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 PGPROC.
95 : */
96 : static Size
97 5994 : PGProcShmemSize(void)
98 : {
99 5994 : Size size = 0;
100 : Size TotalProcs =
101 5994 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
102 :
103 5994 : size = add_size(size, mul_size(TotalProcs, sizeof(PGPROC)));
104 5994 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->xids)));
105 5994 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->subxidStates)));
106 5994 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->statusFlags)));
107 :
108 5994 : return size;
109 : }
110 :
111 : /*
112 : * Report shared-memory space needed by Fast-Path locks.
113 : */
114 : static Size
115 5994 : FastPathLockShmemSize(void)
116 : {
117 5994 : Size size = 0;
118 : Size TotalProcs =
119 5994 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
120 : Size fpLockBitsSize,
121 : fpRelIdSize;
122 :
123 : /*
124 : * Memory needed for PGPROC fast-path lock arrays. Make sure the sizes are
125 : * nicely aligned in each backend.
126 : */
127 5994 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
128 5994 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
129 :
130 5994 : size = add_size(size, mul_size(TotalProcs, (fpLockBitsSize + fpRelIdSize)));
131 :
132 5994 : return size;
133 : }
134 :
135 : /*
136 : * Report shared-memory space needed by InitProcGlobal.
137 : */
138 : Size
139 3898 : ProcGlobalShmemSize(void)
140 : {
141 3898 : Size size = 0;
142 :
143 : /* ProcGlobal */
144 3898 : size = add_size(size, sizeof(PROC_HDR));
145 3898 : size = add_size(size, sizeof(slock_t));
146 :
147 3898 : size = add_size(size, PGProcShmemSize());
148 3898 : size = add_size(size, FastPathLockShmemSize());
149 :
150 3898 : return size;
151 : }
152 :
153 : /*
154 : * Report number of semaphores needed by InitProcGlobal.
155 : */
156 : int
157 3898 : ProcGlobalSemas(void)
158 : {
159 : /*
160 : * We need a sema per backend (including autovacuum), plus one for each
161 : * auxiliary process.
162 : */
163 3898 : return MaxBackends + NUM_AUXILIARY_PROCS;
164 : }
165 :
166 : /*
167 : * InitProcGlobal -
168 : * Initialize the global process table during postmaster or standalone
169 : * backend startup.
170 : *
171 : * We also create all the per-process semaphores we will need to support
172 : * the requested number of backends. We used to allocate semaphores
173 : * only when backends were actually started up, but that is bad because
174 : * it lets Postgres fail under load --- a lot of Unix systems are
175 : * (mis)configured with small limits on the number of semaphores, and
176 : * running out when trying to start another backend is a common failure.
177 : * So, now we grab enough semaphores to support the desired max number
178 : * of backends immediately at initialization --- if the sysadmin has set
179 : * MaxConnections, max_worker_processes, max_wal_senders, or
180 : * autovacuum_worker_slots higher than his kernel will support, he'll
181 : * find out sooner rather than later.
182 : *
183 : * Another reason for creating semaphores here is that the semaphore
184 : * implementation typically requires us to create semaphores in the
185 : * postmaster, not in backends.
186 : *
187 : * Note: this is NOT called by individual backends under a postmaster,
188 : * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
189 : * pointers must be propagated specially for EXEC_BACKEND operation.
190 : */
191 : void
192 2096 : InitProcGlobal(void)
193 : {
194 : PGPROC *procs;
195 : int i,
196 : j;
197 : bool found;
198 2096 : uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
199 :
200 : /* Used for setup of per-backend fast-path slots. */
201 : char *fpPtr,
202 : *fpEndPtr PG_USED_FOR_ASSERTS_ONLY;
203 : Size fpLockBitsSize,
204 : fpRelIdSize;
205 : Size requestSize;
206 : char *ptr;
207 :
208 : /* Create the ProcGlobal shared structure */
209 2096 : ProcGlobal = (PROC_HDR *)
210 2096 : ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
211 : Assert(!found);
212 :
213 : /*
214 : * Initialize the data structures.
215 : */
216 2096 : ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
217 2096 : dlist_init(&ProcGlobal->freeProcs);
218 2096 : dlist_init(&ProcGlobal->autovacFreeProcs);
219 2096 : dlist_init(&ProcGlobal->bgworkerFreeProcs);
220 2096 : dlist_init(&ProcGlobal->walsenderFreeProcs);
221 2096 : ProcGlobal->startupBufferPinWaitBufId = -1;
222 2096 : ProcGlobal->walwriterProc = INVALID_PROC_NUMBER;
223 2096 : ProcGlobal->checkpointerProc = INVALID_PROC_NUMBER;
224 2096 : pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PROC_NUMBER);
225 2096 : pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PROC_NUMBER);
226 :
227 : /*
228 : * Create and initialize all the PGPROC structures we'll need. There are
229 : * six separate consumers: (1) normal backends, (2) autovacuum workers and
230 : * special workers, (3) background workers, (4) walsenders, (5) auxiliary
231 : * processes, and (6) prepared transactions. (For largely-historical
232 : * reasons, we combine autovacuum and special workers into one category
233 : * with a single freelist.) Each PGPROC structure is dedicated to exactly
234 : * one of these purposes, and they do not move between groups.
235 : */
236 2096 : requestSize = PGProcShmemSize();
237 :
238 2096 : ptr = ShmemInitStruct("PGPROC structures",
239 : requestSize,
240 : &found);
241 :
242 2096 : MemSet(ptr, 0, requestSize);
243 :
244 2096 : procs = (PGPROC *) ptr;
245 2096 : ptr = (char *) ptr + TotalProcs * sizeof(PGPROC);
246 :
247 2096 : ProcGlobal->allProcs = procs;
248 : /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
249 2096 : ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
250 :
251 : /*
252 : * Allocate arrays mirroring PGPROC fields in a dense manner. See
253 : * PROC_HDR.
254 : *
255 : * XXX: It might make sense to increase padding for these arrays, given
256 : * how hotly they are accessed.
257 : */
258 2096 : ProcGlobal->xids = (TransactionId *) ptr;
259 2096 : ptr = (char *) ptr + (TotalProcs * sizeof(*ProcGlobal->xids));
260 :
261 2096 : ProcGlobal->subxidStates = (XidCacheStatus *) ptr;
262 2096 : ptr = (char *) ptr + (TotalProcs * sizeof(*ProcGlobal->subxidStates));
263 :
264 2096 : ProcGlobal->statusFlags = (uint8 *) ptr;
265 2096 : ptr = (char *) ptr + (TotalProcs * sizeof(*ProcGlobal->statusFlags));
266 :
267 : /* make sure wer didn't overflow */
268 : Assert((ptr > (char *) procs) && (ptr <= (char *) procs + requestSize));
269 :
270 : /*
271 : * Allocate arrays for fast-path locks. Those are variable-length, so
272 : * can't be included in PGPROC directly. We allocate a separate piece of
273 : * shared memory and then divide that between backends.
