Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * proc.c
4 : * routines to manage per-process shared memory data structure
5 : *
6 : * Portions Copyright (c) 1996-2025, 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 : * JoinWaitQueue(), ProcSleep(), ProcWakeup()
18 : *
19 : * Waiting for a lock causes the backend to be put to sleep. Whoever releases
20 : * the lock wakes the process up again (and gives it an error code so it knows
21 : * whether it was awoken on an error condition).
22 : *
23 : * Interface (b):
24 : *
25 : * ProcReleaseLocks -- frees the locks associated with current transaction
26 : *
27 : * ProcKill -- destroys the shared memory state (and locks)
28 : * associated with the process.
29 : */
30 : #include "postgres.h"
31 :
32 : #include <signal.h>
33 : #include <unistd.h>
34 : #include <sys/time.h>
35 :
36 : #include "access/transam.h"
37 : #include "access/twophase.h"
38 : #include "access/xlogutils.h"
39 : #include "access/xlogwait.h"
40 : #include "miscadmin.h"
41 : #include "pgstat.h"
42 : #include "postmaster/autovacuum.h"
43 : #include "replication/slotsync.h"
44 : #include "replication/syncrep.h"
45 : #include "storage/condition_variable.h"
46 : #include "storage/ipc.h"
47 : #include "storage/lmgr.h"
48 : #include "storage/pmsignal.h"
49 : #include "storage/proc.h"
50 : #include "storage/procarray.h"
51 : #include "storage/procsignal.h"
52 : #include "storage/spin.h"
53 : #include "storage/standby.h"
54 : #include "utils/timeout.h"
55 : #include "utils/timestamp.h"
56 :
57 : /* GUC variables */
58 : int DeadlockTimeout = 1000;
59 : int StatementTimeout = 0;
60 : int LockTimeout = 0;
61 : int IdleInTransactionSessionTimeout = 0;
62 : int TransactionTimeout = 0;
63 : int IdleSessionTimeout = 0;
64 : bool log_lock_waits = true;
65 :
66 : /* Pointer to this process's PGPROC struct, if any */
67 : PGPROC *MyProc = NULL;
68 :
69 : /*
70 : * This spinlock protects the freelist of recycled PGPROC structures.
71 : * We cannot use an LWLock because the LWLock manager depends on already
72 : * having a PGPROC and a wait semaphore! But these structures are touched
73 : * relatively infrequently (only at backend startup or shutdown) and not for
74 : * very long, so a spinlock is okay.
75 : */
76 : NON_EXEC_STATIC slock_t *ProcStructLock = NULL;
77 :
78 : /* Pointers to shared-memory structures */
79 : PROC_HDR *ProcGlobal = NULL;
80 : NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
81 : PGPROC *PreparedXactProcs = NULL;
82 :
83 : static DeadLockState deadlock_state = DS_NOT_YET_CHECKED;
84 :
85 : /* Is a deadlock check pending? */
86 : static volatile sig_atomic_t got_deadlock_timeout;
87 :
88 : static void RemoveProcFromArray(int code, Datum arg);
89 : static void ProcKill(int code, Datum arg);
90 : static void AuxiliaryProcKill(int code, Datum arg);
91 : static void CheckDeadLock(void);
92 :
93 :
94 : /*
95 : * Report shared-memory space needed by PGPROC.
96 : */
97 : static Size
98 6312 : PGProcShmemSize(void)
99 : {
100 6312 : Size size = 0;
101 : Size TotalProcs =
102 6312 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
103 :
104 6312 : size = add_size(size, mul_size(TotalProcs, sizeof(PGPROC)));
105 6312 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->xids)));
106 6312 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->subxidStates)));
107 6312 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->statusFlags)));
108 :
109 6312 : return size;
110 : }
111 :
112 : /*
113 : * Report shared-memory space needed by Fast-Path locks.
114 : */
115 : static Size
116 6312 : FastPathLockShmemSize(void)
117 : {
118 6312 : Size size = 0;
119 : Size TotalProcs =
120 6312 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
121 : Size fpLockBitsSize,
122 : fpRelIdSize;
123 :
124 : /*
125 : * Memory needed for PGPROC fast-path lock arrays. Make sure the sizes are
126 : * nicely aligned in each backend.
127 : */
128 6312 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
129 6312 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
130 :
131 6312 : size = add_size(size, mul_size(TotalProcs, (fpLockBitsSize + fpRelIdSize)));
132 :
133 6312 : return size;
134 : }
135 :
136 : /*
137 : * Report shared-memory space needed by InitProcGlobal.
138 : */
139 : Size
140 4108 : ProcGlobalShmemSize(void)
141 : {
142 4108 : Size size = 0;
143 :
144 : /* ProcGlobal */
145 4108 : size = add_size(size, sizeof(PROC_HDR));
146 4108 : size = add_size(size, sizeof(slock_t));
147 :
148 4108 : size = add_size(size, PGSemaphoreShmemSize(ProcGlobalSemas()));
149 4108 : size = add_size(size, PGProcShmemSize());
150 4108 : size = add_size(size, FastPathLockShmemSize());
151 :
152 4108 : return size;
153 : }
154 :
155 : /*
156 : * Report number of semaphores needed by InitProcGlobal.
157 : */
158 : int
159 8212 : ProcGlobalSemas(void)
160 : {
161 : /*
162 : * We need a sema per backend (including autovacuum), plus one for each
163 : * auxiliary process.
164 : */
165 8212 : return MaxBackends + NUM_AUXILIARY_PROCS;
166 : }
167 :
168 : /*
169 : * InitProcGlobal -
170 : * Initialize the global process table during postmaster or standalone
171 : * backend startup.
172 : *
173 : * We also create all the per-process semaphores we will need to support
174 : * the requested number of backends. We used to allocate semaphores
175 : * only when backends were actually started up, but that is bad because
176 : * it lets Postgres fail under load --- a lot of Unix systems are
177 : * (mis)configured with small limits on the number of semaphores, and
178 : * running out when trying to start another backend is a common failure.
179 : * So, now we grab enough semaphores to support the desired max number
180 : * of backends immediately at initialization --- if the sysadmin has set
181 : * MaxConnections, max_worker_processes, max_wal_senders, or
182 : * autovacuum_worker_slots higher than his kernel will support, he'll
183 : * find out sooner rather than later.
184 : *
185 : * Another reason for creating semaphores here is that the semaphore
186 : * implementation typically requires us to create semaphores in the
187 : * postmaster, not in backends.
188 : *
189 : * Note: this is NOT called by individual backends under a postmaster,
190 : * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
191 : * pointers must be propagated specially for EXEC_BACKEND operation.
192 : */
193 : void
194 2204 : InitProcGlobal(void)
195 : {
196 : PGPROC *procs;
197 : int i,
198 : j;
199 : bool found;
200 2204 : uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
201 :
202 : /* Used for setup of per-backend fast-path slots. */
203 : char *fpPtr,
204 : *fpEndPtr PG_USED_FOR_ASSERTS_ONLY;
205 : Size fpLockBitsSize,
206 : fpRelIdSize;
207 : Size requestSize;
208 : char *ptr;
209 :
210 : /* Create the ProcGlobal shared structure */
211 2204 : ProcGlobal = (PROC_HDR *)
212 2204 : ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
213 : Assert(!found);
214 :
215 : /*
216 : * Initialize the data structures.
217 : */
218 2204 : ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
219 2204 : dlist_init(&ProcGlobal->freeProcs);
220 2204 : dlist_init(&ProcGlobal->autovacFreeProcs);
221 2204 : dlist_init(&ProcGlobal->bgworkerFreeProcs);
222 2204 : dlist_init(&ProcGlobal->walsenderFreeProcs);
223 2204 : ProcGlobal->startupBufferPinWaitBufId = -1;
224 2204 : ProcGlobal->walwriterProc = INVALID_PROC_NUMBER;
225 2204 : ProcGlobal->checkpointerProc = INVALID_PROC_NUMBER;
226 2204 : pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PROC_NUMBER);
227 2204 : pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PROC_NUMBER);
228 :
229 : /*
230 : * Create and initialize all the PGPROC structures we'll need. There are
231 : * six separate consumers: (1) normal backends, (2) autovacuum workers and
232 : * special workers, (3) background workers, (4) walsenders, (5) auxiliary
233 : * processes, and (6) prepared transactions. (For largely-historical
234 : * reasons, we combine autovacuum and special workers into one category
235 : * with a single freelist.) Each PGPROC structure is dedicated to exactly
236 : * one of these purposes, and they do not move between groups.
237 : */
238 2204 : requestSize = PGProcShmemSize();
239 :
240 2204 : ptr = ShmemInitStruct("PGPROC structures",
241 : requestSize,
242 : &found);
243 :
244 2204 : MemSet(ptr, 0, requestSize);
245 :
246 2204 : procs = (PGPROC *) ptr;
247 2204 : ptr = (char *) ptr + TotalProcs * sizeof(PGPROC);
248 :
249 2204 : ProcGlobal->allProcs = procs;
250 : /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
251 2204 : ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
252 :
253 : /*
254 : * Allocate arrays mirroring PGPROC fields in a dense manner. See
255 : * PROC_HDR.
256 : *
257 : * XXX: It might make sense to increase padding for these arrays, given
258 : * how hotly they are accessed.
259 : */
260 2204 : ProcGlobal->xids = (TransactionId *) ptr;
261 2204 : ptr = (char *) ptr + (TotalProcs * sizeof(*ProcGlobal->xids));
262 :
263 2204 : ProcGlobal->subxidStates = (XidCacheStatus *) ptr;
264 2204 : ptr = (char *) ptr + (TotalProcs * sizeof(*ProcGlobal->subxidStates));
265 :
266 2204 : ProcGlobal->statusFlags = (uint8 *) ptr;
267 2204 : ptr = (char *) ptr + (TotalProcs * sizeof(*ProcGlobal->statusFlags));
268 :
269 : /* make sure wer didn't overflow */
270 : Assert((ptr > (char *) procs) && (ptr <= (char *) procs + requestSize));
271 :
272 : /*
273 : * Allocate arrays for fast-path locks. Those are variable-length, so
274 : * can't be included in PGPROC directly. We allocate a separate piece of
275 : * shared memory and then divide that between backends.
