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 "miscadmin.h"
40 : #include "pgstat.h"
41 : #include "postmaster/autovacuum.h"
42 : #include "replication/slotsync.h"
43 : #include "replication/syncrep.h"
44 : #include "storage/condition_variable.h"
45 : #include "storage/ipc.h"
46 : #include "storage/lmgr.h"
47 : #include "storage/pmsignal.h"
48 : #include "storage/proc.h"
49 : #include "storage/procarray.h"
50 : #include "storage/procsignal.h"
51 : #include "storage/spin.h"
52 : #include "storage/standby.h"
53 : #include "utils/timeout.h"
54 : #include "utils/timestamp.h"
55 :
56 : /* GUC variables */
57 : int DeadlockTimeout = 1000;
58 : int StatementTimeout = 0;
59 : int LockTimeout = 0;
60 : int IdleInTransactionSessionTimeout = 0;
61 : int TransactionTimeout = 0;
62 : int IdleSessionTimeout = 0;
63 : bool log_lock_waits = false;
64 :
65 : /* Pointer to this process's PGPROC struct, if any */
66 : PGPROC *MyProc = NULL;
67 :
68 : /*
69 : * This spinlock protects the freelist of recycled PGPROC structures.
70 : * We cannot use an LWLock because the LWLock manager depends on already
71 : * having a PGPROC and a wait semaphore! But these structures are touched
72 : * relatively infrequently (only at backend startup or shutdown) and not for
73 : * very long, so a spinlock is okay.
74 : */
75 : NON_EXEC_STATIC slock_t *ProcStructLock = NULL;
76 :
77 : /* Pointers to shared-memory structures */
78 : PROC_HDR *ProcGlobal = NULL;
79 : NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
80 : PGPROC *PreparedXactProcs = NULL;
81 :
82 : static DeadLockState deadlock_state = DS_NOT_YET_CHECKED;
83 :
84 : /* Is a deadlock check pending? */
85 : static volatile sig_atomic_t got_deadlock_timeout;
86 :
87 : static void RemoveProcFromArray(int code, Datum arg);
88 : static void ProcKill(int code, Datum arg);
89 : static void AuxiliaryProcKill(int code, Datum arg);
90 : static void CheckDeadLock(void);
91 :
92 :
93 : /*
94 : * Report shared-memory space needed by InitProcGlobal.
95 : */
96 : Size
97 3566 : ProcGlobalShmemSize(void)
98 : {
99 3566 : Size size = 0;
100 : Size TotalProcs =
101 3566 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
102 : Size fpLockBitsSize,
103 : fpRelIdSize;
104 :
105 : /* ProcGlobal */
106 3566 : size = add_size(size, sizeof(PROC_HDR));
107 3566 : size = add_size(size, mul_size(TotalProcs, sizeof(PGPROC)));
108 3566 : size = add_size(size, sizeof(slock_t));
109 :
110 3566 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->xids)));
111 3566 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->subxidStates)));
112 3566 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->statusFlags)));
113 :
114 : /*
115 : * Memory needed for PGPROC fast-path lock arrays. Make sure the sizes are
116 : * nicely aligned in each backend.
117 : */
118 3566 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
119 3566 : fpRelIdSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(Oid) * FP_LOCK_SLOTS_PER_GROUP);
120 :
121 3566 : size = add_size(size, mul_size(TotalProcs, (fpLockBitsSize + fpRelIdSize)));
122 :
123 3566 : return size;
124 : }
125 :
126 : /*
127 : * Report number of semaphores needed by InitProcGlobal.
128 : */
129 : int
130 3566 : ProcGlobalSemas(void)
131 : {
132 : /*
133 : * We need a sema per backend (including autovacuum), plus one for each
134 : * auxiliary process.
135 : */
136 3566 : return MaxBackends + NUM_AUXILIARY_PROCS;
137 : }
138 :
139 : /*
140 : * InitProcGlobal -
141 : * Initialize the global process table during postmaster or standalone
142 : * backend startup.
143 : *
144 : * We also create all the per-process semaphores we will need to support
145 : * the requested number of backends. We used to allocate semaphores
146 : * only when backends were actually started up, but that is bad because
147 : * it lets Postgres fail under load --- a lot of Unix systems are
148 : * (mis)configured with small limits on the number of semaphores, and
149 : * running out when trying to start another backend is a common failure.
150 : * So, now we grab enough semaphores to support the desired max number
151 : * of backends immediately at initialization --- if the sysadmin has set
152 : * MaxConnections, max_worker_processes, max_wal_senders, or
153 : * autovacuum_worker_slots higher than his kernel will support, he'll
154 : * find out sooner rather than later.
155 : *
156 : * Another reason for creating semaphores here is that the semaphore
157 : * implementation typically requires us to create semaphores in the
158 : * postmaster, not in backends.
159 : *
160 : * Note: this is NOT called by individual backends under a postmaster,
161 : * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
162 : * pointers must be propagated specially for EXEC_BACKEND operation.
163 : */
164 : void
165 1918 : InitProcGlobal(void)
166 : {
167 : PGPROC *procs;
168 : int i,
169 : j;
170 : bool found;
171 1918 : uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
172 :
173 : /* Used for setup of per-backend fast-path slots. */
174 : char *fpPtr,
175 : *fpEndPtr PG_USED_FOR_ASSERTS_ONLY;
176 : Size fpLockBitsSize,
177 : fpRelIdSize;
178 :
179 : /* Create the ProcGlobal shared structure */
180 1918 : ProcGlobal = (PROC_HDR *)
181 1918 : ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
182 : Assert(!found);
183 :
184 : /*
185 : * Initialize the data structures.
186 : */
187 1918 : ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
188 1918 : dlist_init(&ProcGlobal->freeProcs);
189 1918 : dlist_init(&ProcGlobal->autovacFreeProcs);
190 1918 : dlist_init(&ProcGlobal->bgworkerFreeProcs);
191 1918 : dlist_init(&ProcGlobal->walsenderFreeProcs);
192 1918 : ProcGlobal->startupBufferPinWaitBufId = -1;
193 1918 : ProcGlobal->walwriterProc = INVALID_PROC_NUMBER;
194 1918 : ProcGlobal->checkpointerProc = INVALID_PROC_NUMBER;
195 1918 : pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PROC_NUMBER);
196 1918 : pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PROC_NUMBER);
197 :
198 : /*
199 : * Create and initialize all the PGPROC structures we'll need. There are
200 : * six separate consumers: (1) normal backends, (2) autovacuum workers and
201 : * special workers, (3) background workers, (4) walsenders, (5) auxiliary
202 : * processes, and (6) prepared transactions. (For largely-historical
203 : * reasons, we combine autovacuum and special workers into one category
204 : * with a single freelist.) Each PGPROC structure is dedicated to exactly
205 : * one of these purposes, and they do not move between groups.
206 : */
207 1918 : procs = (PGPROC *) ShmemAlloc(TotalProcs * sizeof(PGPROC));
208 1918 : MemSet(procs, 0, TotalProcs * sizeof(PGPROC));
209 1918 : ProcGlobal->allProcs = procs;
210 : /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
211 1918 : ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
212 :
213 : /*
214 : * Allocate arrays mirroring PGPROC fields in a dense manner. See
215 : * PROC_HDR.
216 : *
217 : * XXX: It might make sense to increase padding for these arrays, given
218 : * how hotly they are accessed.
219 : */
220 3836 : ProcGlobal->xids =
221 1918 : (TransactionId *) ShmemAlloc(TotalProcs * sizeof(*ProcGlobal->xids));
222 77470 : MemSet(ProcGlobal->xids, 0, TotalProcs * sizeof(*ProcGlobal->xids));
223 1918 : ProcGlobal->subxidStates = (XidCacheStatus *) ShmemAlloc(TotalProcs * sizeof(*ProcGlobal->subxidStates));
224 25008 : MemSet(ProcGlobal->subxidStates, 0, TotalProcs * sizeof(*ProcGlobal->subxidStates));
225 1918 : ProcGlobal->statusFlags = (uint8 *) ShmemAlloc(TotalProcs * sizeof(*ProcGlobal->statusFlags));
226 5240 : MemSet(ProcGlobal->statusFlags, 0, TotalProcs * sizeof(*ProcGlobal->statusFlags));
227 :
228 : /*
229 : * Allocate arrays for fast-path locks. Those are variable-length, so
230 : * can't be included in PGPROC directly. We allocate a separate piece of
231 : * shared memory and then divide that between backends.
232 : */
233 1918 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
234 1918 : fpRelIdSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(Oid) * FP_LOCK_SLOTS_PER_GROUP);
235 :
236 1918 : fpPtr = ShmemAlloc(TotalProcs * (fpLockBitsSize + fpRelIdSize));
237 1918 : MemSet(fpPtr, 0, TotalProcs * (fpLockBitsSize + fpRelIdSize));
238 :
239 : /* For asserts checking we did not overflow. */
240 1918 : fpEndPtr = fpPtr + (TotalProcs * (fpLockBitsSize + fpRelIdSize));
241 :
242 191792 : for (i = 0; i < TotalProcs; i++)
243 : {
244 189874 : PGPROC *proc = &procs[i];
245 :
246 : /* Common initialization for all PGPROCs, regardless of type. */
247 :
248 : /*
249 : * Set the fast-path lock arrays, and move the pointer. We interleave
250 : * the two arrays, to (hopefully) get some locality for each backend.