274 : */
275 2096 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
276 2096 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
277 :
278 2096 : requestSize = FastPathLockShmemSize();
279 :
280 2096 : fpPtr = ShmemInitStruct("Fast-Path Lock Array",
281 : requestSize,
282 : &found);
283 :
284 2096 : MemSet(fpPtr, 0, requestSize);
285 :
286 : /* For asserts checking we did not overflow. */
287 2096 : fpEndPtr = fpPtr + requestSize;
288 :
289 280528 : for (i = 0; i < TotalProcs; i++)
290 : {
291 278432 : PGPROC *proc = &procs[i];
292 :
293 : /* Common initialization for all PGPROCs, regardless of type. */
294 :
295 : /*
296 : * Set the fast-path lock arrays, and move the pointer. We interleave
297 : * the two arrays, to (hopefully) get some locality for each backend.
298 : */
299 278432 : proc->fpLockBits = (uint64 *) fpPtr;
300 278432 : fpPtr += fpLockBitsSize;
301 :
302 278432 : proc->fpRelId = (Oid *) fpPtr;
303 278432 : fpPtr += fpRelIdSize;
304 :
305 : Assert(fpPtr <= fpEndPtr);
306 :
307 : /*
308 : * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
309 : * dummy PGPROCs don't need these though - they're never associated
310 : * with a real process
311 : */
312 278432 : if (i < MaxBackends + NUM_AUXILIARY_PROCS)
313 : {
314 276750 : proc->sem = PGSemaphoreCreate();
315 276750 : InitSharedLatch(&(proc->procLatch));
316 276750 : LWLockInitialize(&(proc->fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
317 : }
318 :
319 : /*
320 : * Newly created PGPROCs for normal backends, autovacuum workers,
321 : * special workers, bgworkers, and walsenders must be queued up on the
322 : * appropriate free list. Because there can only ever be a small,
323 : * fixed number of auxiliary processes, no free list is used in that
324 : * case; InitAuxiliaryProcess() instead uses a linear search. PGPROCs
325 : * for prepared transactions are added to a free list by
326 : * TwoPhaseShmemInit().
327 : */
328 278432 : if (i < MaxConnections)
329 : {
330 : /* PGPROC for normal backend, add to freeProcs list */
331 139282 : dlist_push_tail(&ProcGlobal->freeProcs, &proc->links);
332 139282 : proc->procgloballist = &ProcGlobal->freeProcs;
333 : }
334 139150 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS)
335 : {
336 : /* PGPROC for AV or special worker, add to autovacFreeProcs list */
337 27640 : dlist_push_tail(&ProcGlobal->autovacFreeProcs, &proc->links);
338 27640 : proc->procgloballist = &ProcGlobal->autovacFreeProcs;
339 : }
340 111510 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS + max_worker_processes)
341 : {
342 : /* PGPROC for bgworker, add to bgworkerFreeProcs list */
343 16764 : dlist_push_tail(&ProcGlobal->bgworkerFreeProcs, &proc->links);
344 16764 : proc->procgloballist = &ProcGlobal->bgworkerFreeProcs;
345 : }
346 94746 : else if (i < MaxBackends)
347 : {
348 : /* PGPROC for walsender, add to walsenderFreeProcs list */
349 13416 : dlist_push_tail(&ProcGlobal->walsenderFreeProcs, &proc->links);
350 13416 : proc->procgloballist = &ProcGlobal->walsenderFreeProcs;
351 : }
352 :
353 : /* Initialize myProcLocks[] shared memory queues. */
354 4733344 : for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
355 4454912 : dlist_init(&(proc->myProcLocks[j]));
356 :
357 : /* Initialize lockGroupMembers list. */
358 278432 : dlist_init(&proc->lockGroupMembers);
359 :
360 : /*
361 : * Initialize the atomic variables, otherwise, it won't be safe to
362 : * access them for backends that aren't currently in use.
363 : */
364 278432 : pg_atomic_init_u32(&(proc->procArrayGroupNext), INVALID_PROC_NUMBER);
365 278432 : pg_atomic_init_u32(&(proc->clogGroupNext), INVALID_PROC_NUMBER);
366 278432 : pg_atomic_init_u64(&(proc->waitStart), 0);
367 : }
368 :
369 : /* Should have consumed exactly the expected amount of fast-path memory. */
370 : Assert(fpPtr == fpEndPtr);
371 :
372 : /*
373 : * Save pointers to the blocks of PGPROC structures reserved for auxiliary
374 : * processes and prepared transactions.
375 : */
376 2096 : AuxiliaryProcs = &procs[MaxBackends];
377 2096 : PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
378 :
379 : /* Create ProcStructLock spinlock, too */
380 2096 : ProcStructLock = (slock_t *) ShmemInitStruct("ProcStructLock spinlock",
381 : sizeof(slock_t),
382 : &found);
383 2096 : SpinLockInit(ProcStructLock);
384 2096 : }
385 :
386 : /*
387 : * InitProcess -- initialize a per-process PGPROC entry for this backend
388 : */
389 : void
390 33736 : InitProcess(void)
391 : {
392 : dlist_head *procgloballist;
393 :
394 : /*
395 : * ProcGlobal should be set up already (if we are a backend, we inherit
396 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
397 : */
398 33736 : if (ProcGlobal == NULL)
399 0 : elog(PANIC, "proc header uninitialized");
400 :
401 33736 : if (MyProc != NULL)
402 0 : elog(ERROR, "you already exist");
403 :
404 : /*
405 : * Before we start accessing the shared memory in a serious way, mark
406 : * ourselves as an active postmaster child; this is so that the postmaster
407 : * can detect it if we exit without cleaning up.
408 : */
409 33736 : if (IsUnderPostmaster)
410 33522 : RegisterPostmasterChildActive();
411 :
412 : /*
413 : * Decide which list should supply our PGPROC. This logic must match the
414 : * way the freelists were constructed in InitProcGlobal().
