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
945
946 auxproc = &AuxiliaryProcs[proctype];
947
948 Assert(MyProc == auxproc);
949
950 /* Release any LW locks I am holding (see notes above) */
951 LWLockReleaseAll();
952
953 /* Cancel any pending condition variable sleep, too */
954 ConditionVariableCancelSleep();
955
956 /*
957 * Reset MyLatch to the process local one. This is so that signal
958 * handlers et al can continue using the latch after the shared latch
959 * isn't ours anymore. After that clear MyProc and disown the shared
960 * latch.
961 */
962 SwitchBackToLocalLatch();
963 proc = MyProc;
964 MyProc = NULL;
965 DisownLatch(&proc->procLatch);
966
967 SpinLockAcquire(ProcStructLock);
968
969 /* Mark auxiliary proc no longer in use */
970 proc->pid = 0;
971
972 /* Update shared estimate of spins_per_delay */
973 ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
974
975 SpinLockRelease(ProcStructLock);
976 }
977
978 /*
979 * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
980 * given its PID
981 *
982 * Returns NULL if not found.
983 */
984 PGPROC *
AuxiliaryPidGetProc(int pid)985 AuxiliaryPidGetProc(int pid)
986 {
987 PGPROC *result = NULL;
988 int index;
989
990 if (pid == 0) /* never match dummy PGPROCs */
991 return NULL;
992
993 for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
994 {
995 PGPROC *proc = &AuxiliaryProcs[index];
996
997 if (proc->pid == pid)
998 {
999 result = proc;
1000 break;
1001 }
1002 }
1003 return result;
1004 }
1005
1006 /*
1007 * ProcQueue package: routines for putting processes to sleep
1008 * and waking them up
1009 */
1010
1011 /*
1012 * ProcQueueAlloc -- alloc/attach to a shared memory process queue
1013 *
1014 * Returns: a pointer to the queue
1015 * Side Effects: Initializes the queue if it wasn't there before
1016 */
1017 #ifdef NOT_USED
1018 PROC_QUEUE *
ProcQueueAlloc(const char * name)1019 ProcQueueAlloc(const char *name)
1020 {
1021 PROC_QUEUE *queue;
1022 bool found;
1023
1024 queue = (PROC_QUEUE *)
1025 ShmemInitStruct(name, sizeof(PROC_QUEUE), &found);
1026
1027 if (!found)
1028 ProcQueueInit(queue);
1029
1030 return queue;
1031 }
1032 #endif
1033
1034 /*
1035 * ProcQueueInit -- initialize a shared memory process queue
1036 */
1037 void
ProcQueueInit(PROC_QUEUE * queue)1038 ProcQueueInit(PROC_QUEUE *queue)
1039 {
1040 SHMQueueInit(&(queue->links));
1041 queue->size = 0;
1042 }
1043
1044
1045 /*
1046 * ProcSleep -- put a process to sleep on the specified lock
1047 *
1048 * Caller must have set MyProc->heldLocks to reflect locks already held
1049 * on the lockable object by this process (under all XIDs).
1050 *
1051 * The lock table's partition lock must be held at entry, and will be held
1052 * at exit.
1053 *
1054 * Result: STATUS_OK if we acquired the lock, STATUS_ERROR if not (deadlock).
1055 *
1056 * ASSUME: that no one will fiddle with the queue until after
1057 * we release the partition lock.
1058 *
1059 * NOTES: The process queue is now a priority queue for locking.
1060 */
1061 int
ProcSleep(LOCALLOCK * locallock,LockMethod lockMethodTable)1062 ProcSleep(LOCALLOCK *locallock, LockMethod lockMethodTable)
1063 {
1064 LOCKMODE lockmode = locallock->tag.mode;
1065 LOCK *lock = locallock->lock;
1066 PROCLOCK *proclock = locallock->proclock;
1067 uint32 hashcode = locallock->hashcode;
1068 LWLock *partitionLock = LockHashPartitionLock(hashcode);
1069 PROC_QUEUE *waitQueue = &(lock->waitProcs);
1070 LOCKMASK myHeldLocks = MyProc->heldLocks;
1071 bool early_deadlock = false;
1072 bool allow_autovacuum_cancel = true;
1073 int myWaitStatus;
1074 PGPROC *proc;
1075 PGPROC *leader = MyProc->lockGroupLeader;
1076 int i;
1077
1078 /*
1079 * If group locking is in use, locks held by members of my locking group
1080 * need to be included in myHeldLocks. This is not required for relation
1081 * extension or page locks which conflict among group members. However,
1082 * including them in myHeldLocks will give group members the priority to
1083 * get those locks as compared to other backends which are also trying to
1084 * acquire those locks. OTOH, we can avoid giving priority to group
1085 * members for that kind of locks, but there doesn't appear to be a clear
1086 * advantage of the same.