251 : */
252 189874 : proc->fpLockBits = (uint64 *) fpPtr;
253 189874 : fpPtr += fpLockBitsSize;
254 :
255 189874 : proc->fpRelId = (Oid *) fpPtr;
256 189874 : fpPtr += fpRelIdSize;
257 :
258 : Assert(fpPtr <= fpEndPtr);
259 :
260 : /*
261 : * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
262 : * dummy PGPROCs don't need these though - they're never associated
263 : * with a real process
264 : */
265 189874 : if (i < MaxBackends + NUM_AUXILIARY_PROCS)
266 : {
267 188224 : proc->sem = PGSemaphoreCreate();
268 188224 : InitSharedLatch(&(proc->procLatch));
269 188224 : LWLockInitialize(&(proc->fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
270 : }
271 :
272 : /*
273 : * Newly created PGPROCs for normal backends, autovacuum workers,
274 : * special workers, bgworkers, and walsenders must be queued up on the
275 : * appropriate free list. Because there can only ever be a small,
276 : * fixed number of auxiliary processes, no free list is used in that
277 : * case; InitAuxiliaryProcess() instead uses a linear search. PGPROCs
278 : * for prepared transactions are added to a free list by
279 : * TwoPhaseShmemInit().
280 : */
281 189874 : if (i < MaxConnections)
282 : {
283 : /* PGPROC for normal backend, add to freeProcs list */
284 123822 : dlist_push_tail(&ProcGlobal->freeProcs, &proc->links);
285 123822 : proc->procgloballist = &ProcGlobal->freeProcs;
286 : }
287 66052 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS)
288 : {
289 : /* PGPROC for AV or special worker, add to autovacFreeProcs list */
290 24774 : dlist_push_tail(&ProcGlobal->autovacFreeProcs, &proc->links);
291 24774 : proc->procgloballist = &ProcGlobal->autovacFreeProcs;
292 : }
293 41278 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS + max_worker_processes)
294 : {
295 : /* PGPROC for bgworker, add to bgworkerFreeProcs list */
296 15340 : dlist_push_tail(&ProcGlobal->bgworkerFreeProcs, &proc->links);
297 15340 : proc->procgloballist = &ProcGlobal->bgworkerFreeProcs;
298 : }
299 25938 : else if (i < MaxBackends)
300 : {
301 : /* PGPROC for walsender, add to walsenderFreeProcs list */
302 12780 : dlist_push_tail(&ProcGlobal->walsenderFreeProcs, &proc->links);
303 12780 : proc->procgloballist = &ProcGlobal->walsenderFreeProcs;
304 : }
305 :
306 : /* Initialize myProcLocks[] shared memory queues. */
307 3227858 : for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
308 3037984 : dlist_init(&(proc->myProcLocks[j]));
309 :
310 : /* Initialize lockGroupMembers list. */
311 189874 : dlist_init(&proc->lockGroupMembers);
312 :
313 : /*
314 : * Initialize the atomic variables, otherwise, it won't be safe to
315 : * access them for backends that aren't currently in use.
316 : */
317 189874 : pg_atomic_init_u32(&(proc->procArrayGroupNext), INVALID_PROC_NUMBER);
318 189874 : pg_atomic_init_u32(&(proc->clogGroupNext), INVALID_PROC_NUMBER);
319 189874 : pg_atomic_init_u64(&(proc->waitStart), 0);
320 : }
321 :
322 : /* Should have consumed exactly the expected amount of fast-path memory. */
323 : Assert(fpPtr == fpEndPtr);
324 :
325 : /*
326 : * Save pointers to the blocks of PGPROC structures reserved for auxiliary
327 : * processes and prepared transactions.
328 : */
329 1918 : AuxiliaryProcs = &procs[MaxBackends];
330 1918 : PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
331 :
332 : /* Create ProcStructLock spinlock, too */
333 1918 : ProcStructLock = (slock_t *) ShmemAlloc(sizeof(slock_t));
334 1918 : SpinLockInit(ProcStructLock);
335 1918 : }
336 :
337 : /*
338 : * InitProcess -- initialize a per-process PGPROC entry for this backend
339 : */
340 : void
341 30210 : InitProcess(void)
342 : {
343 : dlist_head *procgloballist;
344 :
345 : /*
346 : * ProcGlobal should be set up already (if we are a backend, we inherit
347 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
348 : */
349 30210 : if (ProcGlobal == NULL)
350 0 : elog(PANIC, "proc header uninitialized");
351 :
352 30210 : if (MyProc != NULL)
353 0 : elog(ERROR, "you already exist");
354 :
355 : /*
356 : * Before we start accessing the shared memory in a serious way, mark
357 : * ourselves as an active postmaster child; this is so that the postmaster
358 : * can detect it if we exit without cleaning up.
359 : */
360 30210 : if (IsUnderPostmaster)
361 30012 : RegisterPostmasterChildActive();
362 :
363 : /*
364 : * Decide which list should supply our PGPROC. This logic must match the
365 : * way the freelists were constructed in InitProcGlobal().
366 : */
367 30210 : if (AmAutoVacuumWorkerProcess() || AmSpecialWorkerProcess())
368 2976 : procgloballist = &ProcGlobal->autovacFreeProcs;
369 27234 : else if (AmBackgroundWorkerProcess())
370 4290 : procgloballist = &ProcGlobal->bgworkerFreeProcs;
371 22944 : else if (AmWalSenderProcess())
372 2128 : procgloballist = &ProcGlobal->walsenderFreeProcs;
373 : else
374 20816 : procgloballist = &ProcGlobal->freeProcs;
375 :
376 : /*
377 : * Try to get a proc struct from the appropriate free list. If this
378 : * fails, we must be out of PGPROC structures (not to mention semaphores).
379 : *
380 : * While we are holding the ProcStructLock, also copy the current shared
381 : * estimate of spins_per_delay to local storage.
382 : */
383 30210 : SpinLockAcquire(ProcStructLock);
384 :
385 30210 : set_spins_per_delay(ProcGlobal->spins_per_delay);
386 :
387 30210 : if (!dlist_is_empty(procgloballist))
388 : {
389 30204 : MyProc = dlist_container(PGPROC, links, dlist_pop_head_node(procgloballist));
390 30204 : SpinLockRelease(ProcStructLock);
391 : }
392 : else
393 : {
394 : /*
395 : * If we reach here, all the PGPROCs are in use. This is one of the
396 : * possible places to detect "too many backends", so give the standard
397 : * error message. XXX do we need to give a different failure message
398 : * in the autovacuum case?
399 : */
400 6 : SpinLockRelease(ProcStructLock);
401 6 : if (AmWalSenderProcess())
402 4 : ereport(FATAL,
403 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
404 : errmsg("number of requested standby connections exceeds \"max_wal_senders\" (currently %d)",
405 : max_wal_senders)));
406 2 : ereport(FATAL,
407 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
408 : errmsg("sorry, too many clients already")));
409 : }
410 30204 : MyProcNumber = GetNumberFromPGProc(MyProc);
411 :
412 : /*
413 : * Cross-check that the PGPROC is of the type we expect; if this were not
414 : * the case, it would get returned to the wrong list.
415 : */
416 : Assert(MyProc->procgloballist == procgloballist);
417 :
418 : /*
419 : * Initialize all fields of MyProc, except for those previously
420 : * initialized by InitProcGlobal.
421 : */
422 30204 : dlist_node_init(&MyProc->links);
423 30204 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
424 30204 : MyProc->fpVXIDLock = false;
425 30204 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
426 30204 : MyProc->xid = InvalidTransactionId;
427 30204 : MyProc->xmin = InvalidTransactionId;
428 30204 : MyProc->pid = MyProcPid;
429 30204 : MyProc->vxid.procNumber = MyProcNumber;
430 30204 : MyProc->vxid.lxid = InvalidLocalTransactionId;
431 : /* databaseId and roleId will be filled in later */
432 30204 : MyProc->databaseId = InvalidOid;
433 30204 : MyProc->roleId = InvalidOid;
434 30204 : MyProc->tempNamespaceId = InvalidOid;
435 30204 : MyProc->isRegularBackend = AmRegularBackendProcess();
436 30204 : MyProc->delayChkptFlags = 0;
437 30204 : MyProc->statusFlags = 0;
438 : /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
439 30204 : if (AmAutoVacuumWorkerProcess())
440 2268 : MyProc->statusFlags |= PROC_IS_AUTOVACUUM;
441 30204 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
442 30204 : MyProc->lwWaitMode = 0;
443 30204 : MyProc->waitLock = NULL;
444 30204 : MyProc->waitProcLock = NULL;
445 30204 : pg_atomic_write_u64(&MyProc->waitStart, 0);
446 : #ifdef USE_ASSERT_CHECKING
447 : {
448 : int i;
449 :
450 : /* Last process should have released all locks. */
451 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
452 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
453 : }
454 : #endif
455 30204 : MyProc->recoveryConflictPending = false;
456 :
457 : /* Initialize fields for sync rep */
458 30204 : MyProc->waitLSN = 0;
459 30204 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
460 30204 : dlist_node_init(&MyProc->syncRepLinks);
461 :
462 : /* Initialize fields for group XID clearing. */
463 30204 : MyProc->procArrayGroupMember = false;
464 30204 : MyProc->procArrayGroupMemberXid = InvalidTransactionId;
465 : Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PROC_NUMBER);
466 :
467 : /* Check that group locking fields are in a proper initial state. */
468 : Assert(MyProc->lockGroupLeader == NULL);
469 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
470 :
471 : /* Initialize wait event information. */
472 30204 : MyProc->wait_event_info = 0;
473 :
474 : /* Initialize fields for group transaction status update. */
475 30204 : MyProc->clogGroupMember = false;
476 30204 : MyProc->clogGroupMemberXid = InvalidTransactionId;
477 30204 : MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
478 30204 : MyProc->clogGroupMemberPage = -1;
479 30204 : MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
480 : Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PROC_NUMBER);
481 :
482 : /*
483 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
484 : * on it. That allows us to repoint the process latch, which so far
485 : * points to process local one, to the shared one.