415 : */
416 33736 : if (AmAutoVacuumWorkerProcess() || AmSpecialWorkerProcess())
417 4226 : procgloballist = &ProcGlobal->autovacFreeProcs;
418 29510 : else if (AmBackgroundWorkerProcess())
419 4434 : procgloballist = &ProcGlobal->bgworkerFreeProcs;
420 25076 : else if (AmWalSenderProcess())
421 2208 : procgloballist = &ProcGlobal->walsenderFreeProcs;
422 : else
423 22868 : procgloballist = &ProcGlobal->freeProcs;
424 :
425 : /*
426 : * Try to get a proc struct from the appropriate free list. If this
427 : * fails, we must be out of PGPROC structures (not to mention semaphores).
428 : *
429 : * While we are holding the ProcStructLock, also copy the current shared
430 : * estimate of spins_per_delay to local storage.
431 : */
432 33736 : SpinLockAcquire(ProcStructLock);
433 :
434 33736 : set_spins_per_delay(ProcGlobal->spins_per_delay);
435 :
436 33736 : if (!dlist_is_empty(procgloballist))
437 : {
438 33730 : MyProc = dlist_container(PGPROC, links, dlist_pop_head_node(procgloballist));
439 33730 : SpinLockRelease(ProcStructLock);
440 : }
441 : else
442 : {
443 : /*
444 : * If we reach here, all the PGPROCs are in use. This is one of the
445 : * possible places to detect "too many backends", so give the standard
446 : * error message. XXX do we need to give a different failure message
447 : * in the autovacuum case?
448 : */
449 6 : SpinLockRelease(ProcStructLock);
450 6 : if (AmWalSenderProcess())
451 4 : ereport(FATAL,
452 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
453 : errmsg("number of requested standby connections exceeds \"max_wal_senders\" (currently %d)",
454 : max_wal_senders)));
455 2 : ereport(FATAL,
456 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
457 : errmsg("sorry, too many clients already")));
458 : }
459 33730 : MyProcNumber = GetNumberFromPGProc(MyProc);
460 :
461 : /*
462 : * Cross-check that the PGPROC is of the type we expect; if this were not
463 : * the case, it would get returned to the wrong list.
464 : */
465 : Assert(MyProc->procgloballist == procgloballist);
466 :
467 : /*
468 : * Initialize all fields of MyProc, except for those previously
469 : * initialized by InitProcGlobal.
470 : */
471 33730 : dlist_node_init(&MyProc->links);
472 33730 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
473 33730 : MyProc->fpVXIDLock = false;
474 33730 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
475 33730 : MyProc->xid = InvalidTransactionId;
476 33730 : MyProc->xmin = InvalidTransactionId;
477 33730 : MyProc->pid = MyProcPid;
478 33730 : MyProc->vxid.procNumber = MyProcNumber;
479 33730 : MyProc->vxid.lxid = InvalidLocalTransactionId;
480 : /* databaseId and roleId will be filled in later */
481 33730 : MyProc->databaseId = InvalidOid;
482 33730 : MyProc->roleId = InvalidOid;
483 33730 : MyProc->tempNamespaceId = InvalidOid;
484 33730 : MyProc->isRegularBackend = AmRegularBackendProcess();
485 33730 : MyProc->delayChkptFlags = 0;
486 33730 : MyProc->statusFlags = 0;
487 : /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
488 33730 : if (AmAutoVacuumWorkerProcess())
489 3472 : MyProc->statusFlags |= PROC_IS_AUTOVACUUM;
490 33730 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
491 33730 : MyProc->lwWaitMode = 0;
492 33730 : MyProc->waitLock = NULL;
493 33730 : MyProc->waitProcLock = NULL;
494 33730 : pg_atomic_write_u64(&MyProc->waitStart, 0);
495 : #ifdef USE_ASSERT_CHECKING
496 : {
497 : int i;
498 :
499 : /* Last process should have released all locks. */
500 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
501 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
502 : }
503 : #endif
504 33730 : MyProc->recoveryConflictPending = false;
505 :
506 : /* Initialize fields for sync rep */
507 33730 : MyProc->waitLSN = 0;
508 33730 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
509 33730 : dlist_node_init(&MyProc->syncRepLinks);
510 :
511 : /* Initialize fields for group XID clearing. */
512 33730 : MyProc->procArrayGroupMember = false;
513 33730 : MyProc->procArrayGroupMemberXid = InvalidTransactionId;
514 : Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PROC_NUMBER);
515 :
516 : /* Check that group locking fields are in a proper initial state. */
517 : Assert(MyProc->lockGroupLeader == NULL);
518 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
519 :
520 : /* Initialize wait event information. */
521 33730 : MyProc->wait_event_info = 0;
522 :
523 : /* Initialize fields for group transaction status update. */
524 33730 : MyProc->clogGroupMember = false;
525 33730 : MyProc->clogGroupMemberXid = InvalidTransactionId;
526 33730 : MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
527 33730 : MyProc->clogGroupMemberPage = -1;
528 33730 : MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
529 : Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PROC_NUMBER);
530 :
531 : /*
532 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
533 : * on it. That allows us to repoint the process latch, which so far
534 : * points to process local one, to the shared one.
535 : */
536 33730 : OwnLatch(&MyProc->procLatch);
537 33730 : SwitchToSharedLatch();
538 :
539 : /* now that we have a proc, report wait events to shared memory */
540 33730 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
541 :
542 : /*
543 : * We might be reusing a semaphore that belonged to a failed process. So
544 : * be careful and reinitialize its value here. (This is not strictly
545 : * necessary anymore, but seems like a good idea for cleanliness.)
546 : */
547 33730 : PGSemaphoreReset(MyProc->sem);
548 :
549 : /*
550 : * Arrange to clean up at backend exit.
551 : */
552 33730 : on_shmem_exit(ProcKill, 0);
553 :
554 : /*
555 : * Now that we have a PGPROC, we could try to acquire locks, so initialize
556 : * local state needed for LWLocks, and the deadlock checker.
557 : */
558 33730 : InitLWLockAccess();
559 33730 : InitDeadLockChecking();
560 :
561 : #ifdef EXEC_BACKEND
562 :
563 : /*
564 : * Initialize backend-local pointers to all the shared data structures.
565 : * (We couldn't do this until now because it needs LWLocks.)
566 : */
567 : if (IsUnderPostmaster)
568 : AttachSharedMemoryStructs();
569 : #endif
570 33730 : }
571 :
572 : /*
573 : * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
574 : *
575 : * This is separate from InitProcess because we can't acquire LWLocks until
576 : * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
577 : * work until after we've done AttachSharedMemoryStructs.