1087 */
1088 if (leader != NULL)
1089 {
1090 SHM_QUEUE *procLocks = &(lock->procLocks);
1091 PROCLOCK *otherproclock;
1092
1093 otherproclock = (PROCLOCK *)
1094 SHMQueueNext(procLocks, procLocks, offsetof(PROCLOCK, lockLink));
1095 while (otherproclock != NULL)
1096 {
1097 if (otherproclock->groupLeader == leader)
1098 myHeldLocks |= otherproclock->holdMask;
1099 otherproclock = (PROCLOCK *)
1100 SHMQueueNext(procLocks, &otherproclock->lockLink,
1101 offsetof(PROCLOCK, lockLink));
1102 }
1103 }
1104
1105 /*
1106 * Determine where to add myself in the wait queue.
1107 *
1108 * Normally I should go at the end of the queue. However, if I already
1109 * hold locks that conflict with the request of any previous waiter, put
1110 * myself in the queue just in front of the first such waiter. This is not
1111 * a necessary step, since deadlock detection would move me to before that
1112 * waiter anyway; but it's relatively cheap to detect such a conflict
1113 * immediately, and avoid delaying till deadlock timeout.
1114 *
1115 * Special case: if I find I should go in front of some waiter, check to
1116 * see if I conflict with already-held locks or the requests before that
1117 * waiter. If not, then just grant myself the requested lock immediately.
1118 * This is the same as the test for immediate grant in LockAcquire, except
1119 * we are only considering the part of the wait queue before my insertion
1120 * point.
1121 */
1122 if (myHeldLocks != 0)
1123 {
1124 LOCKMASK aheadRequests = 0;
1125
1126 proc = (PGPROC *) waitQueue->links.next;
1127 for (i = 0; i < waitQueue->size; i++)
1128 {
1129 /*
1130 * If we're part of the same locking group as this waiter, its
1131 * locks neither conflict with ours nor contribute to
1132 * aheadRequests.
1133 */
1134 if (leader != NULL && leader == proc->lockGroupLeader)
1135 {
1136 proc = (PGPROC *) proc->links.next;
1137 continue;
1138 }
1139 /* Must he wait for me? */
1140 if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1141 {
1142 /* Must I wait for him ? */
1143 if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1144 {
1145 /*
1146 * Yes, so we have a deadlock. Easiest way to clean up
1147 * correctly is to call RemoveFromWaitQueue(), but we
1148 * can't do that until we are *on* the wait queue. So, set
1149 * a flag to check below, and break out of loop. Also,
1150 * record deadlock info for later message.
1151 */
1152 RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1153 early_deadlock = true;
1154 break;
1155 }
1156 /* I must go before this waiter. Check special case. */
1157 if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1158 !LockCheckConflicts(lockMethodTable, lockmode, lock,
1159 proclock))
1160 {
1161 /* Skip the wait and just grant myself the lock. */
1162 GrantLock(lock, proclock, lockmode);
1163 GrantAwaitedLock();
1164 return STATUS_OK;
1165 }
1166 /* Break out of loop to put myself before him */
1167 break;
1168 }
1169 /* Nope, so advance to next waiter */
1170 aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1171 proc = (PGPROC *) proc->links.next;
1172 }
1173
1174 /*
1175 * If we fall out of loop normally, proc points to waitQueue head, so
1176 * we will insert at tail of queue as desired.
1177 */
1178 }
1179 else
1180 {
1181 /* I hold no locks, so I can't push in front of anyone. */
1182 proc = (PGPROC *) &(waitQueue->links);
1183 }
1184
1185 /*
1186 * Insert self into queue, ahead of the given proc (or at tail of queue).