486 : */
487 30204 : OwnLatch(&MyProc->procLatch);
488 30204 : SwitchToSharedLatch();
489 :
490 : /* now that we have a proc, report wait events to shared memory */
491 30204 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
492 :
493 : /*
494 : * We might be reusing a semaphore that belonged to a failed process. So
495 : * be careful and reinitialize its value here. (This is not strictly
496 : * necessary anymore, but seems like a good idea for cleanliness.)
497 : */
498 30204 : PGSemaphoreReset(MyProc->sem);
499 :
500 : /*
501 : * Arrange to clean up at backend exit.
502 : */
503 30204 : on_shmem_exit(ProcKill, 0);
504 :
505 : /*
506 : * Now that we have a PGPROC, we could try to acquire locks, so initialize
507 : * local state needed for LWLocks, and the deadlock checker.
508 : */
509 30204 : InitLWLockAccess();
510 30204 : InitDeadLockChecking();
511 :
512 : #ifdef EXEC_BACKEND
513 :
514 : /*
515 : * Initialize backend-local pointers to all the shared data structures.
516 : * (We couldn't do this until now because it needs LWLocks.)
517 : */
518 : if (IsUnderPostmaster)
519 : AttachSharedMemoryStructs();
520 : #endif
521 30204 : }
522 :
523 : /*
524 : * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
525 : *
526 : * This is separate from InitProcess because we can't acquire LWLocks until
527 : * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
528 : * work until after we've done AttachSharedMemoryStructs.
529 : */
530 : void
531 30186 : InitProcessPhase2(void)
532 : {
533 : Assert(MyProc != NULL);
534 :
535 : /*
536 : * Add our PGPROC to the PGPROC array in shared memory.
537 : */
538 30186 : ProcArrayAdd(MyProc);
539 :
540 : /*
541 : * Arrange to clean that up at backend exit.
542 : */
543 30186 : on_shmem_exit(RemoveProcFromArray, 0);
544 30186 : }
545 :
546 : /*
547 : * InitAuxiliaryProcess -- create a PGPROC entry for an auxiliary process
548 : *
549 : * This is called by bgwriter and similar processes so that they will have a
550 : * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
551 : * and sema that are assigned are one of the extra ones created during
552 : * InitProcGlobal.
553 : *
554 : * Auxiliary processes are presently not expected to wait for real (lockmgr)
555 : * locks, so we need not set up the deadlock checker. They are never added
556 : * to the ProcArray or the sinval messaging mechanism, either. They also
557 : * don't get a VXID assigned, since this is only useful when we actually
558 : * hold lockmgr locks.
559 : *
560 : * Startup process however uses locks but never waits for them in the
561 : * normal backend sense. Startup process also takes part in sinval messaging
562 : * as a sendOnly process, so never reads messages from sinval queue. So
563 : * Startup process does have a VXID and does show up in pg_locks.
564 : */
565 : void
566 4496 : InitAuxiliaryProcess(void)
567 : {
568 : PGPROC *auxproc;
569 : int proctype;
570 :
571 : /*
572 : * ProcGlobal should be set up already (if we are a backend, we inherit
573 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
574 : */
575 4496 : if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
576 0 : elog(PANIC, "proc header uninitialized");
577 :
578 4496 : if (MyProc != NULL)
579 0 : elog(ERROR, "you already exist");
580 :
581 4496 : if (IsUnderPostmaster)
582 4496 : RegisterPostmasterChildActive();
583 :
584 : /*
585 : * We use the ProcStructLock to protect assignment and releasing of
586 : * AuxiliaryProcs entries.
587 : *
588 : * While we are holding the ProcStructLock, also copy the current shared
589 : * estimate of spins_per_delay to local storage.
590 : */
591 4496 : SpinLockAcquire(ProcStructLock);
592 :
593 4496 : set_spins_per_delay(ProcGlobal->spins_per_delay);
594 :
595 : /*
596 : * Find a free auxproc ... *big* trouble if there isn't one ...
597 : */
598 11176 : for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
599 : {
600 11176 : auxproc = &AuxiliaryProcs[proctype];
601 11176 : if (auxproc->pid == 0)
602 4496 : break;
603 : }
604 4496 : if (proctype >= NUM_AUXILIARY_PROCS)
605 : {
606 0 : SpinLockRelease(ProcStructLock);
607 0 : elog(FATAL, "all AuxiliaryProcs are in use");
608 : }
609 :
610 : /* Mark auxiliary proc as in use by me */
611 : /* use volatile pointer to prevent code rearrangement */
612 4496 : ((volatile PGPROC *) auxproc)->pid = MyProcPid;
613 :
614 4496 : SpinLockRelease(ProcStructLock);
615 :
616 4496 : MyProc = auxproc;
617 4496 : MyProcNumber = GetNumberFromPGProc(MyProc);
618 :
619 : /*
620 : * Initialize all fields of MyProc, except for those previously
621 : * initialized by InitProcGlobal.
622 : */
623 4496 : dlist_node_init(&MyProc->links);
624 4496 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
625 4496 : MyProc->fpVXIDLock = false;
626 4496 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
627 4496 : MyProc->xid = InvalidTransactionId;
628 4496 : MyProc->xmin = InvalidTransactionId;
629 4496 : MyProc->vxid.procNumber = INVALID_PROC_NUMBER;
630 4496 : MyProc->vxid.lxid = InvalidLocalTransactionId;
631 4496 : MyProc->databaseId = InvalidOid;
632 4496 : MyProc->roleId = InvalidOid;
633 4496 : MyProc->tempNamespaceId = InvalidOid;
634 4496 : MyProc->isRegularBackend = false;
635 4496 : MyProc->delayChkptFlags = 0;
636 4496 : MyProc->statusFlags = 0;
637 4496 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
638 4496 : MyProc->lwWaitMode = 0;
639 4496 : MyProc->waitLock = NULL;
640 4496 : MyProc->waitProcLock = NULL;
641 4496 : pg_atomic_write_u64(&MyProc->waitStart, 0);
642 : #ifdef USE_ASSERT_CHECKING
643 : {
644 : int i;
645 :
646 : /* Last process should have released all locks. */
647 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
648 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
649 : }
650 : #endif
651 :
652 : /*
653 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
654 : * on it. That allows us to repoint the process latch, which so far
655 : * points to process local one, to the shared one.
656 : */
657 4496 : OwnLatch(&MyProc->procLatch);
658 4496 : SwitchToSharedLatch();
659 :
660 : /* now that we have a proc, report wait events to shared memory */
661 4496 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
662 :
663 : /* Check that group locking fields are in a proper initial state. */
664 : Assert(MyProc->lockGroupLeader == NULL);
665 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
666 :
667 : /*
668 : * We might be reusing a semaphore that belonged to a failed process. So
669 : * be careful and reinitialize its value here. (This is not strictly
670 : * necessary anymore, but seems like a good idea for cleanliness.)
671 : */
672 4496 : PGSemaphoreReset(MyProc->sem);
673 :
674 : /*
675 : * Arrange to clean up at process exit.
676 : */
677 4496 : on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
678 :
679 : /*
680 : * Now that we have a PGPROC, we could try to acquire lightweight locks.
681 : * Initialize local state needed for them. (Heavyweight locks cannot be
682 : * acquired in aux processes.)
683 : */
684 4496 : InitLWLockAccess();
685 :
686 : #ifdef EXEC_BACKEND
687 :
688 : /*
689 : * Initialize backend-local pointers to all the shared data structures.
690 : * (We couldn't do this until now because it needs LWLocks.)
691 : */
692 : if (IsUnderPostmaster)
693 : AttachSharedMemoryStructs();
694 : #endif
695 4496 : }
696 :
697 : /*
698 : * Used from bufmgr to share the value of the buffer that Startup waits on,
699 : * or to reset the value to "not waiting" (-1). This allows processing
700 : * of recovery conflicts for buffer pins. Set is made before backends look
701 : * at this value, so locking not required, especially since the set is
702 : * an atomic integer set operation.
703 : */
704 : void
705 40 : SetStartupBufferPinWaitBufId(int bufid)
706 : {
707 : /* use volatile pointer to prevent code rearrangement */
708 40 : volatile PROC_HDR *procglobal = ProcGlobal;
709 :
710 40 : procglobal->startupBufferPinWaitBufId = bufid;
711 40 : }
712 :
713 : /*
714 : * Used by backends when they receive a request to check for buffer pin waits.
715 : */
716 : int
717 10 : GetStartupBufferPinWaitBufId(void)
718 : {
719 : /* use volatile pointer to prevent code rearrangement */
720 10 : volatile PROC_HDR *procglobal = ProcGlobal;
721 :
722 10 : return procglobal->startupBufferPinWaitBufId;
723 : }
724 :
725 : /*
726 : * Check whether there are at least N free PGPROC objects. If false is
727 : * returned, *nfree will be set to the number of free PGPROC objects.
728 : * Otherwise, *nfree will be set to n.
729 : *
730 : * Note: this is designed on the assumption that N will generally be small.
731 : */
732 : bool
733 446 : HaveNFreeProcs(int n, int *nfree)
734 : {
735 : dlist_iter iter;
736 :
737 : Assert(n > 0);
738 : Assert(nfree);
739 :
740 446 : SpinLockAcquire(ProcStructLock);
741 :
742 446 : *nfree = 0;
743 1332 : dlist_foreach(iter, &ProcGlobal->freeProcs)
744 : {
745 1324 : (*nfree)++;
746 1324 : if (*nfree == n)
747 438 : break;
748 : }
749 :
750 446 : SpinLockRelease(ProcStructLock);
751 :
752 446 : return (*nfree == n);
753 : }
754 :
755 : /*
756 : * Cancel any pending wait for lock, when aborting a transaction, and revert
757 : * any strong lock count acquisition for a lock being acquired.