276 : */
277 2204 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
278 2204 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
279 :
280 2204 : requestSize = FastPathLockShmemSize();
281 :
282 2204 : fpPtr = ShmemInitStruct("Fast-Path Lock Array",
283 : requestSize,
284 : &found);
285 :
286 2204 : MemSet(fpPtr, 0, requestSize);
287 :
288 : /* For asserts checking we did not overflow. */
289 2204 : fpEndPtr = fpPtr + requestSize;
290 :
291 : /* Reserve space for semaphores. */
292 2204 : PGReserveSemaphores(ProcGlobalSemas());
293 :
294 291134 : for (i = 0; i < TotalProcs; i++)
295 : {
296 288930 : PGPROC *proc = &procs[i];
297 :
298 : /* Common initialization for all PGPROCs, regardless of type. */
299 :
300 : /*
301 : * Set the fast-path lock arrays, and move the pointer. We interleave
302 : * the two arrays, to (hopefully) get some locality for each backend.
303 : */
304 288930 : proc->fpLockBits = (uint64 *) fpPtr;
305 288930 : fpPtr += fpLockBitsSize;
306 :
307 288930 : proc->fpRelId = (Oid *) fpPtr;
308 288930 : fpPtr += fpRelIdSize;
309 :
310 : Assert(fpPtr <= fpEndPtr);
311 :
312 : /*
313 : * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
314 : * dummy PGPROCs don't need these though - they're never associated
315 : * with a real process
316 : */
317 288930 : if (i < MaxBackends + NUM_AUXILIARY_PROCS)
318 : {
319 287222 : proc->sem = PGSemaphoreCreate();
320 287222 : InitSharedLatch(&(proc->procLatch));
321 287222 : LWLockInitialize(&(proc->fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
322 : }
323 :
324 : /*
325 : * Newly created PGPROCs for normal backends, autovacuum workers,
326 : * special workers, bgworkers, and walsenders must be queued up on the
327 : * appropriate free list. Because there can only ever be a small,
328 : * fixed number of auxiliary processes, no free list is used in that
329 : * case; InitAuxiliaryProcess() instead uses a linear search. PGPROCs
330 : * for prepared transactions are added to a free list by
331 : * TwoPhaseShmemInit().
332 : */
333 288930 : if (i < MaxConnections)
334 : {
335 : /* PGPROC for normal backend, add to freeProcs list */
336 142882 : dlist_push_tail(&ProcGlobal->freeProcs, &proc->links);
337 142882 : proc->procgloballist = &ProcGlobal->freeProcs;
338 : }
339 146048 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS)
340 : {
341 : /* PGPROC for AV or special worker, add to autovacFreeProcs list */
342 28544 : dlist_push_tail(&ProcGlobal->autovacFreeProcs, &proc->links);
343 28544 : proc->procgloballist = &ProcGlobal->autovacFreeProcs;
344 : }
345 117504 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS + max_worker_processes)
346 : {
347 : /* PGPROC for bgworker, add to bgworkerFreeProcs list */
348 17628 : dlist_push_tail(&ProcGlobal->bgworkerFreeProcs, &proc->links);
349 17628 : proc->procgloballist = &ProcGlobal->bgworkerFreeProcs;
350 : }
351 99876 : else if (i < MaxBackends)
352 : {
353 : /* PGPROC for walsender, add to walsenderFreeProcs list */
354 14416 : dlist_push_tail(&ProcGlobal->walsenderFreeProcs, &proc->links);
355 14416 : proc->procgloballist = &ProcGlobal->walsenderFreeProcs;
356 : }
357 :
358 : /* Initialize myProcLocks[] shared memory queues. */
359 4911810 : for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
360 4622880 : dlist_init(&(proc->myProcLocks[j]));
361 :
362 : /* Initialize lockGroupMembers list. */
363 288930 : dlist_init(&proc->lockGroupMembers);
364 :
365 : /*
366 : * Initialize the atomic variables, otherwise, it won't be safe to
367 : * access them for backends that aren't currently in use.
368 : */
369 288930 : pg_atomic_init_u32(&(proc->procArrayGroupNext), INVALID_PROC_NUMBER);
370 288930 : pg_atomic_init_u32(&(proc->clogGroupNext), INVALID_PROC_NUMBER);
371 288930 : pg_atomic_init_u64(&(proc->waitStart), 0);
372 : }
373 :
374 : /* Should have consumed exactly the expected amount of fast-path memory. */
375 : Assert(fpPtr == fpEndPtr);
376 :
377 : /*
378 : * Save pointers to the blocks of PGPROC structures reserved for auxiliary
379 : * processes and prepared transactions.
380 : */
381 2204 : AuxiliaryProcs = &procs[MaxBackends];
382 2204 : PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
383 :
384 : /* Create ProcStructLock spinlock, too */
385 2204 : ProcStructLock = (slock_t *) ShmemInitStruct("ProcStructLock spinlock",
386 : sizeof(slock_t),
387 : &found);
388 2204 : SpinLockInit(ProcStructLock);
389 2204 : }
390 :
391 : /*
392 : * InitProcess -- initialize a per-process PGPROC entry for this backend
393 : */
394 : void
395 36416 : InitProcess(void)
396 : {
397 : dlist_head *procgloballist;
398 :
399 : /*
400 : * ProcGlobal should be set up already (if we are a backend, we inherit
401 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
402 : */
403 36416 : if (ProcGlobal == NULL)
404 0 : elog(PANIC, "proc header uninitialized");
405 :
406 36416 : if (MyProc != NULL)
407 0 : elog(ERROR, "you already exist");
408 :
409 : /*
410 : * Before we start accessing the shared memory in a serious way, mark
411 : * ourselves as an active postmaster child; this is so that the postmaster
412 : * can detect it if we exit without cleaning up.
413 : */
414 36416 : if (IsUnderPostmaster)
415 36178 : RegisterPostmasterChildActive();
416 :
417 : /*
418 : * Decide which list should supply our PGPROC. This logic must match the
419 : * way the freelists were constructed in InitProcGlobal().
420 : */
421 36416 : if (AmAutoVacuumWorkerProcess() || AmSpecialWorkerProcess())
422 4986 : procgloballist = &ProcGlobal->autovacFreeProcs;
423 31430 : else if (AmBackgroundWorkerProcess())
424 4772 : procgloballist = &ProcGlobal->bgworkerFreeProcs;
425 26658 : else if (AmWalSenderProcess())
426 2358 : procgloballist = &ProcGlobal->walsenderFreeProcs;
427 : else
428 24300 : procgloballist = &ProcGlobal->freeProcs;
429 :
430 : /*
431 : * Try to get a proc struct from the appropriate free list. If this
432 : * fails, we must be out of PGPROC structures (not to mention semaphores).
433 : *
434 : * While we are holding the ProcStructLock, also copy the current shared
435 : * estimate of spins_per_delay to local storage.
436 : */
437 36416 : SpinLockAcquire(ProcStructLock);
438 :
439 36416 : set_spins_per_delay(ProcGlobal->spins_per_delay);
440 :
441 36416 : if (!dlist_is_empty(procgloballist))
442 : {
443 36410 : MyProc = dlist_container(PGPROC, links, dlist_pop_head_node(procgloballist));
444 36410 : SpinLockRelease(ProcStructLock);
445 : }
446 : else
447 : {
448 : /*
449 : * If we reach here, all the PGPROCs are in use. This is one of the
450 : * possible places to detect "too many backends", so give the standard
451 : * error message. XXX do we need to give a different failure message
452 : * in the autovacuum case?
453 : */
454 6 : SpinLockRelease(ProcStructLock);
455 6 : if (AmWalSenderProcess())
456 4 : ereport(FATAL,
457 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
458 : errmsg("number of requested standby connections exceeds \"max_wal_senders\" (currently %d)",
459 : max_wal_senders)));
460 2 : ereport(FATAL,
461 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
462 : errmsg("sorry, too many clients already")));
463 : }
464 36410 : MyProcNumber = GetNumberFromPGProc(MyProc);
465 :
466 : /*
467 : * Cross-check that the PGPROC is of the type we expect; if this were not
468 : * the case, it would get returned to the wrong list.
469 : */
470 : Assert(MyProc->procgloballist == procgloballist);
471 :
472 : /*
473 : * Initialize all fields of MyProc, except for those previously
474 : * initialized by InitProcGlobal.
475 : */
476 36410 : dlist_node_init(&MyProc->links);
477 36410 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
478 36410 : MyProc->fpVXIDLock = false;
479 36410 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
480 36410 : MyProc->xid = InvalidTransactionId;
481 36410 : MyProc->xmin = InvalidTransactionId;
482 36410 : MyProc->pid = MyProcPid;
483 36410 : MyProc->vxid.procNumber = MyProcNumber;
484 36410 : MyProc->vxid.lxid = InvalidLocalTransactionId;
485 : /* databaseId and roleId will be filled in later */
486 36410 : MyProc->databaseId = InvalidOid;
487 36410 : MyProc->roleId = InvalidOid;
488 36410 : MyProc->tempNamespaceId = InvalidOid;
489 36410 : MyProc->isRegularBackend = AmRegularBackendProcess();
490 36410 : MyProc->delayChkptFlags = 0;
491 36410 : MyProc->statusFlags = 0;
492 : /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
493 36410 : if (AmAutoVacuumWorkerProcess())
494 4172 : MyProc->statusFlags |= PROC_IS_AUTOVACUUM;
495 36410 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
496 36410 : MyProc->lwWaitMode = 0;
497 36410 : MyProc->waitLock = NULL;
498 36410 : MyProc->waitProcLock = NULL;
499 36410 : pg_atomic_write_u64(&MyProc->waitStart, 0);
500 : #ifdef USE_ASSERT_CHECKING
501 : {
502 : int i;
503 :
504 : /* Last process should have released all locks. */
505 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
506 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
507 : }
508 : #endif
509 36410 : MyProc->recoveryConflictPending = false;
510 :
511 : /* Initialize fields for sync rep */
512 36410 : MyProc->waitLSN = 0;
513 36410 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
514 36410 : dlist_node_init(&MyProc->syncRepLinks);
515 :
516 : /* Initialize fields for group XID clearing. */
517 36410 : MyProc->procArrayGroupMember = false;
518 36410 : MyProc->procArrayGroupMemberXid = InvalidTransactionId;
519 : Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PROC_NUMBER);
520 :
521 : /* Check that group locking fields are in a proper initial state. */
522 : Assert(MyProc->lockGroupLeader == NULL);
523 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
524 :
525 : /* Initialize wait event information. */
526 36410 : MyProc->wait_event_info = 0;
527 :
528 : /* Initialize fields for group transaction status update. */
529 36410 : MyProc->clogGroupMember = false;
530 36410 : MyProc->clogGroupMemberXid = InvalidTransactionId;
531 36410 : MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
532 36410 : MyProc->clogGroupMemberPage = -1;
533 36410 : MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
534 : Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PROC_NUMBER);
535 :
536 : /*
537 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
538 : * on it. That allows us to repoint the process latch, which so far
539 : * points to process local one, to the shared one.