1187 */
1188 SHMQueueInsertBefore(&(proc->links), &(MyProc->links));
1189 waitQueue->size++;
1190
1191 lock->waitMask |= LOCKBIT_ON(lockmode);
1192
1193 /* Set up wait information in PGPROC object, too */
1194 MyProc->waitLock = lock;
1195 MyProc->waitProcLock = proclock;
1196 MyProc->waitLockMode = lockmode;
1197
1198 MyProc->waitStatus = STATUS_WAITING;
1199
1200 /*
1201 * If we detected deadlock, give up without waiting. This must agree with
1202 * CheckDeadLock's recovery code.
1203 */
1204 if (early_deadlock)
1205 {
1206 RemoveFromWaitQueue(MyProc, hashcode);
1207 return STATUS_ERROR;
1208 }
1209
1210 /* mark that we are waiting for a lock */
1211 lockAwaited = locallock;
1212
1213 /*
1214 * Release the lock table's partition lock.
1215 *
1216 * NOTE: this may also cause us to exit critical-section state, possibly
1217 * allowing a cancel/die interrupt to be accepted. This is OK because we
1218 * have recorded the fact that we are waiting for a lock, and so
1219 * LockErrorCleanup will clean up if cancel/die happens.
1220 */
1221 LWLockRelease(partitionLock);
1222
1223 /*
1224 * Also, now that we will successfully clean up after an ereport, it's
1225 * safe to check to see if there's a buffer pin deadlock against the
1226 * Startup process. Of course, that's only necessary if we're doing Hot
1227 * Standby and are not the Startup process ourselves.
1228 */
1229 if (RecoveryInProgress() && !InRecovery)
1230 CheckRecoveryConflictDeadlock();
1231
1232 /* Reset deadlock_state before enabling the timeout handler */
1233 deadlock_state = DS_NOT_YET_CHECKED;
1234 got_deadlock_timeout = false;
1235
1236 /*
1237 * Set timer so we can wake up after awhile and check for a deadlock. If a
1238 * deadlock is detected, the handler sets MyProc->waitStatus =
1239 * STATUS_ERROR, allowing us to know that we must report failure rather
1240 * than success.
1241 *
1242 * By delaying the check until we've waited for a bit, we can avoid
1243 * running the rather expensive deadlock-check code in most cases.
1244 *
1245 * If LockTimeout is set, also enable the timeout for that. We can save a
1246 * few cycles by enabling both timeout sources in one call.
1247 *
1248 * If InHotStandby we set lock waits slightly later for clarity with other
1249 * code.
1250 */
1251 if (!InHotStandby)
1252 {
1253 if (LockTimeout > 0)
1254 {
1255 EnableTimeoutParams timeouts[2];
1256
1257 timeouts[0].id = DEADLOCK_TIMEOUT;
1258 timeouts[0].type = TMPARAM_AFTER;
1259 timeouts[0].delay_ms = DeadlockTimeout;
1260 timeouts[1].id = LOCK_TIMEOUT;
1261 timeouts[1].type = TMPARAM_AFTER;
1262 timeouts[1].delay_ms = LockTimeout;
1263 enable_timeouts(timeouts, 2);
1264 }
1265 else
1266 enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1267 }
1268
1269 /*
1270 * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1271 * will not wait. But a set latch does not necessarily mean that the lock
1272 * is free now, as there are many other sources for latch sets than
1273 * somebody releasing the lock.
1274 *
1275 * We process interrupts whenever the latch has been set, so cancel/die
1276 * interrupts are processed quickly. This means we must not mind losing
1277 * control to a cancel/die interrupt here. We don't, because we have no
1278 * shared-state-change work to do after being granted the lock (the
1279 * grantor did it all). We do have to worry about canceling the deadlock
1280 * timeout and updating the locallock table, but if we lose control to an
1281 * error, LockErrorCleanup will fix that up.
1282 */
1283 do
1284 {
1285 if (InHotStandby)
1286 {
1287 /* Set a timer and wait for that or for the Lock to be granted */
1288 ResolveRecoveryConflictWithLock(locallock->tag.lock);
1289 }
1290 else
1291 {
1292 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1293 PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1294 ResetLatch(MyLatch);
1295 /* check for deadlocks first, as that's probably log-worthy */
1296 if (got_deadlock_timeout)
1297 {
1298 CheckDeadLock();
1299 got_deadlock_timeout = false;
1300 }
1301 CHECK_FOR_INTERRUPTS();
1302 }
1303
1304 /*
1305 * waitStatus could change from STATUS_WAITING to something else
1306 * asynchronously. Read it just once per loop to prevent surprising
1307 * behavior (such as missing log messages).