758 : *
759 : * (Normally, this would only happen if we accept a cancel/die
760 : * interrupt while waiting; but an ereport(ERROR) before or during the lock
761 : * wait is within the realm of possibility, too.)
762 : */
763 : void
764 843614 : LockErrorCleanup(void)
765 : {
766 : LOCALLOCK *lockAwaited;
767 : LWLock *partitionLock;
768 : DisableTimeoutParams timeouts[2];
769 :
770 843614 : HOLD_INTERRUPTS();
771 :
772 843614 : AbortStrongLockAcquire();
773 :
774 : /* Nothing to do if we weren't waiting for a lock */
775 843614 : lockAwaited = GetAwaitedLock();
776 843614 : if (lockAwaited == NULL)
777 : {
778 843266 : RESUME_INTERRUPTS();
779 843266 : return;
780 : }
781 :
782 : /*
783 : * Turn off the deadlock and lock timeout timers, if they are still
784 : * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
785 : * indicator flag, since this function is executed before
786 : * ProcessInterrupts when responding to SIGINT; else we'd lose the
787 : * knowledge that the SIGINT came from a lock timeout and not an external
788 : * source.
789 : */
790 348 : timeouts[0].id = DEADLOCK_TIMEOUT;
791 348 : timeouts[0].keep_indicator = false;
792 348 : timeouts[1].id = LOCK_TIMEOUT;
793 348 : timeouts[1].keep_indicator = true;
794 348 : disable_timeouts(timeouts, 2);
795 :
796 : /* Unlink myself from the wait queue, if on it (might not be anymore!) */
797 348 : partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
798 348 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
799 :
800 348 : if (!dlist_node_is_detached(&MyProc->links))
801 : {
802 : /* We could not have been granted the lock yet */
803 80 : RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
804 : }
805 : else
806 : {
807 : /*
808 : * Somebody kicked us off the lock queue already. Perhaps they
809 : * granted us the lock, or perhaps they detected a deadlock. If they
810 : * did grant us the lock, we'd better remember it in our local lock
811 : * table.
812 : */
813 268 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
814 4 : GrantAwaitedLock();
815 : }
816 :
817 348 : LWLockRelease(partitionLock);
818 :
819 348 : RESUME_INTERRUPTS();
820 : }
821 :
822 :
823 : /*
824 : * ProcReleaseLocks() -- release locks associated with current transaction
825 : * at main transaction commit or abort
826 : *
827 : * At main transaction commit, we release standard locks except session locks.
828 : * At main transaction abort, we release all locks including session locks.
829 : *
830 : * Advisory locks are released only if they are transaction-level;
831 : * session-level holds remain, whether this is a commit or not.
832 : *
833 : * At subtransaction commit, we don't release any locks (so this func is not
834 : * needed at all); we will defer the releasing to the parent transaction.
835 : * At subtransaction abort, we release all locks held by the subtransaction;
836 : * this is implemented by retail releasing of the locks under control of
837 : * the ResourceOwner mechanism.
838 : */
839 : void
840 786098 : ProcReleaseLocks(bool isCommit)
841 : {
842 786098 : if (!MyProc)
843 0 : return;
844 : /* If waiting, get off wait queue (should only be needed after error) */
845 786098 : LockErrorCleanup();
846 : /* Release standard locks, including session-level if aborting */
847 786098 : LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
848 : /* Release transaction-level advisory locks */
849 786098 : LockReleaseAll(USER_LOCKMETHOD, false);
850 : }
851 :
852 :
853 : /*
854 : * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
855 : */
856 : static void
857 30186 : RemoveProcFromArray(int code, Datum arg)
858 : {
859 : Assert(MyProc != NULL);
860 30186 : ProcArrayRemove(MyProc, InvalidTransactionId);
861 30186 : }
862 :
863 : /*
864 : * ProcKill() -- Destroy the per-proc data structure for
865 : * this process. Release any of its held LW locks.
866 : */
867 : static void
868 30204 : ProcKill(int code, Datum arg)
869 : {
870 : PGPROC *proc;
871 : dlist_head *procgloballist;
872 :
873 : Assert(MyProc != NULL);
874 :
875 : /* not safe if forked by system(), etc. */
876 30204 : if (MyProc->pid != (int) getpid())
877 0 : elog(PANIC, "ProcKill() called in child process");
878 :
879 : /* Make sure we're out of the sync rep lists */
880 30204 : SyncRepCleanupAtProcExit();
881 :
882 : #ifdef USE_ASSERT_CHECKING
883 : {
884 : int i;
885 :
886 : /* Last process should have released all locks. */
887 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
888 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
889 : }
890 : #endif
891 :
892 : /*
893 : * Release any LW locks I am holding. There really shouldn't be any, but
894 : * it's cheap to check again before we cut the knees off the LWLock
895 : * facility by releasing our PGPROC ...
896 : */
897 30204 : LWLockReleaseAll();
898 :
899 : /* Cancel any pending condition variable sleep, too */
900 30204 : ConditionVariableCancelSleep();
901 :
902 : /*
903 : * Detach from any lock group of which we are a member. If the leader
904 : * exits before all other group members, its PGPROC will remain allocated
905 : * until the last group process exits; that process must return the
906 : * leader's PGPROC to the appropriate list.
907 : */
908 30204 : if (MyProc->lockGroupLeader != NULL)
909 : {
910 2850 : PGPROC *leader = MyProc->lockGroupLeader;
911 2850 : LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
912 :
913 2850 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
914 : Assert(!dlist_is_empty(&leader->lockGroupMembers));
915 2850 : dlist_delete(&MyProc->lockGroupLink);
916 2850 : if (dlist_is_empty(&leader->lockGroupMembers))
917 : {
918 138 : leader->lockGroupLeader = NULL;
919 138 : if (leader != MyProc)
920 : {
921 0 : procgloballist = leader->procgloballist;
922 :
923 : /* Leader exited first; return its PGPROC. */
924 0 : SpinLockAcquire(ProcStructLock);
925 0 : dlist_push_head(procgloballist, &leader->links);
926 0 : SpinLockRelease(ProcStructLock);
927 : }
928 : }
929 2712 : else if (leader != MyProc)
930 2712 : MyProc->lockGroupLeader = NULL;
931 2850 : LWLockRelease(leader_lwlock);
932 : }
933 :
934 : /*
935 : * Reset MyLatch to the process local one. This is so that signal
936 : * handlers et al can continue using the latch after the shared latch
937 : * isn't ours anymore.
938 : *
939 : * Similarly, stop reporting wait events to MyProc->wait_event_info.
940 : *
941 : * After that clear MyProc and disown the shared latch.
942 : */
943 30204 : SwitchBackToLocalLatch();
944 30204 : pgstat_reset_wait_event_storage();
945 :
946 30204 : proc = MyProc;
947 30204 : MyProc = NULL;
948 30204 : MyProcNumber = INVALID_PROC_NUMBER;
949 30204 : DisownLatch(&proc->procLatch);
950 :
951 : /* Mark the proc no longer in use */
952 30204 : proc->pid = 0;
953 30204 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
954 30204 : proc->vxid.lxid = InvalidTransactionId;
955 :
956 30204 : procgloballist = proc->procgloballist;
957 30204 : SpinLockAcquire(ProcStructLock);
958 :
959 : /*
960 : * If we're still a member of a locking group, that means we're a leader
961 : * which has somehow exited before its children. The last remaining child
962 : * will release our PGPROC. Otherwise, release it now.
963 : */
964 30204 : if (proc->lockGroupLeader == NULL)
965 : {
966 : /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
967 : Assert(dlist_is_empty(&proc->lockGroupMembers));
968 :
969 : /* Return PGPROC structure (and semaphore) to appropriate freelist */
970 30204 : dlist_push_tail(procgloballist, &proc->links);
971 : }
972 :
973 : /* Update shared estimate of spins_per_delay */
974 30204 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
975 :
976 30204 : SpinLockRelease(ProcStructLock);
977 :
978 : /* wake autovac launcher if needed -- see comments in FreeWorkerInfo */
979 30204 : if (AutovacuumLauncherPid != 0)
980 2264 : kill(AutovacuumLauncherPid, SIGUSR2);
981 30204 : }
982 :
983 : /*
984 : * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
985 : * processes (bgwriter, etc). The PGPROC and sema are not released, only
986 : * marked as not-in-use.
987 : */
988 : static void
989 4496 : AuxiliaryProcKill(int code, Datum arg)
990 : {
991 4496 : int proctype = DatumGetInt32(arg);
992 : PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
993 : PGPROC *proc;
994 :
995 : Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
996 :
997 : /* not safe if forked by system(), etc. */
998 4496 : if (MyProc->pid != (int) getpid())
999 0 : elog(PANIC, "AuxiliaryProcKill() called in child process");
1000 :
1001 4496 : auxproc = &AuxiliaryProcs[proctype];
1002 :
1003 : Assert(MyProc == auxproc);
1004 :
1005 : /* Release any LW locks I am holding (see notes above) */
1006 4496 : LWLockReleaseAll();
1007 :
1008 : /* Cancel any pending condition variable sleep, too */
1009 4496 : ConditionVariableCancelSleep();
1010 :
1011 : /* look at the equivalent ProcKill() code for comments */
1012 4496 : SwitchBackToLocalLatch();
1013 4496 : pgstat_reset_wait_event_storage();
1014 :
1015 4496 : proc = MyProc;
1016 4496 : MyProc = NULL;
1017 4496 : MyProcNumber = INVALID_PROC_NUMBER;
1018 4496 : DisownLatch(&proc->procLatch);
1019 :
1020 4496 : SpinLockAcquire(ProcStructLock);
1021 :
1022 : /* Mark auxiliary proc no longer in use */
1023 4496 : proc->pid = 0;
1024 4496 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1025 4496 : proc->vxid.lxid = InvalidTransactionId;
1026 :
1027 : /* Update shared estimate of spins_per_delay */
1028 4496 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1029 :
1030 4496 : SpinLockRelease(ProcStructLock);
1031 4496 : }
1032 :
1033 : /*
1034 : * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
1035 : * given its PID
1036 : *
1037 : * Returns NULL if not found.