578 : */
579 : void
580 33712 : InitProcessPhase2(void)
581 : {
582 : Assert(MyProc != NULL);
583 :
584 : /*
585 : * Add our PGPROC to the PGPROC array in shared memory.
586 : */
587 33712 : ProcArrayAdd(MyProc);
588 :
589 : /*
590 : * Arrange to clean that up at backend exit.
591 : */
592 33712 : on_shmem_exit(RemoveProcFromArray, 0);
593 33712 : }
594 :
595 : /*
596 : * InitAuxiliaryProcess -- create a PGPROC entry for an auxiliary process
597 : *
598 : * This is called by bgwriter and similar processes so that they will have a
599 : * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
600 : * and sema that are assigned are one of the extra ones created during
601 : * InitProcGlobal.
602 : *
603 : * Auxiliary processes are presently not expected to wait for real (lockmgr)
604 : * locks, so we need not set up the deadlock checker. They are never added
605 : * to the ProcArray or the sinval messaging mechanism, either. They also
606 : * don't get a VXID assigned, since this is only useful when we actually
607 : * hold lockmgr locks.
608 : *
609 : * Startup process however uses locks but never waits for them in the
610 : * normal backend sense. Startup process also takes part in sinval messaging
611 : * as a sendOnly process, so never reads messages from sinval queue. So
612 : * Startup process does have a VXID and does show up in pg_locks.
613 : */
614 : void
615 8092 : InitAuxiliaryProcess(void)
616 : {
617 : PGPROC *auxproc;
618 : int proctype;
619 :
620 : /*
621 : * ProcGlobal should be set up already (if we are a backend, we inherit
622 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
623 : */
624 8092 : if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
625 0 : elog(PANIC, "proc header uninitialized");
626 :
627 8092 : if (MyProc != NULL)
628 0 : elog(ERROR, "you already exist");
629 :
630 8092 : if (IsUnderPostmaster)
631 8092 : RegisterPostmasterChildActive();
632 :
633 : /*
634 : * We use the ProcStructLock to protect assignment and releasing of
635 : * AuxiliaryProcs entries.
636 : *
637 : * While we are holding the ProcStructLock, also copy the current shared
638 : * estimate of spins_per_delay to local storage.
639 : */
640 8092 : SpinLockAcquire(ProcStructLock);
641 :
642 8092 : set_spins_per_delay(ProcGlobal->spins_per_delay);
643 :
644 : /*
645 : * Find a free auxproc ... *big* trouble if there isn't one ...
646 : */
647 31910 : for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
648 : {
649 31910 : auxproc = &AuxiliaryProcs[proctype];
650 31910 : if (auxproc->pid == 0)
651 8092 : break;
652 : }
653 8092 : if (proctype >= NUM_AUXILIARY_PROCS)
654 : {
655 0 : SpinLockRelease(ProcStructLock);
656 0 : elog(FATAL, "all AuxiliaryProcs are in use");
657 : }
658 :
659 : /* Mark auxiliary proc as in use by me */
660 : /* use volatile pointer to prevent code rearrangement */
661 8092 : ((volatile PGPROC *) auxproc)->pid = MyProcPid;
662 :
663 8092 : SpinLockRelease(ProcStructLock);
664 :
665 8092 : MyProc = auxproc;
666 8092 : MyProcNumber = GetNumberFromPGProc(MyProc);
667 :
668 : /*
669 : * Initialize all fields of MyProc, except for those previously
670 : * initialized by InitProcGlobal.
671 : */
672 8092 : dlist_node_init(&MyProc->links);
673 8092 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
674 8092 : MyProc->fpVXIDLock = false;
675 8092 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
676 8092 : MyProc->xid = InvalidTransactionId;
677 8092 : MyProc->xmin = InvalidTransactionId;
678 8092 : MyProc->vxid.procNumber = INVALID_PROC_NUMBER;
679 8092 : MyProc->vxid.lxid = InvalidLocalTransactionId;
680 8092 : MyProc->databaseId = InvalidOid;
681 8092 : MyProc->roleId = InvalidOid;
682 8092 : MyProc->tempNamespaceId = InvalidOid;
683 8092 : MyProc->isRegularBackend = false;
684 8092 : MyProc->delayChkptFlags = 0;
685 8092 : MyProc->statusFlags = 0;
686 8092 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
687 8092 : MyProc->lwWaitMode = 0;
688 8092 : MyProc->waitLock = NULL;
689 8092 : MyProc->waitProcLock = NULL;
690 8092 : pg_atomic_write_u64(&MyProc->waitStart, 0);
691 : #ifdef USE_ASSERT_CHECKING
692 : {
693 : int i;
694 :
695 : /* Last process should have released all locks. */
696 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
697 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
698 : }
699 : #endif
700 :
701 : /*
702 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
703 : * on it. That allows us to repoint the process latch, which so far
704 : * points to process local one, to the shared one.
705 : */
706 8092 : OwnLatch(&MyProc->procLatch);
707 8092 : SwitchToSharedLatch();
708 :
709 : /* now that we have a proc, report wait events to shared memory */
710 8092 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
711 :
712 : /* Check that group locking fields are in a proper initial state. */
713 : Assert(MyProc->lockGroupLeader == NULL);
714 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
715 :
716 : /*
717 : * We might be reusing a semaphore that belonged to a failed process. So
718 : * be careful and reinitialize its value here. (This is not strictly
719 : * necessary anymore, but seems like a good idea for cleanliness.)
720 : */
721 8092 : PGSemaphoreReset(MyProc->sem);
722 :
723 : /*
724 : * Arrange to clean up at process exit.
725 : */
726 8092 : on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
727 :
728 : /*
729 : * Now that we have a PGPROC, we could try to acquire lightweight locks.
730 : * Initialize local state needed for them. (Heavyweight locks cannot be
731 : * acquired in aux processes.)
732 : */
733 8092 : InitLWLockAccess();
734 :
735 : #ifdef EXEC_BACKEND
736 :
737 : /*
738 : * Initialize backend-local pointers to all the shared data structures.
739 : * (We couldn't do this until now because it needs LWLocks.)
740 : */
741 : if (IsUnderPostmaster)
742 : AttachSharedMemoryStructs();
743 : #endif
744 8092 : }
745 :
746 : /*
747 : * Used from bufmgr to share the value of the buffer that Startup waits on,
748 : * or to reset the value to "not waiting" (-1). This allows processing
749 : * of recovery conflicts for buffer pins. Set is made before backends look
750 : * at this value, so locking not required, especially since the set is
751 : * an atomic integer set operation.