540 : */
541 36410 : OwnLatch(&MyProc->procLatch);
542 36410 : SwitchToSharedLatch();
543 :
544 : /* now that we have a proc, report wait events to shared memory */
545 36410 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
546 :
547 : /*
548 : * We might be reusing a semaphore that belonged to a failed process. So
549 : * be careful and reinitialize its value here. (This is not strictly
550 : * necessary anymore, but seems like a good idea for cleanliness.)
551 : */
552 36410 : PGSemaphoreReset(MyProc->sem);
553 :
554 : /*
555 : * Arrange to clean up at backend exit.
556 : */
557 36410 : on_shmem_exit(ProcKill, 0);
558 :
559 : /*
560 : * Now that we have a PGPROC, we could try to acquire locks, so initialize
561 : * local state needed for LWLocks, and the deadlock checker.
562 : */
563 36410 : InitLWLockAccess();
564 36410 : InitDeadLockChecking();
565 :
566 : #ifdef EXEC_BACKEND
567 :
568 : /*
569 : * Initialize backend-local pointers to all the shared data structures.
570 : * (We couldn't do this until now because it needs LWLocks.)
571 : */
572 : if (IsUnderPostmaster)
573 : AttachSharedMemoryStructs();
574 : #endif
575 36410 : }
576 :
577 : /*
578 : * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
579 : *
580 : * This is separate from InitProcess because we can't acquire LWLocks until
581 : * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
582 : * work until after we've done AttachSharedMemoryStructs.
583 : */
584 : void
585 36392 : InitProcessPhase2(void)
586 : {
587 : Assert(MyProc != NULL);
588 :
589 : /*
590 : * Add our PGPROC to the PGPROC array in shared memory.
591 : */
592 36392 : ProcArrayAdd(MyProc);
593 :
594 : /*
595 : * Arrange to clean that up at backend exit.
596 : */
597 36392 : on_shmem_exit(RemoveProcFromArray, 0);
598 36392 : }
599 :
600 : /*
601 : * InitAuxiliaryProcess -- create a PGPROC entry for an auxiliary process
602 : *
603 : * This is called by bgwriter and similar processes so that they will have a
604 : * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
605 : * and sema that are assigned are one of the extra ones created during
606 : * InitProcGlobal.
607 : *
608 : * Auxiliary processes are presently not expected to wait for real (lockmgr)
609 : * locks, so we need not set up the deadlock checker. They are never added
610 : * to the ProcArray or the sinval messaging mechanism, either. They also
611 : * don't get a VXID assigned, since this is only useful when we actually
612 : * hold lockmgr locks.
613 : *
614 : * Startup process however uses locks but never waits for them in the
615 : * normal backend sense. Startup process also takes part in sinval messaging
616 : * as a sendOnly process, so never reads messages from sinval queue. So
617 : * Startup process does have a VXID and does show up in pg_locks.
618 : */
619 : void
620 8692 : InitAuxiliaryProcess(void)
621 : {
622 : PGPROC *auxproc;
623 : int proctype;
624 :
625 : /*
626 : * ProcGlobal should be set up already (if we are a backend, we inherit
627 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
628 : */
629 8692 : if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
630 0 : elog(PANIC, "proc header uninitialized");
631 :
632 8692 : if (MyProc != NULL)
633 0 : elog(ERROR, "you already exist");
634 :
635 8692 : if (IsUnderPostmaster)
636 8692 : RegisterPostmasterChildActive();
637 :
638 : /*
639 : * We use the ProcStructLock to protect assignment and releasing of
640 : * AuxiliaryProcs entries.
641 : *
642 : * While we are holding the ProcStructLock, also copy the current shared
643 : * estimate of spins_per_delay to local storage.
644 : */
645 8692 : SpinLockAcquire(ProcStructLock);
646 :
647 8692 : set_spins_per_delay(ProcGlobal->spins_per_delay);
648 :
649 : /*
650 : * Find a free auxproc ... *big* trouble if there isn't one ...
651 : */
652 34128 : for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
653 : {
654 34128 : auxproc = &AuxiliaryProcs[proctype];
655 34128 : if (auxproc->pid == 0)
656 8692 : break;
657 : }
658 8692 : if (proctype >= NUM_AUXILIARY_PROCS)
659 : {
660 0 : SpinLockRelease(ProcStructLock);
661 0 : elog(FATAL, "all AuxiliaryProcs are in use");
662 : }
663 :
664 : /* Mark auxiliary proc as in use by me */
665 : /* use volatile pointer to prevent code rearrangement */
666 8692 : ((volatile PGPROC *) auxproc)->pid = MyProcPid;
667 :
668 8692 : SpinLockRelease(ProcStructLock);
669 :
670 8692 : MyProc = auxproc;
671 8692 : MyProcNumber = GetNumberFromPGProc(MyProc);
672 :
673 : /*
674 : * Initialize all fields of MyProc, except for those previously
675 : * initialized by InitProcGlobal.
676 : */
677 8692 : dlist_node_init(&MyProc->links);
678 8692 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
679 8692 : MyProc->fpVXIDLock = false;
680 8692 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
681 8692 : MyProc->xid = InvalidTransactionId;
682 8692 : MyProc->xmin = InvalidTransactionId;
683 8692 : MyProc->vxid.procNumber = INVALID_PROC_NUMBER;
684 8692 : MyProc->vxid.lxid = InvalidLocalTransactionId;
685 8692 : MyProc->databaseId = InvalidOid;
686 8692 : MyProc->roleId = InvalidOid;
687 8692 : MyProc->tempNamespaceId = InvalidOid;
688 8692 : MyProc->isRegularBackend = false;
689 8692 : MyProc->delayChkptFlags = 0;
690 8692 : MyProc->statusFlags = 0;
691 8692 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
692 8692 : MyProc->lwWaitMode = 0;
693 8692 : MyProc->waitLock = NULL;
694 8692 : MyProc->waitProcLock = NULL;
695 8692 : pg_atomic_write_u64(&MyProc->waitStart, 0);
696 : #ifdef USE_ASSERT_CHECKING
697 : {
698 : int i;
699 :
700 : /* Last process should have released all locks. */
701 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
702 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
703 : }
704 : #endif
705 :
706 : /*
707 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
708 : * on it. That allows us to repoint the process latch, which so far
709 : * points to process local one, to the shared one.
710 : */
711 8692 : OwnLatch(&MyProc->procLatch);
712 8692 : SwitchToSharedLatch();
713 :
714 : /* now that we have a proc, report wait events to shared memory */
715 8692 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
716 :
717 : /* Check that group locking fields are in a proper initial state. */
718 : Assert(MyProc->lockGroupLeader == NULL);
719 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
720 :
721 : /*
722 : * We might be reusing a semaphore that belonged to a failed process. So
723 : * be careful and reinitialize its value here. (This is not strictly
724 : * necessary anymore, but seems like a good idea for cleanliness.)
725 : */
726 8692 : PGSemaphoreReset(MyProc->sem);
727 :
728 : /*
729 : * Arrange to clean up at process exit.
730 : */
731 8692 : on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
732 :
733 : /*
734 : * Now that we have a PGPROC, we could try to acquire lightweight locks.
735 : * Initialize local state needed for them. (Heavyweight locks cannot be
736 : * acquired in aux processes.)
737 : */
738 8692 : InitLWLockAccess();
739 :
740 : #ifdef EXEC_BACKEND
741 :
742 : /*
743 : * Initialize backend-local pointers to all the shared data structures.
744 : * (We couldn't do this until now because it needs LWLocks.)
745 : */
746 : if (IsUnderPostmaster)
747 : AttachSharedMemoryStructs();
748 : #endif
749 8692 : }
750 :
751 : /*
752 : * Used from bufmgr to share the value of the buffer that Startup waits on,
753 : * or to reset the value to "not waiting" (-1). This allows processing
754 : * of recovery conflicts for buffer pins. Set is made before backends look
755 : * at this value, so locking not required, especially since the set is
756 : * an atomic integer set operation.
757 : */
758 : void
759 40 : SetStartupBufferPinWaitBufId(int bufid)
760 : {
761 : /* use volatile pointer to prevent code rearrangement */
762 40 : volatile PROC_HDR *procglobal = ProcGlobal;
763 :
764 40 : procglobal->startupBufferPinWaitBufId = bufid;
765 40 : }
766 :
767 : /*
768 : * Used by backends when they receive a request to check for buffer pin waits.
769 : */
770 : int
771 6 : GetStartupBufferPinWaitBufId(void)
772 : {
773 : /* use volatile pointer to prevent code rearrangement */
774 6 : volatile PROC_HDR *procglobal = ProcGlobal;
775 :
776 6 : return procglobal->startupBufferPinWaitBufId;
777 : }
778 :
779 : /*
780 : * Check whether there are at least N free PGPROC objects. If false is
781 : * returned, *nfree will be set to the number of free PGPROC objects.
782 : * Otherwise, *nfree will be set to n.
783 : *
784 : * Note: this is designed on the assumption that N will generally be small.