1308 */
1309 myWaitStatus = *((volatile int *) &MyProc->waitStatus);
1310
1311 /*
1312 * If we are not deadlocked, but are waiting on an autovacuum-induced
1313 * task, send a signal to interrupt it.
1314 */
1315 if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1316 {
1317 PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1318 PGXACT *autovac_pgxact = &ProcGlobal->allPgXact[autovac->pgprocno];
1319
1320 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1321
1322 /*
1323 * Only do it if the worker is not working to protect against Xid
1324 * wraparound.
1325 */
1326 if ((autovac_pgxact->vacuumFlags & PROC_IS_AUTOVACUUM) &&
1327 !(autovac_pgxact->vacuumFlags & PROC_VACUUM_FOR_WRAPAROUND))
1328 {
1329 int pid = autovac->pid;
1330 StringInfoData locktagbuf;
1331 StringInfoData logbuf; /* errdetail for server log */
1332
1333 initStringInfo(&locktagbuf);
1334 initStringInfo(&logbuf);
1335 DescribeLockTag(&locktagbuf, &lock->tag);
1336 appendStringInfo(&logbuf,
1337 _("Process %d waits for %s on %s."),
1338 MyProcPid,
1339 GetLockmodeName(lock->tag.locktag_lockmethodid,
1340 lockmode),
1341 locktagbuf.data);
1342
1343 /* release lock as quickly as possible */
1344 LWLockRelease(ProcArrayLock);
1345
1346 /* send the autovacuum worker Back to Old Kent Road */
1347 ereport(DEBUG1,
1348 (errmsg("sending cancel to blocking autovacuum PID %d",
1349 pid),
1350 errdetail_log("%s", logbuf.data)));
1351
1352 if (kill(pid, SIGINT) < 0)
1353 {
1354 /*
1355 * There's a race condition here: once we release the
1356 * ProcArrayLock, it's possible for the autovac worker to
1357 * close up shop and exit before we can do the kill().
1358 * Therefore, we do not whinge about no-such-process.
1359 * Other errors such as EPERM could conceivably happen if
1360 * the kernel recycles the PID fast enough, but such cases
1361 * seem improbable enough that it's probably best to issue
1362 * a warning if we see some other errno.
1363 */
1364 if (errno != ESRCH)
1365 ereport(WARNING,
1366 (errmsg("could not send signal to process %d: %m",
1367 pid)));
1368 }
1369
1370 pfree(logbuf.data);
1371 pfree(locktagbuf.data);
1372 }
1373 else
1374 LWLockRelease(ProcArrayLock);
1375
1376 /* prevent signal from being sent again more than once */
1377 allow_autovacuum_cancel = false;
1378 }
1379
1380 /*
1381 * If awoken after the deadlock check interrupt has run, and
1382 * log_lock_waits is on, then report about the wait.
1383 */
1384 if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1385 {
1386 StringInfoData buf,
1387 lock_waiters_sbuf,
1388 lock_holders_sbuf;
1389 const char *modename;
1390 long secs;
1391 int usecs;
1392 long msecs;
1393 SHM_QUEUE *procLocks;
1394 PROCLOCK *proclock;
1395 bool first_holder = true,
1396 first_waiter = true;
1397 int lockHoldersNum = 0;
1398
1399 initStringInfo(&buf);
1400 initStringInfo(&lock_waiters_sbuf);
1401 initStringInfo(&lock_holders_sbuf);
1402
1403 DescribeLockTag(&buf, &locallock->tag.lock);
1404 modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1405 lockmode);
1406 TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1407 GetCurrentTimestamp(),
1408 &secs, &usecs);
1409 msecs = secs * 1000 + usecs / 1000;
1410 usecs = usecs % 1000;
1411
1412 /*
1413 * we loop over the lock's procLocks to gather a list of all
1414 * holders and waiters. Thus we will be able to provide more
1415 * detailed information for lock debugging purposes.
1416 *
1417 * lock->procLocks contains all processes which hold or wait for
1418 * this lock.