752 : */
753 : void
754 40 : SetStartupBufferPinWaitBufId(int bufid)
755 : {
756 : /* use volatile pointer to prevent code rearrangement */
757 40 : volatile PROC_HDR *procglobal = ProcGlobal;
758 :
759 40 : procglobal->startupBufferPinWaitBufId = bufid;
760 40 : }
761 :
762 : /*
763 : * Used by backends when they receive a request to check for buffer pin waits.
764 : */
765 : int
766 6 : GetStartupBufferPinWaitBufId(void)
767 : {
768 : /* use volatile pointer to prevent code rearrangement */
769 6 : volatile PROC_HDR *procglobal = ProcGlobal;
770 :
771 6 : return procglobal->startupBufferPinWaitBufId;
772 : }
773 :
774 : /*
775 : * Check whether there are at least N free PGPROC objects. If false is
776 : * returned, *nfree will be set to the number of free PGPROC objects.
777 : * Otherwise, *nfree will be set to n.
778 : *
779 : * Note: this is designed on the assumption that N will generally be small.
780 : */
781 : bool
782 474 : HaveNFreeProcs(int n, int *nfree)
783 : {
784 : dlist_iter iter;
785 :
786 : Assert(n > 0);
787 : Assert(nfree);
788 :
789 474 : SpinLockAcquire(ProcStructLock);
790 :
791 474 : *nfree = 0;
792 1418 : dlist_foreach(iter, &ProcGlobal->freeProcs)
793 : {
794 1412 : (*nfree)++;
795 1412 : if (*nfree == n)
796 468 : break;
797 : }
798 :
799 474 : SpinLockRelease(ProcStructLock);
800 :
801 474 : return (*nfree == n);
802 : }
803 :
804 : /*
805 : * Cancel any pending wait for lock, when aborting a transaction, and revert
806 : * any strong lock count acquisition for a lock being acquired.
807 : *
808 : * (Normally, this would only happen if we accept a cancel/die
809 : * interrupt while waiting; but an ereport(ERROR) before or during the lock
810 : * wait is within the realm of possibility, too.)
811 : */
812 : void
813 1133400 : LockErrorCleanup(void)
814 : {
815 : LOCALLOCK *lockAwaited;
816 : LWLock *partitionLock;
817 : DisableTimeoutParams timeouts[2];
818 :
819 1133400 : HOLD_INTERRUPTS();
820 :
821 1133400 : AbortStrongLockAcquire();
822 :
823 : /* Nothing to do if we weren't waiting for a lock */
824 1133400 : lockAwaited = GetAwaitedLock();
825 1133400 : if (lockAwaited == NULL)
826 : {
827 1133312 : RESUME_INTERRUPTS();
828 1133312 : return;
829 : }
830 :
831 : /*
832 : * Turn off the deadlock and lock timeout timers, if they are still
833 : * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
834 : * indicator flag, since this function is executed before
835 : * ProcessInterrupts when responding to SIGINT; else we'd lose the
836 : * knowledge that the SIGINT came from a lock timeout and not an external
837 : * source.
838 : */
839 88 : timeouts[0].id = DEADLOCK_TIMEOUT;
840 88 : timeouts[0].keep_indicator = false;
841 88 : timeouts[1].id = LOCK_TIMEOUT;
842 88 : timeouts[1].keep_indicator = true;
843 88 : disable_timeouts(timeouts, 2);
844 :
845 : /* Unlink myself from the wait queue, if on it (might not be anymore!) */
846 88 : partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
847 88 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
848 :
849 88 : if (!dlist_node_is_detached(&MyProc->links))
850 : {
851 : /* We could not have been granted the lock yet */
852 84 : RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
853 : }
854 : else
855 : {
856 : /*
857 : * Somebody kicked us off the lock queue already. Perhaps they
858 : * granted us the lock, or perhaps they detected a deadlock. If they
859 : * did grant us the lock, we'd better remember it in our local lock
860 : * table.
861 : */
862 4 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
863 4 : GrantAwaitedLock();
864 : }
865 :
866 88 : ResetAwaitedLock();
867 :
868 88 : LWLockRelease(partitionLock);
869 :
870 88 : RESUME_INTERRUPTS();
871 : }
872 :
873 :
874 : /*
875 : * ProcReleaseLocks() -- release locks associated with current transaction
876 : * at main transaction commit or abort
877 : *
878 : * At main transaction commit, we release standard locks except session locks.
879 : * At main transaction abort, we release all locks including session locks.
880 : *
881 : * Advisory locks are released only if they are transaction-level;
882 : * session-level holds remain, whether this is a commit or not.
883 : *
884 : * At subtransaction commit, we don't release any locks (so this func is not
885 : * needed at all); we will defer the releasing to the parent transaction.
886 : * At subtransaction abort, we release all locks held by the subtransaction;
887 : * this is implemented by retail releasing of the locks under control of
888 : * the ResourceOwner mechanism.
889 : */
890 : void
891 1073552 : ProcReleaseLocks(bool isCommit)
892 : {
893 1073552 : if (!MyProc)
894 0 : return;
895 : /* If waiting, get off wait queue (should only be needed after error) */
896 1073552 : LockErrorCleanup();
897 : /* Release standard locks, including session-level if aborting */
898 1073552 : LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
899 : /* Release transaction-level advisory locks */
900 1073552 : LockReleaseAll(USER_LOCKMETHOD, false);
901 : }
902 :
903 :
904 : /*
905 : * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
906 : */
907 : static void
908 33712 : RemoveProcFromArray(int code, Datum arg)
909 : {
910 : Assert(MyProc != NULL);
911 33712 : ProcArrayRemove(MyProc, InvalidTransactionId);
912 33712 : }
913 :
914 : /*
915 : * ProcKill() -- Destroy the per-proc data structure for
916 : * this process. Release any of its held LW locks.
917 : */
918 : static void
919 33730 : ProcKill(int code, Datum arg)
920 : {
921 : PGPROC *proc;
922 : dlist_head *procgloballist;
923 :
924 : Assert(MyProc != NULL);
925 :
926 : /* not safe if forked by system(), etc. */
927 33730 : if (MyProc->pid != (int) getpid())
928 0 : elog(PANIC, "ProcKill() called in child process");
929 :
930 : /* Make sure we're out of the sync rep lists */
931 33730 : SyncRepCleanupAtProcExit();
932 :
933 : #ifdef USE_ASSERT_CHECKING
934 : {
935 : int i;
936 :
937 : /* Last process should have released all locks. */
938 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
939 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
940 : }
941 : #endif
942 :
943 : /*
944 : * Release any LW locks I am holding. There really shouldn't be any, but
945 : * it's cheap to check again before we cut the knees off the LWLock
946 : * facility by releasing our PGPROC ...