785 : */
786 : bool
787 488 : HaveNFreeProcs(int n, int *nfree)
788 : {
789 : dlist_iter iter;
790 :
791 : Assert(n > 0);
792 : Assert(nfree);
793 :
794 488 : SpinLockAcquire(ProcStructLock);
795 :
796 488 : *nfree = 0;
797 1458 : dlist_foreach(iter, &ProcGlobal->freeProcs)
798 : {
799 1450 : (*nfree)++;
800 1450 : if (*nfree == n)
801 480 : break;
802 : }
803 :
804 488 : SpinLockRelease(ProcStructLock);
805 :
806 488 : return (*nfree == n);
807 : }
808 :
809 : /*
810 : * Cancel any pending wait for lock, when aborting a transaction, and revert
811 : * any strong lock count acquisition for a lock being acquired.
812 : *
813 : * (Normally, this would only happen if we accept a cancel/die
814 : * interrupt while waiting; but an ereport(ERROR) before or during the lock
815 : * wait is within the realm of possibility, too.)
816 : */
817 : void
818 1167382 : LockErrorCleanup(void)
819 : {
820 : LOCALLOCK *lockAwaited;
821 : LWLock *partitionLock;
822 : DisableTimeoutParams timeouts[2];
823 :
824 1167382 : HOLD_INTERRUPTS();
825 :
826 1167382 : AbortStrongLockAcquire();
827 :
828 : /* Nothing to do if we weren't waiting for a lock */
829 1167382 : lockAwaited = GetAwaitedLock();
830 1167382 : if (lockAwaited == NULL)
831 : {
832 1167300 : RESUME_INTERRUPTS();
833 1167300 : return;
834 : }
835 :
836 : /*
837 : * Turn off the deadlock and lock timeout timers, if they are still
838 : * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
839 : * indicator flag, since this function is executed before
840 : * ProcessInterrupts when responding to SIGINT; else we'd lose the
841 : * knowledge that the SIGINT came from a lock timeout and not an external
842 : * source.
843 : */
844 82 : timeouts[0].id = DEADLOCK_TIMEOUT;
845 82 : timeouts[0].keep_indicator = false;
846 82 : timeouts[1].id = LOCK_TIMEOUT;
847 82 : timeouts[1].keep_indicator = true;
848 82 : disable_timeouts(timeouts, 2);
849 :
850 : /* Unlink myself from the wait queue, if on it (might not be anymore!) */
851 82 : partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
852 82 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
853 :
854 82 : if (!dlist_node_is_detached(&MyProc->links))
855 : {
856 : /* We could not have been granted the lock yet */
857 80 : RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
858 : }
859 : else
860 : {
861 : /*
862 : * Somebody kicked us off the lock queue already. Perhaps they
863 : * granted us the lock, or perhaps they detected a deadlock. If they
864 : * did grant us the lock, we'd better remember it in our local lock
865 : * table.
866 : */
867 2 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
868 2 : GrantAwaitedLock();
869 : }
870 :
871 82 : ResetAwaitedLock();
872 :
873 82 : LWLockRelease(partitionLock);
874 :
875 82 : RESUME_INTERRUPTS();
876 : }
877 :
878 :
879 : /*
880 : * ProcReleaseLocks() -- release locks associated with current transaction
881 : * at main transaction commit or abort
882 : *
883 : * At main transaction commit, we release standard locks except session locks.
884 : * At main transaction abort, we release all locks including session locks.
885 : *
886 : * Advisory locks are released only if they are transaction-level;
887 : * session-level holds remain, whether this is a commit or not.
888 : *
889 : * At subtransaction commit, we don't release any locks (so this func is not
890 : * needed at all); we will defer the releasing to the parent transaction.
891 : * At subtransaction abort, we release all locks held by the subtransaction;
892 : * this is implemented by retail releasing of the locks under control of
893 : * the ResourceOwner mechanism.
894 : */
895 : void
896 1105678 : ProcReleaseLocks(bool isCommit)
897 : {
898 1105678 : if (!MyProc)
899 0 : return;
900 : /* If waiting, get off wait queue (should only be needed after error) */
901 1105678 : LockErrorCleanup();
902 : /* Release standard locks, including session-level if aborting */
903 1105678 : LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
904 : /* Release transaction-level advisory locks */
905 1105678 : LockReleaseAll(USER_LOCKMETHOD, false);
906 : }
907 :
908 :
909 : /*
910 : * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
911 : */
912 : static void
913 36392 : RemoveProcFromArray(int code, Datum arg)
914 : {
915 : Assert(MyProc != NULL);
916 36392 : ProcArrayRemove(MyProc, InvalidTransactionId);
917 36392 : }
918 :
919 : /*
920 : * ProcKill() -- Destroy the per-proc data structure for
921 : * this process. Release any of its held LW locks.
922 : */
923 : static void
924 36410 : ProcKill(int code, Datum arg)
925 : {
926 : PGPROC *proc;
927 : dlist_head *procgloballist;
928 :
929 : Assert(MyProc != NULL);
930 :
931 : /* not safe if forked by system(), etc. */
932 36410 : if (MyProc->pid != (int) getpid())
933 0 : elog(PANIC, "ProcKill() called in child process");
934 :
935 : /* Make sure we're out of the sync rep lists */
936 36410 : SyncRepCleanupAtProcExit();
937 :
938 : #ifdef USE_ASSERT_CHECKING
939 : {
940 : int i;
941 :
942 : /* Last process should have released all locks. */
943 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
944 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
945 : }
946 : #endif
947 :
948 : /*
949 : * Release any LW locks I am holding. There really shouldn't be any, but
950 : * it's cheap to check again before we cut the knees off the LWLock
951 : * facility by releasing our PGPROC ...
952 : */
953 36410 : LWLockReleaseAll();
954 :
955 : /*
956 : * Cleanup waiting for LSN if any.
957 : */
958 36410 : WaitLSNCleanup();
959 :
960 : /* Cancel any pending condition variable sleep, too */
961 36410 : ConditionVariableCancelSleep();
962 :
963 : /*
964 : * Detach from any lock group of which we are a member. If the leader
965 : * exits before all other group members, its PGPROC will remain allocated
966 : * until the last group process exits; that process must return the
967 : * leader's PGPROC to the appropriate list.
968 : */
969 36410 : if (MyProc->lockGroupLeader != NULL)
970 : {
971 2892 : PGPROC *leader = MyProc->lockGroupLeader;
972 2892 : LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
973 :
974 2892 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
975 : Assert(!dlist_is_empty(&leader->lockGroupMembers));
976 2892 : dlist_delete(&MyProc->lockGroupLink);
977 2892 : if (dlist_is_empty(&leader->lockGroupMembers))
978 : {
979 144 : leader->lockGroupLeader = NULL;
980 144 : if (leader != MyProc)
981 : {
982 0 : procgloballist = leader->procgloballist;
983 :
984 : /* Leader exited first; return its PGPROC. */
985 0 : SpinLockAcquire(ProcStructLock);
986 0 : dlist_push_head(procgloballist, &leader->links);
987 0 : SpinLockRelease(ProcStructLock);
988 : }
989 : }
990 2748 : else if (leader != MyProc)
991 2748 : MyProc->lockGroupLeader = NULL;
992 2892 : LWLockRelease(leader_lwlock);
993 : }
994 :
995 : /*
996 : * Reset MyLatch to the process local one. This is so that signal
997 : * handlers et al can continue using the latch after the shared latch
998 : * isn't ours anymore.
999 : *
1000 : * Similarly, stop reporting wait events to MyProc->wait_event_info.
1001 : *
1002 : * After that clear MyProc and disown the shared latch.
1003 : */
1004 36410 : SwitchBackToLocalLatch();
1005 36410 : pgstat_reset_wait_event_storage();
1006 :
1007 36410 : proc = MyProc;
1008 36410 : MyProc = NULL;
1009 36410 : MyProcNumber = INVALID_PROC_NUMBER;
1010 36410 : DisownLatch(&proc->procLatch);
1011 :
1012 : /* Mark the proc no longer in use */
1013 36410 : proc->pid = 0;
1014 36410 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1015 36410 : proc->vxid.lxid = InvalidTransactionId;
1016 :
1017 36410 : procgloballist = proc->procgloballist;
1018 36410 : SpinLockAcquire(ProcStructLock);
1019 :
1020 : /*
1021 : * If we're still a member of a locking group, that means we're a leader
1022 : * which has somehow exited before its children. The last remaining child
1023 : * will release our PGPROC. Otherwise, release it now.
1024 : */
1025 36410 : if (proc->lockGroupLeader == NULL)
1026 : {
1027 : /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
1028 : Assert(dlist_is_empty(&proc->lockGroupMembers));
1029 :
1030 : /* Return PGPROC structure (and semaphore) to appropriate freelist */
1031 36410 : dlist_push_tail(procgloballist, &proc->links);
1032 : }
1033 :
1034 : /* Update shared estimate of spins_per_delay */
1035 36410 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1036 :
1037 36410 : SpinLockRelease(ProcStructLock);
1038 :
1039 : /* wake autovac launcher if needed -- see comments in FreeWorkerInfo */
1040 36410 : if (AutovacuumLauncherPid != 0)
1041 4170 : kill(AutovacuumLauncherPid, SIGUSR2);
1042 36410 : }
1043 :
1044 : /*
1045 : * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
1046 : * processes (bgwriter, etc). The PGPROC and sema are not released, only
1047 : * marked as not-in-use.