1038 : */
1039 : PGPROC *
1040 4434 : AuxiliaryPidGetProc(int pid)
1041 : {
1042 4434 : PGPROC *result = NULL;
1043 : int index;
1044 :
1045 4434 : if (pid == 0) /* never match dummy PGPROCs */
1046 0 : return NULL;
1047 :
1048 9054 : for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
1049 : {
1050 9054 : PGPROC *proc = &AuxiliaryProcs[index];
1051 :
1052 9054 : if (proc->pid == pid)
1053 : {
1054 4434 : result = proc;
1055 4434 : break;
1056 : }
1057 : }
1058 4434 : return result;
1059 : }
1060 :
1061 :
1062 : /*
1063 : * JoinWaitQueue -- join the wait queue on the specified lock
1064 : *
1065 : * It's not actually guaranteed that we need to wait when this function is
1066 : * called, because it could be that when we try to find a position at which
1067 : * to insert ourself into the wait queue, we discover that we must be inserted
1068 : * ahead of everyone who wants a lock that conflict with ours. In that case,
1069 : * we get the lock immediately. Because of this, it's sensible for this function
1070 : * to have a dontWait argument, despite the name.
1071 : *
1072 : * On entry, the caller has already set up LOCK and PROCLOCK entries to
1073 : * reflect that we have "requested" the lock. The caller is responsible for
1074 : * cleaning that up, if we end up not joining the queue after all.
1075 : *
1076 : * The lock table's partition lock must be held at entry, and is still held
1077 : * at exit. The caller must release it before calling ProcSleep().
1078 : *
1079 : * Result is one of the following:
1080 : *
1081 : * PROC_WAIT_STATUS_OK - lock was immediately granted
1082 : * PROC_WAIT_STATUS_WAITING - joined the wait queue; call ProcSleep()
1083 : * PROC_WAIT_STATUS_ERROR - immediate deadlock was detected, or would
1084 : * need to wait and dontWait == true
1085 : *
1086 : * NOTES: The process queue is now a priority queue for locking.
1087 : */
1088 : ProcWaitStatus
1089 3830 : JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
1090 : {
1091 3830 : LOCKMODE lockmode = locallock->tag.mode;
1092 3830 : LOCK *lock = locallock->lock;
1093 3830 : PROCLOCK *proclock = locallock->proclock;
1094 3830 : uint32 hashcode = locallock->hashcode;
1095 3830 : LWLock *partitionLock PG_USED_FOR_ASSERTS_ONLY = LockHashPartitionLock(hashcode);
1096 3830 : dclist_head *waitQueue = &lock->waitProcs;
1097 3830 : PGPROC *insert_before = NULL;
1098 : LOCKMASK myProcHeldLocks;
1099 : LOCKMASK myHeldLocks;
1100 3830 : bool early_deadlock = false;
1101 3830 : PGPROC *leader = MyProc->lockGroupLeader;
1102 :
1103 : Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
1104 :
1105 : /*
1106 : * Set bitmask of locks this process already holds on this object.
1107 : */
1108 3830 : myHeldLocks = MyProc->heldLocks = proclock->holdMask;
1109 :
1110 : /*
1111 : * Determine which locks we're already holding.
1112 : *
1113 : * If group locking is in use, locks held by members of my locking group
1114 : * need to be included in myHeldLocks. This is not required for relation
1115 : * extension lock which conflict among group members. However, including
1116 : * them in myHeldLocks will give group members the priority to get those
1117 : * locks as compared to other backends which are also trying to acquire
1118 : * those locks. OTOH, we can avoid giving priority to group members for
1119 : * that kind of locks, but there doesn't appear to be a clear advantage of
1120 : * the same.
1121 : */
1122 3830 : myProcHeldLocks = proclock->holdMask;
1123 3830 : myHeldLocks = myProcHeldLocks;
1124 3830 : if (leader != NULL)
1125 : {
1126 : dlist_iter iter;
1127 :
1128 66 : dlist_foreach(iter, &lock->procLocks)
1129 : {
1130 : PROCLOCK *otherproclock;
1131 :
1132 50 : otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
1133 :
1134 50 : if (otherproclock->groupLeader == leader)
1135 24 : myHeldLocks |= otherproclock->holdMask;
1136 : }
1137 : }
1138 :
1139 : /*
1140 : * Determine where to add myself in the wait queue.
1141 : *
1142 : * Normally I should go at the end of the queue. However, if I already
1143 : * hold locks that conflict with the request of any previous waiter, put
1144 : * myself in the queue just in front of the first such waiter. This is not
1145 : * a necessary step, since deadlock detection would move me to before that
1146 : * waiter anyway; but it's relatively cheap to detect such a conflict
1147 : * immediately, and avoid delaying till deadlock timeout.
1148 : *
1149 : * Special case: if I find I should go in front of some waiter, check to
1150 : * see if I conflict with already-held locks or the requests before that
1151 : * waiter. If not, then just grant myself the requested lock immediately.
1152 : * This is the same as the test for immediate grant in LockAcquire, except
1153 : * we are only considering the part of the wait queue before my insertion
1154 : * point.
1155 : */
1156 3830 : if (myHeldLocks != 0 && !dclist_is_empty(waitQueue))
1157 : {
1158 12 : LOCKMASK aheadRequests = 0;
1159 : dlist_iter iter;
1160 :
1161 12 : dclist_foreach(iter, waitQueue)
1162 : {
1163 12 : PGPROC *proc = dlist_container(PGPROC, links, iter.cur);
1164 :
1165 : /*
1166 : * If we're part of the same locking group as this waiter, its
1167 : * locks neither conflict with ours nor contribute to
1168 : * aheadRequests.
1169 : */
1170 12 : if (leader != NULL && leader == proc->lockGroupLeader)
1171 0 : continue;
1172 :
1173 : /* Must he wait for me? */
1174 12 : if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1175 : {
1176 : /* Must I wait for him ? */
1177 12 : if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1178 : {
1179 : /*
1180 : * Yes, so we have a deadlock. Easiest way to clean up
1181 : * correctly is to call RemoveFromWaitQueue(), but we
1182 : * can't do that until we are *on* the wait queue. So, set
1183 : * a flag to check below, and break out of loop. Also,
1184 : * record deadlock info for later message.
1185 : */
1186 2 : RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1187 2 : early_deadlock = true;
1188 2 : break;
1189 : }
1190 : /* I must go before this waiter. Check special case. */
1191 10 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1192 10 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1193 : proclock))
1194 : {
1195 : /* Skip the wait and just grant myself the lock. */
1196 10 : GrantLock(lock, proclock, lockmode);
1197 10 : return PROC_WAIT_STATUS_OK;
1198 : }
1199 :
1200 : /* Put myself into wait queue before conflicting process */
1201 0 : insert_before = proc;
1202 0 : break;
1203 : }
1204 : /* Nope, so advance to next waiter */
1205 0 : aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1206 : }
1207 : }
1208 :
1209 : /*
1210 : * If we detected deadlock, give up without waiting. This must agree with
1211 : * CheckDeadLock's recovery code.
1212 : */
1213 3820 : if (early_deadlock)
1214 2 : return PROC_WAIT_STATUS_ERROR;
1215 :
1216 : /*
1217 : * At this point we know that we'd really need to sleep. If we've been
1218 : * commanded not to do that, bail out.
1219 : */
1220 3818 : if (dontWait)
1221 1288 : return PROC_WAIT_STATUS_ERROR;
1222 :
1223 : /*
1224 : * Insert self into queue, at the position determined above.
1225 : */
1226 2530 : if (insert_before)
1227 0 : dclist_insert_before(waitQueue, &insert_before->links, &MyProc->links);
1228 : else
1229 2530 : dclist_push_tail(waitQueue, &MyProc->links);
1230 :
1231 2530 : lock->waitMask |= LOCKBIT_ON(lockmode);
1232 :
1233 : /* Set up wait information in PGPROC object, too */
1234 2530 : MyProc->heldLocks = myProcHeldLocks;
1235 2530 : MyProc->waitLock = lock;
1236 2530 : MyProc->waitProcLock = proclock;
1237 2530 : MyProc->waitLockMode = lockmode;
1238 :
1239 2530 : MyProc->waitStatus = PROC_WAIT_STATUS_WAITING;
1240 :
1241 2530 : return PROC_WAIT_STATUS_WAITING;
1242 : }
1243 :
1244 : /*
1245 : * ProcSleep -- put process to sleep waiting on lock
1246 : *
1247 : * This must be called when JoinWaitQueue() returns PROC_WAIT_STATUS_WAITING.
1248 : * Returns after the lock has been granted, or if a deadlock is detected. Can
1249 : * also bail out with ereport(ERROR), if some other error condition, or a
1250 : * timeout or cancellation is triggered.