1419 */
1420
1421 LWLockAcquire(partitionLock, LW_SHARED);
1422
1423 procLocks = &(lock->procLocks);
1424 proclock = (PROCLOCK *) SHMQueueNext(procLocks, procLocks,
1425 offsetof(PROCLOCK, lockLink));
1426
1427 while (proclock)
1428 {
1429 /*
1430 * we are a waiter if myProc->waitProcLock == proclock; we are
1431 * a holder if it is NULL or something different
1432 */
1433 if (proclock->tag.myProc->waitProcLock == proclock)
1434 {
1435 if (first_waiter)
1436 {
1437 appendStringInfo(&lock_waiters_sbuf, "%d",
1438 proclock->tag.myProc->pid);
1439 first_waiter = false;
1440 }
1441 else
1442 appendStringInfo(&lock_waiters_sbuf, ", %d",
1443 proclock->tag.myProc->pid);
1444 }
1445 else
1446 {
1447 if (first_holder)
1448 {
1449 appendStringInfo(&lock_holders_sbuf, "%d",
1450 proclock->tag.myProc->pid);
1451 first_holder = false;
1452 }
1453 else
1454 appendStringInfo(&lock_holders_sbuf, ", %d",
1455 proclock->tag.myProc->pid);
1456
1457 lockHoldersNum++;
1458 }
1459
1460 proclock = (PROCLOCK *) SHMQueueNext(procLocks, &proclock->lockLink,
1461 offsetof(PROCLOCK, lockLink));
1462 }
1463
1464 LWLockRelease(partitionLock);
1465
1466 if (deadlock_state == DS_SOFT_DEADLOCK)
1467 ereport(LOG,
1468 (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1469 MyProcPid, modename, buf.data, msecs, usecs),
1470 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1471 "Processes holding the lock: %s. Wait queue: %s.",
1472 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1473 else if (deadlock_state == DS_HARD_DEADLOCK)
1474 {
1475 /*
1476 * This message is a bit redundant with the error that will be
1477 * reported subsequently, but in some cases the error report
1478 * might not make it to the log (eg, if it's caught by an
1479 * exception handler), and we want to ensure all long-wait
1480 * events get logged.
1481 */
1482 ereport(LOG,
1483 (errmsg("process %d detected deadlock while waiting for %s on %s after %ld.%03d ms",
1484 MyProcPid, modename, buf.data, msecs, usecs),
1485 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1486 "Processes holding the lock: %s. Wait queue: %s.",
1487 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1488 }
1489
1490 if (myWaitStatus == STATUS_WAITING)
1491 ereport(LOG,
1492 (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1493 MyProcPid, modename, buf.data, msecs, usecs),
1494 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1495 "Processes holding the lock: %s. Wait queue: %s.",
1496 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1497 else if (myWaitStatus == STATUS_OK)
1498 ereport(LOG,
1499 (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1500 MyProcPid, modename, buf.data, msecs, usecs)));
1501 else
1502 {
1503 Assert(myWaitStatus == STATUS_ERROR);
1504
1505 /*
1506 * Currently, the deadlock checker always kicks its own
1507 * process, which means that we'll only see STATUS_ERROR when
1508 * deadlock_state == DS_HARD_DEADLOCK, and there's no need to
1509 * print redundant messages. But for completeness and
1510 * future-proofing, print a message if it looks like someone
1511 * else kicked us off the lock.
1512 */
1513 if (deadlock_state != DS_HARD_DEADLOCK)
1514 ereport(LOG,
1515 (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1516 MyProcPid, modename, buf.data, msecs, usecs),
1517 (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1518 "Processes holding the lock: %s. Wait queue: %s.",
1519 lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1520 }
1521
1522 /*
1523 * At this point we might still need to wait for the lock. Reset
1524 * state so we don't print the above messages again.
1525 */
1526 deadlock_state = DS_NO_DEADLOCK;
1527
1528 pfree(buf.data);
1529 pfree(lock_holders_sbuf.data);
1530 pfree(lock_waiters_sbuf.data);
1531 }
1532 } while (myWaitStatus == STATUS_WAITING);
1533
1534 /*
1535 * Disable the timers, if they are still running. As in LockErrorCleanup,
1536 * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1537 * already caused QueryCancelPending to become set, we want the cancel to
1538 * be reported as a lock timeout, not a user cancel.
1539 */
1540 if (!InHotStandby)
1541 {
1542 if (LockTimeout > 0)
1543 {
1544 DisableTimeoutParams timeouts[2];
1545
1546 timeouts[0].id = DEADLOCK_TIMEOUT;
1547 timeouts[0].keep_indicator = false;
1548 timeouts[1].id = LOCK_TIMEOUT;
1549 timeouts[1].keep_indicator = true;
1550 disable_timeouts(timeouts, 2);
1551 }
1552 else
1553 disable_timeout(DEADLOCK_TIMEOUT, false);
1554 }
1555
1556 /*
1557 * Re-acquire the lock table's partition lock. We have to do this to hold
1558 * off cancel/die interrupts before we can mess with lockAwaited (else we
1559 * might have a missed or duplicated locallock update).