1048 : */
1049 : static void
1050 8692 : AuxiliaryProcKill(int code, Datum arg)
1051 : {
1052 8692 : int proctype = DatumGetInt32(arg);
1053 : PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
1054 : PGPROC *proc;
1055 :
1056 : Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
1057 :
1058 : /* not safe if forked by system(), etc. */
1059 8692 : if (MyProc->pid != (int) getpid())
1060 0 : elog(PANIC, "AuxiliaryProcKill() called in child process");
1061 :
1062 8692 : auxproc = &AuxiliaryProcs[proctype];
1063 :
1064 : Assert(MyProc == auxproc);
1065 :
1066 : /* Release any LW locks I am holding (see notes above) */
1067 8692 : LWLockReleaseAll();
1068 :
1069 : /* Cancel any pending condition variable sleep, too */
1070 8692 : ConditionVariableCancelSleep();
1071 :
1072 : /* look at the equivalent ProcKill() code for comments */
1073 8692 : SwitchBackToLocalLatch();
1074 8692 : pgstat_reset_wait_event_storage();
1075 :
1076 8692 : proc = MyProc;
1077 8692 : MyProc = NULL;
1078 8692 : MyProcNumber = INVALID_PROC_NUMBER;
1079 8692 : DisownLatch(&proc->procLatch);
1080 :
1081 8692 : SpinLockAcquire(ProcStructLock);
1082 :
1083 : /* Mark auxiliary proc no longer in use */
1084 8692 : proc->pid = 0;
1085 8692 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1086 8692 : proc->vxid.lxid = InvalidTransactionId;
1087 :
1088 : /* Update shared estimate of spins_per_delay */
1089 8692 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1090 :
1091 8692 : SpinLockRelease(ProcStructLock);
1092 8692 : }
1093 :
1094 : /*
1095 : * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
1096 : * given its PID
1097 : *
1098 : * Returns NULL if not found.
1099 : */
1100 : PGPROC *
1101 12554 : AuxiliaryPidGetProc(int pid)
1102 : {
1103 12554 : PGPROC *result = NULL;
1104 : int index;
1105 :
1106 12554 : if (pid == 0) /* never match dummy PGPROCs */
1107 6 : return NULL;
1108 :
1109 60740 : for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
1110 : {
1111 60740 : PGPROC *proc = &AuxiliaryProcs[index];
1112 :
1113 60740 : if (proc->pid == pid)
1114 : {
1115 12548 : result = proc;
1116 12548 : break;
1117 : }
1118 : }
1119 12548 : return result;
1120 : }
1121 :
1122 :
1123 : /*
1124 : * JoinWaitQueue -- join the wait queue on the specified lock
1125 : *
1126 : * It's not actually guaranteed that we need to wait when this function is
1127 : * called, because it could be that when we try to find a position at which
1128 : * to insert ourself into the wait queue, we discover that we must be inserted
1129 : * ahead of everyone who wants a lock that conflict with ours. In that case,
1130 : * we get the lock immediately. Because of this, it's sensible for this function
1131 : * to have a dontWait argument, despite the name.
1132 : *
1133 : * On entry, the caller has already set up LOCK and PROCLOCK entries to
1134 : * reflect that we have "requested" the lock. The caller is responsible for
1135 : * cleaning that up, if we end up not joining the queue after all.
1136 : *
1137 : * The lock table's partition lock must be held at entry, and is still held
1138 : * at exit. The caller must release it before calling ProcSleep().
1139 : *
1140 : * Result is one of the following:
1141 : *
1142 : * PROC_WAIT_STATUS_OK - lock was immediately granted
1143 : * PROC_WAIT_STATUS_WAITING - joined the wait queue; call ProcSleep()
1144 : * PROC_WAIT_STATUS_ERROR - immediate deadlock was detected, or would
1145 : * need to wait and dontWait == true
1146 : *
1147 : * NOTES: The process queue is now a priority queue for locking.
1148 : */
1149 : ProcWaitStatus
1150 4124 : JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
1151 : {
1152 4124 : LOCKMODE lockmode = locallock->tag.mode;
1153 4124 : LOCK *lock = locallock->lock;
1154 4124 : PROCLOCK *proclock = locallock->proclock;
1155 4124 : uint32 hashcode = locallock->hashcode;
1156 4124 : LWLock *partitionLock PG_USED_FOR_ASSERTS_ONLY = LockHashPartitionLock(hashcode);
1157 4124 : dclist_head *waitQueue = &lock->waitProcs;
1158 4124 : PGPROC *insert_before = NULL;
1159 : LOCKMASK myProcHeldLocks;
1160 : LOCKMASK myHeldLocks;
1161 4124 : bool early_deadlock = false;
1162 4124 : PGPROC *leader = MyProc->lockGroupLeader;
1163 :
1164 : Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
1165 :
1166 : /*
1167 : * Set bitmask of locks this process already holds on this object.
1168 : */
1169 4124 : myHeldLocks = MyProc->heldLocks = proclock->holdMask;
1170 :
1171 : /*
1172 : * Determine which locks we're already holding.
1173 : *
1174 : * If group locking is in use, locks held by members of my locking group
1175 : * need to be included in myHeldLocks. This is not required for relation
1176 : * extension lock which conflict among group members. However, including
1177 : * them in myHeldLocks will give group members the priority to get those
1178 : * locks as compared to other backends which are also trying to acquire
1179 : * those locks. OTOH, we can avoid giving priority to group members for
1180 : * that kind of locks, but there doesn't appear to be a clear advantage of
1181 : * the same.
1182 : */
1183 4124 : myProcHeldLocks = proclock->holdMask;
1184 4124 : myHeldLocks = myProcHeldLocks;
1185 4124 : if (leader != NULL)
1186 : {
1187 : dlist_iter iter;
1188 :
1189 80 : dlist_foreach(iter, &lock->procLocks)
1190 : {
1191 : PROCLOCK *otherproclock;
1192 :
1193 60 : otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
1194 :
1195 60 : if (otherproclock->groupLeader == leader)
1196 28 : myHeldLocks |= otherproclock->holdMask;
1197 : }
1198 : }
1199 :
1200 : /*
1201 : * Determine where to add myself in the wait queue.
1202 : *
1203 : * Normally I should go at the end of the queue. However, if I already
1204 : * hold locks that conflict with the request of any previous waiter, put
1205 : * myself in the queue just in front of the first such waiter. This is not
1206 : * a necessary step, since deadlock detection would move me to before that
1207 : * waiter anyway; but it's relatively cheap to detect such a conflict
1208 : * immediately, and avoid delaying till deadlock timeout.
1209 : *
1210 : * Special case: if I find I should go in front of some waiter, check to
1211 : * see if I conflict with already-held locks or the requests before that
1212 : * waiter. If not, then just grant myself the requested lock immediately.
1213 : * This is the same as the test for immediate grant in LockAcquire, except
1214 : * we are only considering the part of the wait queue before my insertion
1215 : * point.
1216 : */
1217 4124 : if (myHeldLocks != 0 && !dclist_is_empty(waitQueue))
1218 : {
1219 12 : LOCKMASK aheadRequests = 0;
1220 : dlist_iter iter;
1221 :
1222 12 : dclist_foreach(iter, waitQueue)
1223 : {
1224 12 : PGPROC *proc = dlist_container(PGPROC, links, iter.cur);
1225 :
1226 : /*
1227 : * If we're part of the same locking group as this waiter, its
1228 : * locks neither conflict with ours nor contribute to
1229 : * aheadRequests.
1230 : */
1231 12 : if (leader != NULL && leader == proc->lockGroupLeader)
1232 0 : continue;
1233 :
1234 : /* Must he wait for me? */
1235 12 : if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1236 : {
1237 : /* Must I wait for him ? */
1238 12 : if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1239 : {
1240 : /*
1241 : * Yes, so we have a deadlock. Easiest way to clean up
1242 : * correctly is to call RemoveFromWaitQueue(), but we
1243 : * can't do that until we are *on* the wait queue. So, set
1244 : * a flag to check below, and break out of loop. Also,
1245 : * record deadlock info for later message.
1246 : */
1247 2 : RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1248 2 : early_deadlock = true;
1249 2 : break;
1250 : }
1251 : /* I must go before this waiter. Check special case. */
1252 10 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1253 10 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1254 : proclock))
1255 : {
1256 : /* Skip the wait and just grant myself the lock. */
1257 10 : GrantLock(lock, proclock, lockmode);
1258 10 : return PROC_WAIT_STATUS_OK;
1259 : }
1260 :
1261 : /* Put myself into wait queue before conflicting process */
1262 0 : insert_before = proc;
1263 0 : break;
1264 : }
1265 : /* Nope, so advance to next waiter */
1266 0 : aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1267 : }
1268 : }
1269 :
1270 : /*
1271 : * If we detected deadlock, give up without waiting. This must agree with
1272 : * CheckDeadLock's recovery code.
1273 : */
1274 4114 : if (early_deadlock)
1275 2 : return PROC_WAIT_STATUS_ERROR;
1276 :
1277 : /*
1278 : * At this point we know that we'd really need to sleep. If we've been
1279 : * commanded not to do that, bail out.
1280 : */
1281 4112 : if (dontWait)
1282 1496 : return PROC_WAIT_STATUS_ERROR;
1283 :
1284 : /*
1285 : * Insert self into queue, at the position determined above.
1286 : */
1287 2616 : if (insert_before)
1288 0 : dclist_insert_before(waitQueue, &insert_before->links, &MyProc->links);
1289 : else
1290 2616 : dclist_push_tail(waitQueue, &MyProc->links);
1291 :
1292 2616 : lock->waitMask |= LOCKBIT_ON(lockmode);
1293 :
1294 : /* Set up wait information in PGPROC object, too */
1295 2616 : MyProc->heldLocks = myProcHeldLocks;
1296 2616 : MyProc->waitLock = lock;
1297 2616 : MyProc->waitProcLock = proclock;
1298 2616 : MyProc->waitLockMode = lockmode;
1299 :
1300 2616 : MyProc->waitStatus = PROC_WAIT_STATUS_WAITING;
1301 :
1302 2616 : return PROC_WAIT_STATUS_WAITING;
1303 : }
1304 :
1305 : /*
1306 : * ProcSleep -- put process to sleep waiting on lock
1307 : *
1308 : * This must be called when JoinWaitQueue() returns PROC_WAIT_STATUS_WAITING.
1309 : * Returns after the lock has been granted, or if a deadlock is detected. Can
1310 : * also bail out with ereport(ERROR), if some other error condition, or a
1311 : * timeout or cancellation is triggered.