1251 : *
1252 : * Result is one of the following:
1253 : *
1254 : * PROC_WAIT_STATUS_OK - lock was granted
1255 : * PROC_WAIT_STATUS_ERROR - a deadlock was detected
1256 : */
1257 : ProcWaitStatus
1258 2530 : ProcSleep(LOCALLOCK *locallock)
1259 : {
1260 2530 : LOCKMODE lockmode = locallock->tag.mode;
1261 2530 : LOCK *lock = locallock->lock;
1262 2530 : uint32 hashcode = locallock->hashcode;
1263 2530 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1264 2530 : TimestampTz standbyWaitStart = 0;
1265 2530 : bool allow_autovacuum_cancel = true;
1266 2530 : bool logged_recovery_conflict = false;
1267 : ProcWaitStatus myWaitStatus;
1268 :
1269 : /* The caller must've armed the on-error cleanup mechanism */
1270 : Assert(GetAwaitedLock() == locallock);
1271 : Assert(!LWLockHeldByMe(partitionLock));
1272 :
1273 : /*
1274 : * Now that we will successfully clean up after an ereport, it's safe to
1275 : * check to see if there's a buffer pin deadlock against the Startup
1276 : * process. Of course, that's only necessary if we're doing Hot Standby
1277 : * and are not the Startup process ourselves.
1278 : */
1279 2530 : if (RecoveryInProgress() && !InRecovery)
1280 2 : CheckRecoveryConflictDeadlock();
1281 :
1282 : /* Reset deadlock_state before enabling the timeout handler */
1283 2530 : deadlock_state = DS_NOT_YET_CHECKED;
1284 2530 : got_deadlock_timeout = false;
1285 :
1286 : /*
1287 : * Set timer so we can wake up after awhile and check for a deadlock. If a
1288 : * deadlock is detected, the handler sets MyProc->waitStatus =
1289 : * PROC_WAIT_STATUS_ERROR, allowing us to know that we must report failure
1290 : * rather than success.
1291 : *
1292 : * By delaying the check until we've waited for a bit, we can avoid
1293 : * running the rather expensive deadlock-check code in most cases.
1294 : *
1295 : * If LockTimeout is set, also enable the timeout for that. We can save a
1296 : * few cycles by enabling both timeout sources in one call.
1297 : *
1298 : * If InHotStandby we set lock waits slightly later for clarity with other
1299 : * code.
1300 : */
1301 2530 : if (!InHotStandby)
1302 : {
1303 2526 : if (LockTimeout > 0)
1304 : {
1305 : EnableTimeoutParams timeouts[2];
1306 :
1307 222 : timeouts[0].id = DEADLOCK_TIMEOUT;
1308 222 : timeouts[0].type = TMPARAM_AFTER;
1309 222 : timeouts[0].delay_ms = DeadlockTimeout;
1310 222 : timeouts[1].id = LOCK_TIMEOUT;
1311 222 : timeouts[1].type = TMPARAM_AFTER;
1312 222 : timeouts[1].delay_ms = LockTimeout;
1313 222 : enable_timeouts(timeouts, 2);
1314 : }
1315 : else
1316 2304 : enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1317 :
1318 : /*
1319 : * Use the current time obtained for the deadlock timeout timer as
1320 : * waitStart (i.e., the time when this process started waiting for the
1321 : * lock). Since getting the current time newly can cause overhead, we
1322 : * reuse the already-obtained time to avoid that overhead.
1323 : *
1324 : * Note that waitStart is updated without holding the lock table's
1325 : * partition lock, to avoid the overhead by additional lock
1326 : * acquisition. This can cause "waitstart" in pg_locks to become NULL
1327 : * for a very short period of time after the wait started even though
1328 : * "granted" is false. This is OK in practice because we can assume
1329 : * that users are likely to look at "waitstart" when waiting for the
1330 : * lock for a long time.
1331 : */
1332 2526 : pg_atomic_write_u64(&MyProc->waitStart,
1333 2526 : get_timeout_start_time(DEADLOCK_TIMEOUT));
1334 : }
1335 4 : else if (log_recovery_conflict_waits)
1336 : {
1337 : /*
1338 : * Set the wait start timestamp if logging is enabled and in hot
1339 : * standby.
1340 : */
1341 2 : standbyWaitStart = GetCurrentTimestamp();
1342 : }
1343 :
1344 : /*
1345 : * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1346 : * will not wait. But a set latch does not necessarily mean that the lock
1347 : * is free now, as there are many other sources for latch sets than
1348 : * somebody releasing the lock.
1349 : *
1350 : * We process interrupts whenever the latch has been set, so cancel/die
1351 : * interrupts are processed quickly. This means we must not mind losing
1352 : * control to a cancel/die interrupt here. We don't, because we have no
1353 : * shared-state-change work to do after being granted the lock (the
1354 : * grantor did it all). We do have to worry about canceling the deadlock
1355 : * timeout and updating the locallock table, but if we lose control to an
1356 : * error, LockErrorCleanup will fix that up.
1357 : */
1358 : do
1359 : {
1360 4796 : if (InHotStandby)
1361 : {
1362 12 : bool maybe_log_conflict =
1363 12 : (standbyWaitStart != 0 && !logged_recovery_conflict);
1364 :
1365 : /* Set a timer and wait for that or for the lock to be granted */
1366 12 : ResolveRecoveryConflictWithLock(locallock->tag.lock,
1367 : maybe_log_conflict);
1368 :
1369 : /*
1370 : * Emit the log message if the startup process is waiting longer
1371 : * than deadlock_timeout for recovery conflict on lock.
1372 : */
1373 12 : if (maybe_log_conflict)
1374 : {
1375 4 : TimestampTz now = GetCurrentTimestamp();
1376 :
1377 4 : if (TimestampDifferenceExceeds(standbyWaitStart, now,
1378 : DeadlockTimeout))
1379 : {
1380 : VirtualTransactionId *vxids;
1381 : int cnt;
1382 :
1383 2 : vxids = GetLockConflicts(&locallock->tag.lock,
1384 : AccessExclusiveLock, &cnt);
1385 :
1386 : /*
1387 : * Log the recovery conflict and the list of PIDs of
1388 : * backends holding the conflicting lock. Note that we do
1389 : * logging even if there are no such backends right now
1390 : * because the startup process here has already waited
1391 : * longer than deadlock_timeout.
1392 : */
1393 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1394 : standbyWaitStart, now,
1395 2 : cnt > 0 ? vxids : NULL, true);
1396 2 : logged_recovery_conflict = true;
1397 : }
1398 : }
1399 : }
1400 : else
1401 : {
1402 4784 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1403 4784 : PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1404 4784 : ResetLatch(MyLatch);
1405 : /* check for deadlocks first, as that's probably log-worthy */
1406 4784 : if (got_deadlock_timeout)
1407 : {
1408 58 : CheckDeadLock();
1409 58 : got_deadlock_timeout = false;
1410 : }
1411 4784 : CHECK_FOR_INTERRUPTS();
1412 : }
1413 :
1414 : /*
1415 : * waitStatus could change from PROC_WAIT_STATUS_WAITING to something
1416 : * else asynchronously. Read it just once per loop to prevent
1417 : * surprising behavior (such as missing log messages).
1418 : */
1419 4714 : myWaitStatus = *((volatile ProcWaitStatus *) &MyProc->waitStatus);
1420 :
1421 : /*
1422 : * If we are not deadlocked, but are waiting on an autovacuum-induced
1423 : * task, send a signal to interrupt it.
1424 : */
1425 4714 : if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1426 : {
1427 0 : PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1428 : uint8 statusFlags;
1429 : uint8 lockmethod_copy;
1430 : LOCKTAG locktag_copy;
1431 :
1432 : /*
1433 : * Grab info we need, then release lock immediately. Note this
1434 : * coding means that there is a tiny chance that the process
1435 : * terminates its current transaction and starts a different one
1436 : * before we have a change to send the signal; the worst possible
1437 : * consequence is that a for-wraparound vacuum is canceled. But
1438 : * that could happen in any case unless we were to do kill() with
1439 : * the lock held, which is much more undesirable.
1440 : */
1441 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1442 0 : statusFlags = ProcGlobal->statusFlags[autovac->pgxactoff];
1443 0 : lockmethod_copy = lock->tag.locktag_lockmethodid;
1444 0 : locktag_copy = lock->tag;
1445 0 : LWLockRelease(ProcArrayLock);
1446 :
1447 : /*
1448 : * Only do it if the worker is not working to protect against Xid
1449 : * wraparound.
1450 : */
1451 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
1452 0 : !(statusFlags & PROC_VACUUM_FOR_WRAPAROUND))
1453 : {
1454 0 : int pid = autovac->pid;
1455 :
1456 : /* report the case, if configured to do so */
1457 0 : if (message_level_is_interesting(DEBUG1))
1458 : {
1459 : StringInfoData locktagbuf;
1460 : StringInfoData logbuf; /* errdetail for server log */
1461 :
1462 0 : initStringInfo(&locktagbuf);
1463 0 : initStringInfo(&logbuf);
1464 0 : DescribeLockTag(&locktagbuf, &locktag_copy);
1465 0 : appendStringInfo(&logbuf,
1466 : "Process %d waits for %s on %s.",
1467 : MyProcPid,
1468 : GetLockmodeName(lockmethod_copy, lockmode),
1469 : locktagbuf.data);
1470 :
1471 0 : ereport(DEBUG1,
1472 : (errmsg_internal("sending cancel to blocking autovacuum PID %d",
1473 : pid),
1474 : errdetail_log("%s", logbuf.data)));
1475 :
1476 0 : pfree(locktagbuf.data);
1477 0 : pfree(logbuf.data);
1478 : }
1479 :
1480 : /* send the autovacuum worker Back to Old Kent Road */
1481 0 : if (kill(pid, SIGINT) < 0)
1482 : {
1483 : /*
1484 : * There's a race condition here: once we release the
1485 : * ProcArrayLock, it's possible for the autovac worker to
1486 : * close up shop and exit before we can do the kill().