1560 */
1561 LWLockAcquire(partitionLock, LW_EXCLUSIVE);
1562
1563 /*
1564 * We no longer want LockErrorCleanup to do anything.
1565 */
1566 lockAwaited = NULL;
1567
1568 /*
1569 * If we got the lock, be sure to remember it in the locallock table.
1570 */
1571 if (MyProc->waitStatus == STATUS_OK)
1572 GrantAwaitedLock();
1573
1574 /*
1575 * We don't have to do anything else, because the awaker did all the
1576 * necessary update of the lock table and MyProc.
1577 */
1578 return MyProc->waitStatus;
1579 }
1580
1581
1582 /*
1583 * ProcWakeup -- wake up a process by setting its latch.
1584 *
1585 * Also remove the process from the wait queue and set its links invalid.
1586 * RETURN: the next process in the wait queue.
1587 *
1588 * The appropriate lock partition lock must be held by caller.
1589 *
1590 * XXX: presently, this code is only used for the "success" case, and only
1591 * works correctly for that case. To clean up in failure case, would need
1592 * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1593 * Hence, in practice the waitStatus parameter must be STATUS_OK.
1594 */
1595 PGPROC *
ProcWakeup(PGPROC * proc,int waitStatus)1596 ProcWakeup(PGPROC *proc, int waitStatus)
1597 {
1598 PGPROC *retProc;
1599
1600 /* Proc should be sleeping ... */
1601 if (proc->links.prev == NULL ||
1602 proc->links.next == NULL)
1603 return NULL;
1604 Assert(proc->waitStatus == STATUS_WAITING);
1605
1606 /* Save next process before we zap the list link */
1607 retProc = (PGPROC *) proc->links.next;
1608
1609 /* Remove process from wait queue */
1610 SHMQueueDelete(&(proc->links));
1611 (proc->waitLock->waitProcs.size)--;
1612
1613 /* Clean up process' state and pass it the ok/fail signal */
1614 proc->waitLock = NULL;
1615 proc->waitProcLock = NULL;
1616 proc->waitStatus = waitStatus;
1617
1618 /* And awaken it */
1619 SetLatch(&proc->procLatch);
1620
1621 return retProc;
1622 }
1623
1624 /*
1625 * ProcLockWakeup -- routine for waking up processes when a lock is
1626 * released (or a prior waiter is aborted). Scan all waiters
1627 * for lock, waken any that are no longer blocked.
1628 *
1629 * The appropriate lock partition lock must be held by caller.
1630 */
1631 void
ProcLockWakeup(LockMethod lockMethodTable,LOCK * lock)1632 ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1633 {
1634 PROC_QUEUE *waitQueue = &(lock->waitProcs);
1635 int queue_size = waitQueue->size;
1636 PGPROC *proc;
1637 LOCKMASK aheadRequests = 0;
1638
1639 Assert(queue_size >= 0);
1640
1641 if (queue_size == 0)
1642 return;
1643
1644 proc = (PGPROC *) waitQueue->links.next;
1645
1646 while (queue_size-- > 0)
1647 {
1648 LOCKMODE lockmode = proc->waitLockMode;
1649
1650 /*
1651 * Waken if (a) doesn't conflict with requests of earlier waiters, and
1652 * (b) doesn't conflict with already-held locks.
1653 */
1654 if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1655 !LockCheckConflicts(lockMethodTable, lockmode, lock,
1656 proc->waitProcLock))
1657 {
1658 /* OK to waken */
1659 GrantLock(lock, proc->waitProcLock, lockmode);
1660 proc = ProcWakeup(proc, STATUS_OK);
1661
1662 /*
1663 * ProcWakeup removes proc from the lock's waiting process queue
1664 * and returns the next proc in chain; don't use proc's next-link,
1665 * because it's been cleared.
1666 */
1667 }
1668 else
1669 {
1670 /*
1671 * Cannot wake this guy. Remember his request for later checks.