947 : */
948 33730 : LWLockReleaseAll();
949 :
950 : /* Cancel any pending condition variable sleep, too */
951 33730 : ConditionVariableCancelSleep();
952 :
953 : /*
954 : * Detach from any lock group of which we are a member. If the leader
955 : * exits before all other group members, its PGPROC will remain allocated
956 : * until the last group process exits; that process must return the
957 : * leader's PGPROC to the appropriate list.
958 : */
959 33730 : if (MyProc->lockGroupLeader != NULL)
960 : {
961 2872 : PGPROC *leader = MyProc->lockGroupLeader;
962 2872 : LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
963 :
964 2872 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
965 : Assert(!dlist_is_empty(&leader->lockGroupMembers));
966 2872 : dlist_delete(&MyProc->lockGroupLink);
967 2872 : if (dlist_is_empty(&leader->lockGroupMembers))
968 : {
969 138 : leader->lockGroupLeader = NULL;
970 138 : if (leader != MyProc)
971 : {
972 0 : procgloballist = leader->procgloballist;
973 :
974 : /* Leader exited first; return its PGPROC. */
975 0 : SpinLockAcquire(ProcStructLock);
976 0 : dlist_push_head(procgloballist, &leader->links);
977 0 : SpinLockRelease(ProcStructLock);
978 : }
979 : }
980 2734 : else if (leader != MyProc)
981 2734 : MyProc->lockGroupLeader = NULL;
982 2872 : LWLockRelease(leader_lwlock);
983 : }
984 :
985 : /*
986 : * Reset MyLatch to the process local one. This is so that signal
987 : * handlers et al can continue using the latch after the shared latch
988 : * isn't ours anymore.
989 : *
990 : * Similarly, stop reporting wait events to MyProc->wait_event_info.
991 : *
992 : * After that clear MyProc and disown the shared latch.
993 : */
994 33730 : SwitchBackToLocalLatch();
995 33730 : pgstat_reset_wait_event_storage();
996 :
997 33730 : proc = MyProc;
998 33730 : MyProc = NULL;
999 33730 : MyProcNumber = INVALID_PROC_NUMBER;
1000 33730 : DisownLatch(&proc->procLatch);
1001 :
1002 : /* Mark the proc no longer in use */
1003 33730 : proc->pid = 0;
1004 33730 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1005 33730 : proc->vxid.lxid = InvalidTransactionId;
1006 :
1007 33730 : procgloballist = proc->procgloballist;
1008 33730 : SpinLockAcquire(ProcStructLock);
1009 :
1010 : /*
1011 : * If we're still a member of a locking group, that means we're a leader
1012 : * which has somehow exited before its children. The last remaining child
1013 : * will release our PGPROC. Otherwise, release it now.
1014 : */
1015 33730 : if (proc->lockGroupLeader == NULL)
1016 : {
1017 : /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
1018 : Assert(dlist_is_empty(&proc->lockGroupMembers));
1019 :
1020 : /* Return PGPROC structure (and semaphore) to appropriate freelist */
1021 33730 : dlist_push_tail(procgloballist, &proc->links);
1022 : }
1023 :
1024 : /* Update shared estimate of spins_per_delay */
1025 33730 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1026 :
1027 33730 : SpinLockRelease(ProcStructLock);
1028 :
1029 : /* wake autovac launcher if needed -- see comments in FreeWorkerInfo */
1030 33730 : if (AutovacuumLauncherPid != 0)
1031 3466 : kill(AutovacuumLauncherPid, SIGUSR2);
1032 33730 : }
1033 :
1034 : /*
1035 : * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
1036 : * processes (bgwriter, etc). The PGPROC and sema are not released, only
1037 : * marked as not-in-use.
1038 : */
1039 : static void
1040 8092 : AuxiliaryProcKill(int code, Datum arg)
1041 : {
1042 8092 : int proctype = DatumGetInt32(arg);
1043 : PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
1044 : PGPROC *proc;
1045 :
1046 : Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
1047 :
1048 : /* not safe if forked by system(), etc. */
1049 8092 : if (MyProc->pid != (int) getpid())
1050 0 : elog(PANIC, "AuxiliaryProcKill() called in child process");
1051 :
1052 8092 : auxproc = &AuxiliaryProcs[proctype];
1053 :
1054 : Assert(MyProc == auxproc);
1055 :
1056 : /* Release any LW locks I am holding (see notes above) */
1057 8092 : LWLockReleaseAll();
1058 :
1059 : /* Cancel any pending condition variable sleep, too */
1060 8092 : ConditionVariableCancelSleep();
1061 :
1062 : /* look at the equivalent ProcKill() code for comments */
1063 8092 : SwitchBackToLocalLatch();
1064 8092 : pgstat_reset_wait_event_storage();
1065 :
1066 8092 : proc = MyProc;
1067 8092 : MyProc = NULL;
1068 8092 : MyProcNumber = INVALID_PROC_NUMBER;
1069 8092 : DisownLatch(&proc->procLatch);
1070 :
1071 8092 : SpinLockAcquire(ProcStructLock);
1072 :
1073 : /* Mark auxiliary proc no longer in use */
1074 8092 : proc->pid = 0;
1075 8092 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1076 8092 : proc->vxid.lxid = InvalidTransactionId;
1077 :
1078 : /* Update shared estimate of spins_per_delay */
1079 8092 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1080 :
1081 8092 : SpinLockRelease(ProcStructLock);
1082 8092 : }
1083 :
1084 : /*
1085 : * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
1086 : * given its PID
1087 : *
1088 : * Returns NULL if not found.
1089 : */
1090 : PGPROC *
1091 10640 : AuxiliaryPidGetProc(int pid)
1092 : {
1093 10640 : PGPROC *result = NULL;
1094 : int index;
1095 :
1096 10640 : if (pid == 0) /* never match dummy PGPROCs */
1097 6 : return NULL;
1098 :
1099 51604 : for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
1100 : {
1101 51604 : PGPROC *proc = &AuxiliaryProcs[index];
1102 :
1103 51604 : if (proc->pid == pid)
1104 : {
1105 10634 : result = proc;
1106 10634 : break;
1107 : }
1108 : }
1109 10634 : return result;
1110 : }
1111 :
1112 :
1113 : /*
1114 : * JoinWaitQueue -- join the wait queue on the specified lock
1115 : *
1116 : * It's not actually guaranteed that we need to wait when this function is
1117 : * called, because it could be that when we try to find a position at which
1118 : * to insert ourself into the wait queue, we discover that we must be inserted
1119 : * ahead of everyone who wants a lock that conflict with ours. In that case,
1120 : * we get the lock immediately. Because of this, it's sensible for this function
1121 : * to have a dontWait argument, despite the name.