1312 : *
1313 : * Result is one of the following:
1314 : *
1315 : * PROC_WAIT_STATUS_OK - lock was granted
1316 : * PROC_WAIT_STATUS_ERROR - a deadlock was detected
1317 : */
1318 : ProcWaitStatus
1319 2616 : ProcSleep(LOCALLOCK *locallock)
1320 : {
1321 2616 : LOCKMODE lockmode = locallock->tag.mode;
1322 2616 : LOCK *lock = locallock->lock;
1323 2616 : uint32 hashcode = locallock->hashcode;
1324 2616 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1325 2616 : TimestampTz standbyWaitStart = 0;
1326 2616 : bool allow_autovacuum_cancel = true;
1327 2616 : bool logged_recovery_conflict = false;
1328 : ProcWaitStatus myWaitStatus;
1329 :
1330 : /* The caller must've armed the on-error cleanup mechanism */
1331 : Assert(GetAwaitedLock() == locallock);
1332 : Assert(!LWLockHeldByMe(partitionLock));
1333 :
1334 : /*
1335 : * Now that we will successfully clean up after an ereport, it's safe to
1336 : * check to see if there's a buffer pin deadlock against the Startup
1337 : * process. Of course, that's only necessary if we're doing Hot Standby
1338 : * and are not the Startup process ourselves.
1339 : */
1340 2616 : if (RecoveryInProgress() && !InRecovery)
1341 2 : CheckRecoveryConflictDeadlock();
1342 :
1343 : /* Reset deadlock_state before enabling the timeout handler */
1344 2616 : deadlock_state = DS_NOT_YET_CHECKED;
1345 2616 : got_deadlock_timeout = false;
1346 :
1347 : /*
1348 : * Set timer so we can wake up after awhile and check for a deadlock. If a
1349 : * deadlock is detected, the handler sets MyProc->waitStatus =
1350 : * PROC_WAIT_STATUS_ERROR, allowing us to know that we must report failure
1351 : * rather than success.
1352 : *
1353 : * By delaying the check until we've waited for a bit, we can avoid
1354 : * running the rather expensive deadlock-check code in most cases.
1355 : *
1356 : * If LockTimeout is set, also enable the timeout for that. We can save a
1357 : * few cycles by enabling both timeout sources in one call.
1358 : *
1359 : * If InHotStandby we set lock waits slightly later for clarity with other
1360 : * code.
1361 : */
1362 2616 : if (!InHotStandby)
1363 : {
1364 2614 : if (LockTimeout > 0)
1365 : {
1366 : EnableTimeoutParams timeouts[2];
1367 :
1368 158 : timeouts[0].id = DEADLOCK_TIMEOUT;
1369 158 : timeouts[0].type = TMPARAM_AFTER;
1370 158 : timeouts[0].delay_ms = DeadlockTimeout;
1371 158 : timeouts[1].id = LOCK_TIMEOUT;
1372 158 : timeouts[1].type = TMPARAM_AFTER;
1373 158 : timeouts[1].delay_ms = LockTimeout;
1374 158 : enable_timeouts(timeouts, 2);
1375 : }
1376 : else
1377 2456 : enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1378 :
1379 : /*
1380 : * Use the current time obtained for the deadlock timeout timer as
1381 : * waitStart (i.e., the time when this process started waiting for the
1382 : * lock). Since getting the current time newly can cause overhead, we
1383 : * reuse the already-obtained time to avoid that overhead.
1384 : *
1385 : * Note that waitStart is updated without holding the lock table's
1386 : * partition lock, to avoid the overhead by additional lock
1387 : * acquisition. This can cause "waitstart" in pg_locks to become NULL
1388 : * for a very short period of time after the wait started even though
1389 : * "granted" is false. This is OK in practice because we can assume
1390 : * that users are likely to look at "waitstart" when waiting for the
1391 : * lock for a long time.
1392 : */
1393 2614 : pg_atomic_write_u64(&MyProc->waitStart,
1394 2614 : get_timeout_start_time(DEADLOCK_TIMEOUT));
1395 : }
1396 2 : else if (log_recovery_conflict_waits)
1397 : {
1398 : /*
1399 : * Set the wait start timestamp if logging is enabled and in hot
1400 : * standby.
1401 : */
1402 2 : standbyWaitStart = GetCurrentTimestamp();
1403 : }
1404 :
1405 : /*
1406 : * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1407 : * will not wait. But a set latch does not necessarily mean that the lock
1408 : * is free now, as there are many other sources for latch sets than
1409 : * somebody releasing the lock.
1410 : *
1411 : * We process interrupts whenever the latch has been set, so cancel/die
1412 : * interrupts are processed quickly. This means we must not mind losing
1413 : * control to a cancel/die interrupt here. We don't, because we have no
1414 : * shared-state-change work to do after being granted the lock (the
1415 : * grantor did it all). We do have to worry about canceling the deadlock
1416 : * timeout and updating the locallock table, but if we lose control to an
1417 : * error, LockErrorCleanup will fix that up.
1418 : */
1419 : do
1420 : {
1421 3930 : if (InHotStandby)
1422 : {
1423 6 : bool maybe_log_conflict =
1424 6 : (standbyWaitStart != 0 && !logged_recovery_conflict);
1425 :
1426 : /* Set a timer and wait for that or for the lock to be granted */
1427 6 : ResolveRecoveryConflictWithLock(locallock->tag.lock,
1428 : maybe_log_conflict);
1429 :
1430 : /*
1431 : * Emit the log message if the startup process is waiting longer
1432 : * than deadlock_timeout for recovery conflict on lock.
1433 : */
1434 6 : if (maybe_log_conflict)
1435 : {
1436 2 : TimestampTz now = GetCurrentTimestamp();
1437 :
1438 2 : if (TimestampDifferenceExceeds(standbyWaitStart, now,
1439 : DeadlockTimeout))
1440 : {
1441 : VirtualTransactionId *vxids;
1442 : int cnt;
1443 :
1444 2 : vxids = GetLockConflicts(&locallock->tag.lock,
1445 : AccessExclusiveLock, &cnt);
1446 :
1447 : /*
1448 : * Log the recovery conflict and the list of PIDs of
1449 : * backends holding the conflicting lock. Note that we do
1450 : * logging even if there are no such backends right now
1451 : * because the startup process here has already waited
1452 : * longer than deadlock_timeout.
1453 : */
1454 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1455 : standbyWaitStart, now,
1456 2 : cnt > 0 ? vxids : NULL, true);
1457 2 : logged_recovery_conflict = true;
1458 : }
1459 : }
1460 : }
1461 : else
1462 : {
1463 3924 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1464 3924 : PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1465 3924 : ResetLatch(MyLatch);
1466 : /* check for deadlocks first, as that's probably log-worthy */
1467 3924 : if (got_deadlock_timeout)
1468 : {
1469 62 : CheckDeadLock();
1470 62 : got_deadlock_timeout = false;
1471 : }
1472 3924 : CHECK_FOR_INTERRUPTS();
1473 : }
1474 :
1475 : /*
1476 : * waitStatus could change from PROC_WAIT_STATUS_WAITING to something
1477 : * else asynchronously. Read it just once per loop to prevent
1478 : * surprising behavior (such as missing log messages).
1479 : */
1480 3848 : myWaitStatus = *((volatile ProcWaitStatus *) &MyProc->waitStatus);
1481 :
1482 : /*
1483 : * If we are not deadlocked, but are waiting on an autovacuum-induced
1484 : * task, send a signal to interrupt it.
1485 : */
1486 3848 : if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1487 : {
1488 0 : PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1489 : uint8 statusFlags;
1490 : uint8 lockmethod_copy;
1491 : LOCKTAG locktag_copy;
1492 :
1493 : /*
1494 : * Grab info we need, then release lock immediately. Note this
1495 : * coding means that there is a tiny chance that the process
1496 : * terminates its current transaction and starts a different one
1497 : * before we have a change to send the signal; the worst possible
1498 : * consequence is that a for-wraparound vacuum is canceled. But
1499 : * that could happen in any case unless we were to do kill() with
1500 : * the lock held, which is much more undesirable.
1501 : */
1502 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1503 0 : statusFlags = ProcGlobal->statusFlags[autovac->pgxactoff];
1504 0 : lockmethod_copy = lock->tag.locktag_lockmethodid;
1505 0 : locktag_copy = lock->tag;
1506 0 : LWLockRelease(ProcArrayLock);
1507 :
1508 : /*
1509 : * Only do it if the worker is not working to protect against Xid
1510 : * wraparound.
1511 : */
1512 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
1513 0 : !(statusFlags & PROC_VACUUM_FOR_WRAPAROUND))
1514 : {
1515 0 : int pid = autovac->pid;
1516 :
1517 : /* report the case, if configured to do so */
1518 0 : if (message_level_is_interesting(DEBUG1))
1519 : {
1520 : StringInfoData locktagbuf;
1521 : StringInfoData logbuf; /* errdetail for server log */
1522 :
1523 0 : initStringInfo(&locktagbuf);
1524 0 : initStringInfo(&logbuf);
1525 0 : DescribeLockTag(&locktagbuf, &locktag_copy);
1526 0 : appendStringInfo(&logbuf,
1527 : "Process %d waits for %s on %s.",
1528 : MyProcPid,
1529 : GetLockmodeName(lockmethod_copy, lockmode),
1530 : locktagbuf.data);
1531 :
1532 0 : ereport(DEBUG1,
1533 : (errmsg_internal("sending cancel to blocking autovacuum PID %d",
1534 : pid),
1535 : errdetail_log("%s", logbuf.data)));
1536 :
1537 0 : pfree(locktagbuf.data);
1538 0 : pfree(logbuf.data);
1539 : }
1540 :
1541 : /* send the autovacuum worker Back to Old Kent Road */
1542 0 : if (kill(pid, SIGINT) < 0)
1543 : {
1544 : /*
1545 : * There's a race condition here: once we release the
1546 : * ProcArrayLock, it's possible for the autovac worker to
1547 : * close up shop and exit before we can do the kill().
1548 : * Therefore, we do not whinge about no-such-process.
1549 : * Other errors such as EPERM could conceivably happen if
1550 : * the kernel recycles the PID fast enough, but such cases
1551 : * seem improbable enough that it's probably best to issue
1552 : * a warning if we see some other errno.