1672 */
1673 aheadRequests |= LOCKBIT_ON(lockmode);
1674 proc = (PGPROC *) proc->links.next;
1675 }
1676 }
1677
1678 Assert(waitQueue->size >= 0);
1679 }
1680
1681 /*
1682 * CheckDeadLock
1683 *
1684 * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1685 * lock to be released by some other process. Check if there's a deadlock; if
1686 * not, just return. (But signal ProcSleep to log a message, if
1687 * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1688 * the lock's wait queue and signal an error to ProcSleep.
1689 */
1690 static void
CheckDeadLock(void)1691 CheckDeadLock(void)
1692 {
1693 int i;
1694
1695 /*
1696 * Acquire exclusive lock on the entire shared lock data structures. Must
1697 * grab LWLocks in partition-number order to avoid LWLock deadlock.
1698 *
1699 * Note that the deadlock check interrupt had better not be enabled
1700 * anywhere that this process itself holds lock partition locks, else this
1701 * will wait forever. Also note that LWLockAcquire creates a critical
1702 * section, so that this routine cannot be interrupted by cancel/die
1703 * interrupts.
1704 */
1705 for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1706 LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1707
1708 /*
1709 * Check to see if we've been awoken by anyone in the interim.
1710 *
1711 * If we have, we can return and resume our transaction -- happy day.
1712 * Before we are awoken the process releasing the lock grants it to us so
1713 * we know that we don't have to wait anymore.
1714 *
1715 * We check by looking to see if we've been unlinked from the wait queue.
1716 * This is safe because we hold the lock partition lock.
1717 */
1718 if (MyProc->links.prev == NULL ||
1719 MyProc->links.next == NULL)
1720 goto check_done;
1721
1722 #ifdef LOCK_DEBUG
1723 if (Debug_deadlocks)
1724 DumpAllLocks();
1725 #endif
1726
1727 /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1728 deadlock_state = DeadLockCheck(MyProc);
1729
1730 if (deadlock_state == DS_HARD_DEADLOCK)
1731 {
1732 /*
1733 * Oops. We have a deadlock.
1734 *
1735 * Get this process out of wait state. (Note: we could do this more
1736 * efficiently by relying on lockAwaited, but use this coding to
1737 * preserve the flexibility to kill some other transaction than the
1738 * one detecting the deadlock.)
1739 *
1740 * RemoveFromWaitQueue sets MyProc->waitStatus to STATUS_ERROR, so
1741 * ProcSleep will report an error after we return from the signal
1742 * handler.
1743 */
1744 Assert(MyProc->waitLock != NULL);
1745 RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1746
1747 /*
1748 * We're done here. Transaction abort caused by the error that
1749 * ProcSleep will raise will cause any other locks we hold to be
1750 * released, thus allowing other processes to wake up; we don't need
1751 * to do that here. NOTE: an exception is that releasing locks we
1752 * hold doesn't consider the possibility of waiters that were blocked
1753 * behind us on the lock we just failed to get, and might now be
1754 * wakable because we're not in front of them anymore. However,
1755 * RemoveFromWaitQueue took care of waking up any such processes.
1756 */
1757 }
1758
1759 /*
1760 * And release locks. We do this in reverse order for two reasons: (1)
1761 * Anyone else who needs more than one of the locks will be trying to lock
1762 * them in increasing order; we don't want to release the other process
1763 * until it can get all the locks it needs. (2) This avoids O(N^2)
1764 * behavior inside LWLockRelease.
1765 */
1766 check_done:
1767 for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1768 LWLockRelease(LockHashPartitionLockByIndex(i));
1769 }
1770
1771 /*
1772 * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1773 *
1774 * NB: Runs inside a signal handler, be careful.
1775 */
1776 void
CheckDeadLockAlert(void)1777 CheckDeadLockAlert(void)
1778 {
1779 int save_errno = errno;
1780
1781 got_deadlock_timeout = true;
1782
1783 /*
1784 * Have to set the latch again, even if handle_sig_alarm already did. Back
1785 * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1786 * ever would be a problem, but setting a set latch again is cheap.
1787 *
1788 * Note that, when this function runs inside procsignal_sigusr1_handler(),
1789 * the handler function sets the latch again after the latch is set here.
1790 */
1791 SetLatch(MyLatch);
1792 errno = save_errno;
1793 }
1794
1795 /*
1796 * ProcWaitForSignal - wait for a signal from another backend.