1122 : *
1123 : * On entry, the caller has already set up LOCK and PROCLOCK entries to
1124 : * reflect that we have "requested" the lock. The caller is responsible for
1125 : * cleaning that up, if we end up not joining the queue after all.
1126 : *
1127 : * The lock table's partition lock must be held at entry, and is still held
1128 : * at exit. The caller must release it before calling ProcSleep().
1129 : *
1130 : * Result is one of the following:
1131 : *
1132 : * PROC_WAIT_STATUS_OK - lock was immediately granted
1133 : * PROC_WAIT_STATUS_WAITING - joined the wait queue; call ProcSleep()
1134 : * PROC_WAIT_STATUS_ERROR - immediate deadlock was detected, or would
1135 : * need to wait and dontWait == true
1136 : *
1137 : * NOTES: The process queue is now a priority queue for locking.
1138 : */
1139 : ProcWaitStatus
1140 3976 : JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
1141 : {
1142 3976 : LOCKMODE lockmode = locallock->tag.mode;
1143 3976 : LOCK *lock = locallock->lock;
1144 3976 : PROCLOCK *proclock = locallock->proclock;
1145 3976 : uint32 hashcode = locallock->hashcode;
1146 3976 : LWLock *partitionLock PG_USED_FOR_ASSERTS_ONLY = LockHashPartitionLock(hashcode);
1147 3976 : dclist_head *waitQueue = &lock->waitProcs;
1148 3976 : PGPROC *insert_before = NULL;
1149 : LOCKMASK myProcHeldLocks;
1150 : LOCKMASK myHeldLocks;
1151 3976 : bool early_deadlock = false;
1152 3976 : PGPROC *leader = MyProc->lockGroupLeader;
1153 :
1154 : Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
1155 :
1156 : /*
1157 : * Set bitmask of locks this process already holds on this object.
1158 : */
1159 3976 : myHeldLocks = MyProc->heldLocks = proclock->holdMask;
1160 :
1161 : /*
1162 : * Determine which locks we're already holding.
1163 : *
1164 : * If group locking is in use, locks held by members of my locking group
1165 : * need to be included in myHeldLocks. This is not required for relation
1166 : * extension lock which conflict among group members. However, including
1167 : * them in myHeldLocks will give group members the priority to get those
1168 : * locks as compared to other backends which are also trying to acquire
1169 : * those locks. OTOH, we can avoid giving priority to group members for
1170 : * that kind of locks, but there doesn't appear to be a clear advantage of
1171 : * the same.
1172 : */
1173 3976 : myProcHeldLocks = proclock->holdMask;
1174 3976 : myHeldLocks = myProcHeldLocks;
1175 3976 : if (leader != NULL)
1176 : {
1177 : dlist_iter iter;
1178 :
1179 88 : dlist_foreach(iter, &lock->procLocks)
1180 : {
1181 : PROCLOCK *otherproclock;
1182 :
1183 66 : otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
1184 :
1185 66 : if (otherproclock->groupLeader == leader)
1186 30 : myHeldLocks |= otherproclock->holdMask;
1187 : }
1188 : }
1189 :
1190 : /*
1191 : * Determine where to add myself in the wait queue.
1192 : *
1193 : * Normally I should go at the end of the queue. However, if I already
1194 : * hold locks that conflict with the request of any previous waiter, put
1195 : * myself in the queue just in front of the first such waiter. This is not
1196 : * a necessary step, since deadlock detection would move me to before that
1197 : * waiter anyway; but it's relatively cheap to detect such a conflict
1198 : * immediately, and avoid delaying till deadlock timeout.
1199 : *
1200 : * Special case: if I find I should go in front of some waiter, check to
1201 : * see if I conflict with already-held locks or the requests before that
1202 : * waiter. If not, then just grant myself the requested lock immediately.
1203 : * This is the same as the test for immediate grant in LockAcquire, except
1204 : * we are only considering the part of the wait queue before my insertion
1205 : * point.
1206 : */
1207 3976 : if (myHeldLocks != 0 && !dclist_is_empty(waitQueue))
1208 : {
1209 18 : LOCKMASK aheadRequests = 0;
1210 : dlist_iter iter;
1211 :
1212 18 : dclist_foreach(iter, waitQueue)
1213 : {
1214 18 : PGPROC *proc = dlist_container(PGPROC, links, iter.cur);
1215 :
1216 : /*
1217 : * If we're part of the same locking group as this waiter, its
1218 : * locks neither conflict with ours nor contribute to
1219 : * aheadRequests.
1220 : */
1221 18 : if (leader != NULL && leader == proc->lockGroupLeader)
1222 0 : continue;
1223 :
1224 : /* Must he wait for me? */
1225 18 : if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1226 : {
1227 : /* Must I wait for him ? */
1228 18 : if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1229 : {
1230 : /*
1231 : * Yes, so we have a deadlock. Easiest way to clean up
1232 : * correctly is to call RemoveFromWaitQueue(), but we
1233 : * can't do that until we are *on* the wait queue. So, set
1234 : * a flag to check below, and break out of loop. Also,
1235 : * record deadlock info for later message.
1236 : */
1237 2 : RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1238 2 : early_deadlock = true;
1239 2 : break;
1240 : }
1241 : /* I must go before this waiter. Check special case. */
1242 16 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1243 16 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1244 : proclock))
1245 : {
1246 : /* Skip the wait and just grant myself the lock. */
1247 16 : GrantLock(lock, proclock, lockmode);
1248 16 : return PROC_WAIT_STATUS_OK;
1249 : }
1250 :
1251 : /* Put myself into wait queue before conflicting process */
1252 0 : insert_before = proc;
1253 0 : break;
1254 : }
1255 : /* Nope, so advance to next waiter */
1256 0 : aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1257 : }
1258 : }
1259 :
1260 : /*
1261 : * If we detected deadlock, give up without waiting. This must agree with
1262 : * CheckDeadLock's recovery code.