1553 : */
1554 0 : if (errno != ESRCH)
1555 0 : ereport(WARNING,
1556 : (errmsg("could not send signal to process %d: %m",
1557 : pid)));
1558 : }
1559 : }
1560 :
1561 : /* prevent signal from being sent again more than once */
1562 0 : allow_autovacuum_cancel = false;
1563 : }
1564 :
1565 : /*
1566 : * If awoken after the deadlock check interrupt has run, and
1567 : * log_lock_waits is on, then report about the wait.
1568 : */
1569 3848 : if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1570 : {
1571 : StringInfoData buf,
1572 : lock_waiters_sbuf,
1573 : lock_holders_sbuf;
1574 : const char *modename;
1575 : long secs;
1576 : int usecs;
1577 : long msecs;
1578 1048 : int lockHoldersNum = 0;
1579 :
1580 1048 : initStringInfo(&buf);
1581 1048 : initStringInfo(&lock_waiters_sbuf);
1582 1048 : initStringInfo(&lock_holders_sbuf);
1583 :
1584 1048 : DescribeLockTag(&buf, &locallock->tag.lock);
1585 1048 : modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1586 : lockmode);
1587 1048 : TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1588 : GetCurrentTimestamp(),
1589 : &secs, &usecs);
1590 1048 : msecs = secs * 1000 + usecs / 1000;
1591 1048 : usecs = usecs % 1000;
1592 :
1593 : /* Gather a list of all lock holders and waiters */
1594 1048 : LWLockAcquire(partitionLock, LW_SHARED);
1595 1048 : GetLockHoldersAndWaiters(locallock, &lock_holders_sbuf,
1596 : &lock_waiters_sbuf, &lockHoldersNum);
1597 1048 : LWLockRelease(partitionLock);
1598 :
1599 1048 : if (deadlock_state == DS_SOFT_DEADLOCK)
1600 6 : ereport(LOG,
1601 : (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1602 : MyProcPid, modename, buf.data, msecs, usecs),
1603 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1604 : "Processes holding the lock: %s. Wait queue: %s.",
1605 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1606 1042 : else if (deadlock_state == DS_HARD_DEADLOCK)
1607 : {
1608 : /*
1609 : * This message is a bit redundant with the error that will be
1610 : * reported subsequently, but in some cases the error report
1611 : * might not make it to the log (eg, if it's caught by an
1612 : * exception handler), and we want to ensure all long-wait
1613 : * events get logged.
1614 : */
1615 10 : ereport(LOG,
1616 : (errmsg("process %d detected deadlock while waiting for %s on %s after %ld.%03d ms",
1617 : MyProcPid, modename, buf.data, msecs, usecs),
1618 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1619 : "Processes holding the lock: %s. Wait queue: %s.",
1620 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1621 : }
1622 :
1623 1048 : if (myWaitStatus == PROC_WAIT_STATUS_WAITING)
1624 992 : ereport(LOG,
1625 : (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1626 : MyProcPid, modename, buf.data, msecs, usecs),
1627 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1628 : "Processes holding the lock: %s. Wait queue: %s.",
1629 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1630 56 : else if (myWaitStatus == PROC_WAIT_STATUS_OK)
1631 46 : ereport(LOG,
1632 : (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1633 : MyProcPid, modename, buf.data, msecs, usecs)));
1634 : else
1635 : {
1636 : Assert(myWaitStatus == PROC_WAIT_STATUS_ERROR);
1637 :
1638 : /*
1639 : * Currently, the deadlock checker always kicks its own
1640 : * process, which means that we'll only see
1641 : * PROC_WAIT_STATUS_ERROR when deadlock_state ==
1642 : * DS_HARD_DEADLOCK, and there's no need to print redundant
1643 : * messages. But for completeness and future-proofing, print
1644 : * a message if it looks like someone else kicked us off the
1645 : * lock.
1646 : */
1647 10 : if (deadlock_state != DS_HARD_DEADLOCK)
1648 0 : ereport(LOG,
1649 : (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1650 : MyProcPid, modename, buf.data, msecs, usecs),
1651 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1652 : "Processes holding the lock: %s. Wait queue: %s.",
1653 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1654 : }
1655 :
1656 : /*
1657 : * At this point we might still need to wait for the lock. Reset
1658 : * state so we don't print the above messages again.
1659 : */
1660 1048 : deadlock_state = DS_NO_DEADLOCK;
1661 :
1662 1048 : pfree(buf.data);
1663 1048 : pfree(lock_holders_sbuf.data);
1664 1048 : pfree(lock_waiters_sbuf.data);
1665 : }
1666 3848 : } while (myWaitStatus == PROC_WAIT_STATUS_WAITING);
1667 :
1668 : /*
1669 : * Disable the timers, if they are still running. As in LockErrorCleanup,
1670 : * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1671 : * already caused QueryCancelPending to become set, we want the cancel to
1672 : * be reported as a lock timeout, not a user cancel.
1673 : */
1674 2534 : if (!InHotStandby)
1675 : {
1676 2532 : if (LockTimeout > 0)
1677 : {
1678 : DisableTimeoutParams timeouts[2];
1679 :
1680 146 : timeouts[0].id = DEADLOCK_TIMEOUT;
1681 146 : timeouts[0].keep_indicator = false;
1682 146 : timeouts[1].id = LOCK_TIMEOUT;
1683 146 : timeouts[1].keep_indicator = true;
1684 146 : disable_timeouts(timeouts, 2);
1685 : }
1686 : else
1687 2386 : disable_timeout(DEADLOCK_TIMEOUT, false);
1688 : }
1689 :
1690 : /*
1691 : * Emit the log message if recovery conflict on lock was resolved but the
1692 : * startup process waited longer than deadlock_timeout for it.
1693 : */
1694 2534 : if (InHotStandby && logged_recovery_conflict)
1695 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1696 : standbyWaitStart, GetCurrentTimestamp(),
1697 : NULL, false);
1698 :
1699 : /*
1700 : * We don't have to do anything else, because the awaker did all the
1701 : * necessary updates of the lock table and MyProc. (The caller is
1702 : * responsible for updating the local lock table.)
1703 : */
1704 2534 : return myWaitStatus;
1705 : }
1706 :
1707 :
1708 : /*
1709 : * ProcWakeup -- wake up a process by setting its latch.
1710 : *
1711 : * Also remove the process from the wait queue and set its links invalid.
1712 : *
1713 : * The appropriate lock partition lock must be held by caller.
1714 : *
1715 : * XXX: presently, this code is only used for the "success" case, and only
1716 : * works correctly for that case. To clean up in failure case, would need
1717 : * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1718 : * Hence, in practice the waitStatus parameter must be PROC_WAIT_STATUS_OK.
1719 : */
1720 : void
1721 2526 : ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus)
1722 : {
1723 2526 : if (dlist_node_is_detached(&proc->links))
1724 0 : return;
1725 :
1726 : Assert(proc->waitStatus == PROC_WAIT_STATUS_WAITING);
1727 :
1728 : /* Remove process from wait queue */
1729 2526 : dclist_delete_from_thoroughly(&proc->waitLock->waitProcs, &proc->links);
1730 :
1731 : /* Clean up process' state and pass it the ok/fail signal */
1732 2526 : proc->waitLock = NULL;
1733 2526 : proc->waitProcLock = NULL;
1734 2526 : proc->waitStatus = waitStatus;
1735 2526 : pg_atomic_write_u64(&MyProc->waitStart, 0);
1736 :
1737 : /* And awaken it */
1738 2526 : SetLatch(&proc->procLatch);
1739 : }
1740 :
1741 : /*
1742 : * ProcLockWakeup -- routine for waking up processes when a lock is
1743 : * released (or a prior waiter is aborted). Scan all waiters
1744 : * for lock, waken any that are no longer blocked.
1745 : *
1746 : * The appropriate lock partition lock must be held by caller.
1747 : */
1748 : void
1749 2558 : ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1750 : {
1751 2558 : dclist_head *waitQueue = &lock->waitProcs;
1752 2558 : LOCKMASK aheadRequests = 0;
1753 : dlist_mutable_iter miter;
1754 :
1755 2558 : if (dclist_is_empty(waitQueue))
1756 90 : return;
1757 :
1758 5698 : dclist_foreach_modify(miter, waitQueue)
1759 : {
1760 3230 : PGPROC *proc = dlist_container(PGPROC, links, miter.cur);
1761 3230 : LOCKMODE lockmode = proc->waitLockMode;
1762 :
1763 : /*
1764 : * Waken if (a) doesn't conflict with requests of earlier waiters, and
1765 : * (b) doesn't conflict with already-held locks.
1766 : */
1767 3230 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1768 2928 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1769 : proc->waitProcLock))
1770 : {
1771 : /* OK to waken */
1772 2526 : GrantLock(lock, proc->waitProcLock, lockmode);
1773 : /* removes proc from the lock's waiting process queue */
1774 2526 : ProcWakeup(proc, PROC_WAIT_STATUS_OK);
1775 : }
1776 : else
1777 : {
1778 : /*
1779 : * Lock conflicts: Don't wake, but remember requested mode for
1780 : * later checks.
1781 : */
1782 704 : aheadRequests |= LOCKBIT_ON(lockmode);
1783 : }
1784 : }
1785 : }
1786 :
1787 : /*
1788 : * CheckDeadLock
1789 : *
1790 : * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1791 : * lock to be released by some other process. Check if there's a deadlock; if
1792 : * not, just return. (But signal ProcSleep to log a message, if
1793 : * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1794 : * the lock's wait queue and signal an error to ProcSleep.
1795 : */
1796 : static void
1797 62 : CheckDeadLock(void)
1798 : {
1799 : int i;
1800 :
1801 : /*
1802 : * Acquire exclusive lock on the entire shared lock data structures. Must
1803 : * grab LWLocks in partition-number order to avoid LWLock deadlock.
1804 : *
1805 : * Note that the deadlock check interrupt had better not be enabled
1806 : * anywhere that this process itself holds lock partition locks, else this
1807 : * will wait forever. Also note that LWLockAcquire creates a critical
1808 : * section, so that this routine cannot be interrupted by cancel/die
1809 : * interrupts.
1810 : */
1811 1054 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1812 992 : LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1813 :
1814 : /*
1815 : * Check to see if we've been awoken by anyone in the interim.