1797 *
1798 * As this uses the generic process latch the caller has to be robust against
1799 * unrelated wakeups: Always check that the desired state has occurred, and
1800 * wait again if not.
1801 */
1802 void
ProcWaitForSignal(uint32 wait_event_info)1803 ProcWaitForSignal(uint32 wait_event_info)
1804 {
1805 (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1806 wait_event_info);
1807 ResetLatch(MyLatch);
1808 CHECK_FOR_INTERRUPTS();
1809 }
1810
1811 /*
1812 * ProcSendSignal - send a signal to a backend identified by PID
1813 */
1814 void
ProcSendSignal(int pid)1815 ProcSendSignal(int pid)
1816 {
1817 PGPROC *proc = NULL;
1818
1819 if (RecoveryInProgress())
1820 {
1821 SpinLockAcquire(ProcStructLock);
1822
1823 /*
1824 * Check to see whether it is the Startup process we wish to signal.
1825 * This call is made by the buffer manager when it wishes to wake up a
1826 * process that has been waiting for a pin in so it can obtain a
1827 * cleanup lock using LockBufferForCleanup(). Startup is not a normal
1828 * backend, so BackendPidGetProc() will not return any pid at all. So
1829 * we remember the information for this special case.
1830 */
1831 if (pid == ProcGlobal->startupProcPid)
1832 proc = ProcGlobal->startupProc;
1833
1834 SpinLockRelease(ProcStructLock);
1835 }
1836
1837 if (proc == NULL)
1838 proc = BackendPidGetProc(pid);
1839
1840 if (proc != NULL)
1841 {
1842 SetLatch(&proc->procLatch);
1843 }
1844 }
1845
1846 /*
1847 * BecomeLockGroupLeader - designate process as lock group leader
1848 *
1849 * Once this function has returned, other processes can join the lock group
1850 * by calling BecomeLockGroupMember.
1851 */
1852 void
BecomeLockGroupLeader(void)1853 BecomeLockGroupLeader(void)
1854 {
1855 LWLock *leader_lwlock;
1856
1857 /* If we already did it, we don't need to do it again. */
1858 if (MyProc->lockGroupLeader == MyProc)
1859 return;
1860
1861 /* We had better not be a follower. */
1862 Assert(MyProc->lockGroupLeader == NULL);
1863
1864 /* Create single-member group, containing only ourselves. */
1865 leader_lwlock = LockHashPartitionLockByProc(MyProc);
1866 LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1867 MyProc->lockGroupLeader = MyProc;
1868 dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
1869 LWLockRelease(leader_lwlock);
1870 }
1871
1872 /*
1873 * BecomeLockGroupMember - designate process as lock group member
1874 *
1875 * This is pretty straightforward except for the possibility that the leader
1876 * whose group we're trying to join might exit before we manage to do so;
1877 * and the PGPROC might get recycled for an unrelated process. To avoid
1878 * that, we require the caller to pass the PID of the intended PGPROC as
1879 * an interlock. Returns true if we successfully join the intended lock
1880 * group, and false if not.
1881 */
1882 bool
BecomeLockGroupMember(PGPROC * leader,int pid)1883 BecomeLockGroupMember(PGPROC *leader, int pid)
1884 {
1885 LWLock *leader_lwlock;
1886 bool ok = false;
1887
1888 /* Group leader can't become member of group */
1889 Assert(MyProc != leader);
1890
1891 /* Can't already be a member of a group */
1892 Assert(MyProc->lockGroupLeader == NULL);
1893
1894 /* PID must be valid. */
1895 Assert(pid != 0);
1896
1897 /*
1898 * Get lock protecting the group fields. Note LockHashPartitionLockByProc
1899 * accesses leader->pgprocno in a PGPROC that might be free. This is safe
1900 * because all PGPROCs' pgprocno fields are set during shared memory
1901 * initialization and never change thereafter; so we will acquire the
1902 * correct lock even if the leader PGPROC is in process of being recycled.
1903 */
1904 leader_lwlock = LockHashPartitionLockByProc(leader);
1905 LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1906
1907 /* Is this the leader we're looking for? */
1908 if (leader->pid == pid && leader->lockGroupLeader == leader)
1909 {
1910 /* OK, join the group */
1911 ok = true;
1912 MyProc->lockGroupLeader = leader;
1913 dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
1914 }
1915 LWLockRelease(leader_lwlock);
1916
1917 return ok;
1918 }
1919