1487 : * Therefore, we do not whinge about no-such-process.
1488 : * Other errors such as EPERM could conceivably happen if
1489 : * the kernel recycles the PID fast enough, but such cases
1490 : * seem improbable enough that it's probably best to issue
1491 : * a warning if we see some other errno.
1492 : */
1493 0 : if (errno != ESRCH)
1494 0 : ereport(WARNING,
1495 : (errmsg("could not send signal to process %d: %m",
1496 : pid)));
1497 : }
1498 : }
1499 :
1500 : /* prevent signal from being sent again more than once */
1501 0 : allow_autovacuum_cancel = false;
1502 : }
1503 :
1504 : /*
1505 : * If awoken after the deadlock check interrupt has run, and
1506 : * log_lock_waits is on, then report about the wait.
1507 : */
1508 4714 : if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1509 : {
1510 : StringInfoData buf,
1511 : lock_waiters_sbuf,
1512 : lock_holders_sbuf;
1513 : const char *modename;
1514 : long secs;
1515 : int usecs;
1516 : long msecs;
1517 : dlist_iter proc_iter;
1518 : PROCLOCK *curproclock;
1519 44 : bool first_holder = true,
1520 44 : first_waiter = true;
1521 44 : int lockHoldersNum = 0;
1522 :
1523 44 : initStringInfo(&buf);
1524 44 : initStringInfo(&lock_waiters_sbuf);
1525 44 : initStringInfo(&lock_holders_sbuf);
1526 :
1527 44 : DescribeLockTag(&buf, &locallock->tag.lock);
1528 44 : modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1529 : lockmode);
1530 44 : TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1531 : GetCurrentTimestamp(),
1532 : &secs, &usecs);
1533 44 : msecs = secs * 1000 + usecs / 1000;
1534 44 : usecs = usecs % 1000;
1535 :
1536 : /*
1537 : * we loop over the lock's procLocks to gather a list of all
1538 : * holders and waiters. Thus we will be able to provide more
1539 : * detailed information for lock debugging purposes.
1540 : *
1541 : * lock->procLocks contains all processes which hold or wait for
1542 : * this lock.
1543 : */
1544 :
1545 44 : LWLockAcquire(partitionLock, LW_SHARED);
1546 :
1547 140 : dlist_foreach(proc_iter, &lock->procLocks)
1548 : {
1549 96 : curproclock =
1550 96 : dlist_container(PROCLOCK, lockLink, proc_iter.cur);
1551 :
1552 : /*
1553 : * we are a waiter if myProc->waitProcLock == curproclock; we
1554 : * are a holder if it is NULL or something different
1555 : */
1556 96 : if (curproclock->tag.myProc->waitProcLock == curproclock)
1557 : {
1558 42 : if (first_waiter)
1559 : {
1560 24 : appendStringInfo(&lock_waiters_sbuf, "%d",
1561 24 : curproclock->tag.myProc->pid);
1562 24 : first_waiter = false;
1563 : }
1564 : else
1565 18 : appendStringInfo(&lock_waiters_sbuf, ", %d",
1566 18 : curproclock->tag.myProc->pid);
1567 : }
1568 : else
1569 : {
1570 54 : if (first_holder)
1571 : {
1572 44 : appendStringInfo(&lock_holders_sbuf, "%d",
1573 44 : curproclock->tag.myProc->pid);
1574 44 : first_holder = false;
1575 : }
1576 : else
1577 10 : appendStringInfo(&lock_holders_sbuf, ", %d",
1578 10 : curproclock->tag.myProc->pid);
1579 :
1580 54 : lockHoldersNum++;
1581 : }
1582 : }
1583 :
1584 44 : LWLockRelease(partitionLock);
1585 :
1586 44 : if (deadlock_state == DS_SOFT_DEADLOCK)
1587 6 : ereport(LOG,
1588 : (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1589 : MyProcPid, modename, buf.data, msecs, usecs),
1590 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1591 : "Processes holding the lock: %s. Wait queue: %s.",
1592 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1593 38 : else if (deadlock_state == DS_HARD_DEADLOCK)
1594 : {
1595 : /*
1596 : * This message is a bit redundant with the error that will be
1597 : * reported subsequently, but in some cases the error report
1598 : * might not make it to the log (eg, if it's caught by an
1599 : * exception handler), and we want to ensure all long-wait
1600 : * events get logged.
1601 : */
1602 4 : ereport(LOG,
1603 : (errmsg("process %d detected deadlock while waiting for %s on %s after %ld.%03d ms",
1604 : MyProcPid, modename, buf.data, msecs, usecs),
1605 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1606 : "Processes holding the lock: %s. Wait queue: %s.",
1607 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1608 : }
1609 :
1610 44 : if (myWaitStatus == PROC_WAIT_STATUS_WAITING)
1611 22 : ereport(LOG,
1612 : (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1613 : MyProcPid, modename, buf.data, msecs, usecs),
1614 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1615 : "Processes holding the lock: %s. Wait queue: %s.",
1616 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1617 22 : else if (myWaitStatus == PROC_WAIT_STATUS_OK)
1618 18 : ereport(LOG,
1619 : (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1620 : MyProcPid, modename, buf.data, msecs, usecs)));
1621 : else
1622 : {
1623 : Assert(myWaitStatus == PROC_WAIT_STATUS_ERROR);
1624 :
1625 : /*
1626 : * Currently, the deadlock checker always kicks its own
1627 : * process, which means that we'll only see
1628 : * PROC_WAIT_STATUS_ERROR when deadlock_state ==
1629 : * DS_HARD_DEADLOCK, and there's no need to print redundant
1630 : * messages. But for completeness and future-proofing, print
1631 : * a message if it looks like someone else kicked us off the
1632 : * lock.
1633 : */
1634 4 : if (deadlock_state != DS_HARD_DEADLOCK)
1635 0 : ereport(LOG,
1636 : (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1637 : MyProcPid, modename, buf.data, msecs, usecs),
1638 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1639 : "Processes holding the lock: %s. Wait queue: %s.",
1640 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1641 : }
1642 :
1643 : /*
1644 : * At this point we might still need to wait for the lock. Reset
1645 : * state so we don't print the above messages again.
1646 : */
1647 44 : deadlock_state = DS_NO_DEADLOCK;
1648 :
1649 44 : pfree(buf.data);
1650 44 : pfree(lock_holders_sbuf.data);
1651 44 : pfree(lock_waiters_sbuf.data);
1652 : }
1653 4714 : } while (myWaitStatus == PROC_WAIT_STATUS_WAITING);
1654 :
1655 : /*
1656 : * Disable the timers, if they are still running. As in LockErrorCleanup,
1657 : * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1658 : * already caused QueryCancelPending to become set, we want the cancel to
1659 : * be reported as a lock timeout, not a user cancel.
1660 : */
1661 2448 : if (!InHotStandby)
1662 : {
1663 2444 : if (LockTimeout > 0)
1664 : {
1665 : DisableTimeoutParams timeouts[2];
1666 :
1667 210 : timeouts[0].id = DEADLOCK_TIMEOUT;
1668 210 : timeouts[0].keep_indicator = false;
1669 210 : timeouts[1].id = LOCK_TIMEOUT;
1670 210 : timeouts[1].keep_indicator = true;
1671 210 : disable_timeouts(timeouts, 2);
1672 : }
1673 : else
1674 2234 : disable_timeout(DEADLOCK_TIMEOUT, false);
1675 : }
1676 :
1677 : /*
1678 : * Emit the log message if recovery conflict on lock was resolved but the
1679 : * startup process waited longer than deadlock_timeout for it.
1680 : */
1681 2448 : if (InHotStandby && logged_recovery_conflict)
1682 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1683 : standbyWaitStart, GetCurrentTimestamp(),
1684 : NULL, false);
1685 :
1686 : /*
1687 : * We don't have to do anything else, because the awaker did all the
1688 : * necessary updates of the lock table and MyProc. (The caller is
1689 : * responsible for updating the local lock table.)
1690 : */
1691 2448 : return myWaitStatus;
1692 : }
1693 :
1694 :
1695 : /*
1696 : * ProcWakeup -- wake up a process by setting its latch.
1697 : *
1698 : * Also remove the process from the wait queue and set its links invalid.
1699 : *
1700 : * The appropriate lock partition lock must be held by caller.
1701 : *
1702 : * XXX: presently, this code is only used for the "success" case, and only
1703 : * works correctly for that case. To clean up in failure case, would need
1704 : * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1705 : * Hence, in practice the waitStatus parameter must be PROC_WAIT_STATUS_OK.
1706 : */
1707 : void
1708 2466 : ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus)
1709 : {
1710 2466 : if (dlist_node_is_detached(&proc->links))
1711 0 : return;
1712 :
1713 : Assert(proc->waitStatus == PROC_WAIT_STATUS_WAITING);
1714 :
1715 : /* Remove process from wait queue */
1716 2466 : dclist_delete_from_thoroughly(&proc->waitLock->waitProcs, &proc->links);
1717 :
1718 : /* Clean up process' state and pass it the ok/fail signal */
1719 2466 : proc->waitLock = NULL;
1720 2466 : proc->waitProcLock = NULL;
1721 2466 : proc->waitStatus = waitStatus;
1722 2466 : pg_atomic_write_u64(&MyProc->waitStart, 0);
1723 :
1724 : /* And awaken it */
1725 2466 : SetLatch(&proc->procLatch);
1726 : }
1727 :
1728 : /*
1729 : * ProcLockWakeup -- routine for waking up processes when a lock is
1730 : * released (or a prior waiter is aborted). Scan all waiters
1731 : * for lock, waken any that are no longer blocked.