1816 : *
1817 : * If we have, we can return and resume our transaction -- happy day.
1818 : * Before we are awoken the process releasing the lock grants it to us so
1819 : * we know that we don't have to wait anymore.
1820 : *
1821 : * We check by looking to see if we've been unlinked from the wait queue.
1822 : * This is safe because we hold the lock partition lock.
1823 : */
1824 62 : if (MyProc->links.prev == NULL ||
1825 62 : MyProc->links.next == NULL)
1826 0 : goto check_done;
1827 :
1828 : #ifdef LOCK_DEBUG
1829 : if (Debug_deadlocks)
1830 : DumpAllLocks();
1831 : #endif
1832 :
1833 : /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1834 62 : deadlock_state = DeadLockCheck(MyProc);
1835 :
1836 62 : if (deadlock_state == DS_HARD_DEADLOCK)
1837 : {
1838 : /*
1839 : * Oops. We have a deadlock.
1840 : *
1841 : * Get this process out of wait state. (Note: we could do this more
1842 : * efficiently by relying on lockAwaited, but use this coding to
1843 : * preserve the flexibility to kill some other transaction than the
1844 : * one detecting the deadlock.)
1845 : *
1846 : * RemoveFromWaitQueue sets MyProc->waitStatus to
1847 : * PROC_WAIT_STATUS_ERROR, so ProcSleep will report an error after we
1848 : * return from the signal handler.
1849 : */
1850 : Assert(MyProc->waitLock != NULL);
1851 10 : RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1852 :
1853 : /*
1854 : * We're done here. Transaction abort caused by the error that
1855 : * ProcSleep will raise will cause any other locks we hold to be
1856 : * released, thus allowing other processes to wake up; we don't need
1857 : * to do that here. NOTE: an exception is that releasing locks we
1858 : * hold doesn't consider the possibility of waiters that were blocked
1859 : * behind us on the lock we just failed to get, and might now be
1860 : * wakable because we're not in front of them anymore. However,
1861 : * RemoveFromWaitQueue took care of waking up any such processes.
1862 : */
1863 : }
1864 :
1865 : /*
1866 : * And release locks. We do this in reverse order for two reasons: (1)
1867 : * Anyone else who needs more than one of the locks will be trying to lock
1868 : * them in increasing order; we don't want to release the other process
1869 : * until it can get all the locks it needs. (2) This avoids O(N^2)
1870 : * behavior inside LWLockRelease.
1871 : */
1872 52 : check_done:
1873 1054 : for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1874 992 : LWLockRelease(LockHashPartitionLockByIndex(i));
1875 62 : }
1876 :
1877 : /*
1878 : * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1879 : *
1880 : * NB: Runs inside a signal handler, be careful.
1881 : */
1882 : void
1883 62 : CheckDeadLockAlert(void)
1884 : {
1885 62 : int save_errno = errno;
1886 :
1887 62 : got_deadlock_timeout = true;
1888 :
1889 : /*
1890 : * Have to set the latch again, even if handle_sig_alarm already did. Back
1891 : * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1892 : * ever would be a problem, but setting a set latch again is cheap.
1893 : *
1894 : * Note that, when this function runs inside procsignal_sigusr1_handler(),
1895 : * the handler function sets the latch again after the latch is set here.
1896 : */
1897 62 : SetLatch(MyLatch);
1898 62 : errno = save_errno;
1899 62 : }
1900 :
1901 : /*
1902 : * GetLockHoldersAndWaiters - get lock holders and waiters for a lock
1903 : *
1904 : * Fill lock_holders_sbuf and lock_waiters_sbuf with the PIDs of processes holding
1905 : * and waiting for the lock, and set lockHoldersNum to the number of lock holders.
1906 : *
1907 : * The lock table's partition lock must be held on entry and remains held on exit.
1908 : */
1909 : void
1910 1048 : GetLockHoldersAndWaiters(LOCALLOCK *locallock, StringInfo lock_holders_sbuf,
1911 : StringInfo lock_waiters_sbuf, int *lockHoldersNum)
1912 : {
1913 : dlist_iter proc_iter;
1914 : PROCLOCK *curproclock;
1915 1048 : LOCK *lock = locallock->lock;
1916 1048 : bool first_holder = true,
1917 1048 : first_waiter = true;
1918 :
1919 : #ifdef USE_ASSERT_CHECKING
1920 : {
1921 : uint32 hashcode = locallock->hashcode;
1922 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1923 :
1924 : Assert(LWLockHeldByMe(partitionLock));
1925 : }
1926 : #endif
1927 :
1928 1048 : *lockHoldersNum = 0;
1929 :
1930 : /*
1931 : * Loop over the lock's procLocks to gather a list of all holders and
1932 : * waiters. Thus we will be able to provide more detailed information for
1933 : * lock debugging purposes.
1934 : *
1935 : * lock->procLocks contains all processes which hold or wait for this
1936 : * lock.
1937 : */
1938 3154 : dlist_foreach(proc_iter, &lock->procLocks)
1939 : {
1940 2106 : curproclock =
1941 2106 : dlist_container(PROCLOCK, lockLink, proc_iter.cur);
1942 :
1943 : /*
1944 : * We are a waiter if myProc->waitProcLock == curproclock; we are a
1945 : * holder if it is NULL or something different.
1946 : */
1947 2106 : if (curproclock->tag.myProc->waitProcLock == curproclock)
1948 : {
1949 1032 : if (first_waiter)
1950 : {
1951 994 : appendStringInfo(lock_waiters_sbuf, "%d",
1952 994 : curproclock->tag.myProc->pid);
1953 994 : first_waiter = false;
1954 : }
1955 : else
1956 38 : appendStringInfo(lock_waiters_sbuf, ", %d",
1957 38 : curproclock->tag.myProc->pid);
1958 : }
1959 : else
1960 : {
1961 1074 : if (first_holder)
1962 : {
1963 1048 : appendStringInfo(lock_holders_sbuf, "%d",
1964 1048 : curproclock->tag.myProc->pid);
1965 1048 : first_holder = false;
1966 : }
1967 : else
1968 26 : appendStringInfo(lock_holders_sbuf, ", %d",
1969 26 : curproclock->tag.myProc->pid);
1970 :
1971 1074 : (*lockHoldersNum)++;
1972 : }
1973 : }
1974 1048 : }
1975 :
1976 : /*
1977 : * ProcWaitForSignal - wait for a signal from another backend.
1978 : *
1979 : * As this uses the generic process latch the caller has to be robust against
1980 : * unrelated wakeups: Always check that the desired state has occurred, and
1981 : * wait again if not.
1982 : */
1983 : void
1984 182 : ProcWaitForSignal(uint32 wait_event_info)
1985 : {
1986 182 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1987 : wait_event_info);
1988 182 : ResetLatch(MyLatch);
1989 182 : CHECK_FOR_INTERRUPTS();
1990 182 : }
1991 :
1992 : /*
1993 : * ProcSendSignal - set the latch of a backend identified by ProcNumber
1994 : */
1995 : void
1996 158 : ProcSendSignal(ProcNumber procNumber)
1997 : {
1998 158 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
1999 0 : elog(ERROR, "procNumber out of range");
2000 :
2001 158 : SetLatch(&ProcGlobal->allProcs[procNumber].procLatch);
2002 158 : }
2003 :
2004 : /*
2005 : * BecomeLockGroupLeader - designate process as lock group leader
2006 : *
2007 : * Once this function has returned, other processes can join the lock group
2008 : * by calling BecomeLockGroupMember.
2009 : */
2010 : void
2011 1174 : BecomeLockGroupLeader(void)
2012 : {
2013 : LWLock *leader_lwlock;
2014 :
2015 : /* If we already did it, we don't need to do it again. */
2016 1174 : if (MyProc->lockGroupLeader == MyProc)
2017 1030 : return;
2018 :
2019 : /* We had better not be a follower. */
2020 : Assert(MyProc->lockGroupLeader == NULL);
2021 :
2022 : /* Create single-member group, containing only ourselves. */
2023 144 : leader_lwlock = LockHashPartitionLockByProc(MyProc);
2024 144 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2025 144 : MyProc->lockGroupLeader = MyProc;
2026 144 : dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
2027 144 : LWLockRelease(leader_lwlock);
2028 : }
2029 :
2030 : /*
2031 : * BecomeLockGroupMember - designate process as lock group member
2032 : *
2033 : * This is pretty straightforward except for the possibility that the leader
2034 : * whose group we're trying to join might exit before we manage to do so;
2035 : * and the PGPROC might get recycled for an unrelated process. To avoid
2036 : * that, we require the caller to pass the PID of the intended PGPROC as
2037 : * an interlock. Returns true if we successfully join the intended lock
2038 : * group, and false if not.
2039 : */
2040 : bool
2041 2748 : BecomeLockGroupMember(PGPROC *leader, int pid)
2042 : {
2043 : LWLock *leader_lwlock;
2044 2748 : bool ok = false;
2045 :
2046 : /* Group leader can't become member of group */
2047 : Assert(MyProc != leader);
2048 :
2049 : /* Can't already be a member of a group */
2050 : Assert(MyProc->lockGroupLeader == NULL);
2051 :
2052 : /* PID must be valid. */
2053 : Assert(pid != 0);
2054 :
2055 : /*
2056 : * Get lock protecting the group fields. Note LockHashPartitionLockByProc
2057 : * calculates the proc number based on the PGPROC slot without looking at
2058 : * its contents, so we will acquire the correct lock even if the leader
2059 : * PGPROC is in process of being recycled.
2060 : */
2061 2748 : leader_lwlock = LockHashPartitionLockByProc(leader);
2062 2748 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2063 :
2064 : /* Is this the leader we're looking for? */
2065 2748 : if (leader->pid == pid && leader->lockGroupLeader == leader)
2066 : {
2067 : /* OK, join the group */
2068 2748 : ok = true;
2069 2748 : MyProc->lockGroupLeader = leader;
2070 2748 : dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
2071 : }
2072 2748 : LWLockRelease(leader_lwlock);
2073 :
2074 2748 : return ok;
2075 : }
|