1732 : *
1733 : * The appropriate lock partition lock must be held by caller.
1734 : */
1735 : void
1736 2514 : ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1737 : {
1738 2514 : dclist_head *waitQueue = &lock->waitProcs;
1739 2514 : LOCKMASK aheadRequests = 0;
1740 : dlist_mutable_iter miter;
1741 :
1742 2514 : if (dclist_is_empty(waitQueue))
1743 90 : return;
1744 :
1745 5730 : dclist_foreach_modify(miter, waitQueue)
1746 : {
1747 3306 : PGPROC *proc = dlist_container(PGPROC, links, miter.cur);
1748 3306 : LOCKMODE lockmode = proc->waitLockMode;
1749 :
1750 : /*
1751 : * Waken if (a) doesn't conflict with requests of earlier waiters, and
1752 : * (b) doesn't conflict with already-held locks.
1753 : */
1754 3306 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1755 2900 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1756 : proc->waitProcLock))
1757 : {
1758 : /* OK to waken */
1759 2466 : GrantLock(lock, proc->waitProcLock, lockmode);
1760 : /* removes proc from the lock's waiting process queue */
1761 2466 : ProcWakeup(proc, PROC_WAIT_STATUS_OK);
1762 : }
1763 : else
1764 : {
1765 : /*
1766 : * Lock conflicts: Don't wake, but remember requested mode for
1767 : * later checks.
1768 : */
1769 840 : aheadRequests |= LOCKBIT_ON(lockmode);
1770 : }
1771 : }
1772 : }
1773 :
1774 : /*
1775 : * CheckDeadLock
1776 : *
1777 : * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1778 : * lock to be released by some other process. Check if there's a deadlock; if
1779 : * not, just return. (But signal ProcSleep to log a message, if
1780 : * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1781 : * the lock's wait queue and signal an error to ProcSleep.
1782 : */
1783 : static void
1784 58 : CheckDeadLock(void)
1785 : {
1786 : int i;
1787 :
1788 : /*
1789 : * Acquire exclusive lock on the entire shared lock data structures. Must
1790 : * grab LWLocks in partition-number order to avoid LWLock deadlock.
1791 : *
1792 : * Note that the deadlock check interrupt had better not be enabled
1793 : * anywhere that this process itself holds lock partition locks, else this
1794 : * will wait forever. Also note that LWLockAcquire creates a critical
1795 : * section, so that this routine cannot be interrupted by cancel/die
1796 : * interrupts.
1797 : */
1798 986 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1799 928 : LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1800 :
1801 : /*
1802 : * Check to see if we've been awoken by anyone in the interim.
1803 : *
1804 : * If we have, we can return and resume our transaction -- happy day.
1805 : * Before we are awoken the process releasing the lock grants it to us so
1806 : * we know that we don't have to wait anymore.
1807 : *
1808 : * We check by looking to see if we've been unlinked from the wait queue.
1809 : * This is safe because we hold the lock partition lock.
1810 : */
1811 58 : if (MyProc->links.prev == NULL ||
1812 58 : MyProc->links.next == NULL)
1813 0 : goto check_done;
1814 :
1815 : #ifdef LOCK_DEBUG
1816 : if (Debug_deadlocks)
1817 : DumpAllLocks();
1818 : #endif
1819 :
1820 : /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1821 58 : deadlock_state = DeadLockCheck(MyProc);
1822 :
1823 58 : if (deadlock_state == DS_HARD_DEADLOCK)
1824 : {
1825 : /*
1826 : * Oops. We have a deadlock.
1827 : *
1828 : * Get this process out of wait state. (Note: we could do this more
1829 : * efficiently by relying on lockAwaited, but use this coding to
1830 : * preserve the flexibility to kill some other transaction than the
1831 : * one detecting the deadlock.)
1832 : *
1833 : * RemoveFromWaitQueue sets MyProc->waitStatus to
1834 : * PROC_WAIT_STATUS_ERROR, so ProcSleep will report an error after we
1835 : * return from the signal handler.
1836 : */
1837 : Assert(MyProc->waitLock != NULL);
1838 10 : RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1839 :
1840 : /*
1841 : * We're done here. Transaction abort caused by the error that
1842 : * ProcSleep will raise will cause any other locks we hold to be
1843 : * released, thus allowing other processes to wake up; we don't need
1844 : * to do that here. NOTE: an exception is that releasing locks we
1845 : * hold doesn't consider the possibility of waiters that were blocked
1846 : * behind us on the lock we just failed to get, and might now be
1847 : * wakable because we're not in front of them anymore. However,
1848 : * RemoveFromWaitQueue took care of waking up any such processes.
1849 : */
1850 : }
1851 :
1852 : /*
1853 : * And release locks. We do this in reverse order for two reasons: (1)
1854 : * Anyone else who needs more than one of the locks will be trying to lock
1855 : * them in increasing order; we don't want to release the other process
1856 : * until it can get all the locks it needs. (2) This avoids O(N^2)
1857 : * behavior inside LWLockRelease.
1858 : */
1859 48 : check_done:
1860 986 : for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1861 928 : LWLockRelease(LockHashPartitionLockByIndex(i));
1862 58 : }
1863 :
1864 : /*
1865 : * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1866 : *
1867 : * NB: Runs inside a signal handler, be careful.
1868 : */
1869 : void
1870 58 : CheckDeadLockAlert(void)
1871 : {
1872 58 : int save_errno = errno;
1873 :
1874 58 : got_deadlock_timeout = true;
1875 :
1876 : /*
1877 : * Have to set the latch again, even if handle_sig_alarm already did. Back
1878 : * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1879 : * ever would be a problem, but setting a set latch again is cheap.
1880 : *
1881 : * Note that, when this function runs inside procsignal_sigusr1_handler(),
1882 : * the handler function sets the latch again after the latch is set here.
1883 : */
1884 58 : SetLatch(MyLatch);
1885 58 : errno = save_errno;
1886 58 : }
1887 :
1888 : /*
1889 : * ProcWaitForSignal - wait for a signal from another backend.
1890 : *
1891 : * As this uses the generic process latch the caller has to be robust against
1892 : * unrelated wakeups: Always check that the desired state has occurred, and
1893 : * wait again if not.
1894 : */
1895 : void
1896 38 : ProcWaitForSignal(uint32 wait_event_info)
1897 : {
1898 38 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1899 : wait_event_info);
1900 38 : ResetLatch(MyLatch);
1901 38 : CHECK_FOR_INTERRUPTS();
1902 38 : }
1903 :
1904 : /*
1905 : * ProcSendSignal - set the latch of a backend identified by ProcNumber
1906 : */
1907 : void
1908 8 : ProcSendSignal(ProcNumber procNumber)
1909 : {
1910 8 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
1911 0 : elog(ERROR, "procNumber out of range");
1912 :
1913 8 : SetLatch(&ProcGlobal->allProcs[procNumber].procLatch);
1914 8 : }
1915 :
1916 : /*
1917 : * BecomeLockGroupLeader - designate process as lock group leader
1918 : *
1919 : * Once this function has returned, other processes can join the lock group
1920 : * by calling BecomeLockGroupMember.
1921 : */
1922 : void
1923 1150 : BecomeLockGroupLeader(void)
1924 : {
1925 : LWLock *leader_lwlock;
1926 :
1927 : /* If we already did it, we don't need to do it again. */
1928 1150 : if (MyProc->lockGroupLeader == MyProc)
1929 1012 : return;
1930 :
1931 : /* We had better not be a follower. */
1932 : Assert(MyProc->lockGroupLeader == NULL);
1933 :
1934 : /* Create single-member group, containing only ourselves. */
1935 138 : leader_lwlock = LockHashPartitionLockByProc(MyProc);
1936 138 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1937 138 : MyProc->lockGroupLeader = MyProc;
1938 138 : dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
1939 138 : LWLockRelease(leader_lwlock);
1940 : }
1941 :
1942 : /*
1943 : * BecomeLockGroupMember - designate process as lock group member
1944 : *
1945 : * This is pretty straightforward except for the possibility that the leader
1946 : * whose group we're trying to join might exit before we manage to do so;
1947 : * and the PGPROC might get recycled for an unrelated process. To avoid
1948 : * that, we require the caller to pass the PID of the intended PGPROC as
1949 : * an interlock. Returns true if we successfully join the intended lock
1950 : * group, and false if not.
1951 : */
1952 : bool
1953 2712 : BecomeLockGroupMember(PGPROC *leader, int pid)
1954 : {
1955 : LWLock *leader_lwlock;
1956 2712 : bool ok = false;
1957 :
1958 : /* Group leader can't become member of group */
1959 : Assert(MyProc != leader);
1960 :
1961 : /* Can't already be a member of a group */
1962 : Assert(MyProc->lockGroupLeader == NULL);
1963 :
1964 : /* PID must be valid. */
1965 : Assert(pid != 0);
1966 :
1967 : /*
1968 : * Get lock protecting the group fields. Note LockHashPartitionLockByProc
1969 : * calculates the proc number based on the PGPROC slot without looking at
1970 : * its contents, so we will acquire the correct lock even if the leader
1971 : * PGPROC is in process of being recycled.
1972 : */
1973 2712 : leader_lwlock = LockHashPartitionLockByProc(leader);
1974 2712 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
1975 :
1976 : /* Is this the leader we're looking for? */
1977 2712 : if (leader->pid == pid && leader->lockGroupLeader == leader)
1978 : {
1979 : /* OK, join the group */
1980 2712 : ok = true;
1981 2712 : MyProc->lockGroupLeader = leader;
1982 2712 : dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
1983 : }
1984 2712 : LWLockRelease(leader_lwlock);
1985 :
1986 2712 : return ok;
1987 : }
|