1263 : */
1264 3960 : if (early_deadlock)
1265 2 : return PROC_WAIT_STATUS_ERROR;
1266 :
1267 : /*
1268 : * At this point we know that we'd really need to sleep. If we've been
1269 : * commanded not to do that, bail out.
1270 : */
1271 3958 : if (dontWait)
1272 1390 : return PROC_WAIT_STATUS_ERROR;
1273 :
1274 : /*
1275 : * Insert self into queue, at the position determined above.
1276 : */
1277 2568 : if (insert_before)
1278 0 : dclist_insert_before(waitQueue, &insert_before->links, &MyProc->links);
1279 : else
1280 2568 : dclist_push_tail(waitQueue, &MyProc->links);
1281 :
1282 2568 : lock->waitMask |= LOCKBIT_ON(lockmode);
1283 :
1284 : /* Set up wait information in PGPROC object, too */
1285 2568 : MyProc->heldLocks = myProcHeldLocks;
1286 2568 : MyProc->waitLock = lock;
1287 2568 : MyProc->waitProcLock = proclock;
1288 2568 : MyProc->waitLockMode = lockmode;
1289 :
1290 2568 : MyProc->waitStatus = PROC_WAIT_STATUS_WAITING;
1291 :
1292 2568 : return PROC_WAIT_STATUS_WAITING;
1293 : }
1294 :
1295 : /*
1296 : * ProcSleep -- put process to sleep waiting on lock
1297 : *
1298 : * This must be called when JoinWaitQueue() returns PROC_WAIT_STATUS_WAITING.
1299 : * Returns after the lock has been granted, or if a deadlock is detected. Can
1300 : * also bail out with ereport(ERROR), if some other error condition, or a
1301 : * timeout or cancellation is triggered.
1302 : *
1303 : * Result is one of the following:
1304 : *
1305 : * PROC_WAIT_STATUS_OK - lock was granted
1306 : * PROC_WAIT_STATUS_ERROR - a deadlock was detected
1307 : */
1308 : ProcWaitStatus
1309 2568 : ProcSleep(LOCALLOCK *locallock)
1310 : {
1311 2568 : LOCKMODE lockmode = locallock->tag.mode;
1312 2568 : LOCK *lock = locallock->lock;
1313 2568 : uint32 hashcode = locallock->hashcode;
1314 2568 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1315 2568 : TimestampTz standbyWaitStart = 0;
1316 2568 : bool allow_autovacuum_cancel = true;
1317 2568 : bool logged_recovery_conflict = false;
1318 : ProcWaitStatus myWaitStatus;
1319 :
1320 : /* The caller must've armed the on-error cleanup mechanism */
1321 : Assert(GetAwaitedLock() == locallock);
1322 : Assert(!LWLockHeldByMe(partitionLock));
1323 :
1324 : /*
1325 : * Now that we will successfully clean up after an ereport, it's safe to
1326 : * check to see if there's a buffer pin deadlock against the Startup
1327 : * process. Of course, that's only necessary if we're doing Hot Standby
1328 : * and are not the Startup process ourselves.
1329 : */
1330 2568 : if (RecoveryInProgress() && !InRecovery)
1331 2 : CheckRecoveryConflictDeadlock();
1332 :
1333 : /* Reset deadlock_state before enabling the timeout handler */
1334 2568 : deadlock_state = DS_NOT_YET_CHECKED;
1335 2568 : got_deadlock_timeout = false;
1336 :
1337 : /*
1338 : * Set timer so we can wake up after awhile and check for a deadlock. If a
1339 : * deadlock is detected, the handler sets MyProc->waitStatus =
1340 : * PROC_WAIT_STATUS_ERROR, allowing us to know that we must report failure
1341 : * rather than success.
1342 : *
1343 : * By delaying the check until we've waited for a bit, we can avoid
1344 : * running the rather expensive deadlock-check code in most cases.
1345 : *
1346 : * If LockTimeout is set, also enable the timeout for that. We can save a
1347 : * few cycles by enabling both timeout sources in one call.
1348 : *
1349 : * If InHotStandby we set lock waits slightly later for clarity with other
1350 : * code.
1351 : */
1352 2568 : if (!InHotStandby)
1353 : {
1354 2566 : if (LockTimeout > 0)
1355 : {
1356 : EnableTimeoutParams timeouts[2];
1357 :
1358 230 : timeouts[0].id = DEADLOCK_TIMEOUT;
1359 230 : timeouts[0].type = TMPARAM_AFTER;
1360 230 : timeouts[0].delay_ms = DeadlockTimeout;
1361 230 : timeouts[1].id = LOCK_TIMEOUT;
1362 230 : timeouts[1].type = TMPARAM_AFTER;
1363 230 : timeouts[1].delay_ms = LockTimeout;
1364 230 : enable_timeouts(timeouts, 2);
1365 : }
1366 : else
1367 2336 : enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1368 :
1369 : /*
1370 : * Use the current time obtained for the deadlock timeout timer as
1371 : * waitStart (i.e., the time when this process started waiting for the
1372 : * lock). Since getting the current time newly can cause overhead, we
1373 : * reuse the already-obtained time to avoid that overhead.
1374 : *
1375 : * Note that waitStart is updated without holding the lock table's
1376 : * partition lock, to avoid the overhead by additional lock
1377 : * acquisition. This can cause "waitstart" in pg_locks to become NULL
1378 : * for a very short period of time after the wait started even though
1379 : * "granted" is false. This is OK in practice because we can assume
1380 : * that users are likely to look at "waitstart" when waiting for the
1381 : * lock for a long time.
1382 : */
1383 2566 : pg_atomic_write_u64(&MyProc->waitStart,
1384 2566 : get_timeout_start_time(DEADLOCK_TIMEOUT));
1385 : }
1386 2 : else if (log_recovery_conflict_waits)
1387 : {
1388 : /*
1389 : * Set the wait start timestamp if logging is enabled and in hot
1390 : * standby.
1391 : */
1392 2 : standbyWaitStart = GetCurrentTimestamp();
1393 : }
1394 :
1395 : /*
1396 : * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1397 : * will not wait. But a set latch does not necessarily mean that the lock
1398 : * is free now, as there are many other sources for latch sets than
1399 : * somebody releasing the lock.
1400 : *
1401 : * We process interrupts whenever the latch has been set, so cancel/die
1402 : * interrupts are processed quickly. This means we must not mind losing
1403 : * control to a cancel/die interrupt here. We don't, because we have no
1404 : * shared-state-change work to do after being granted the lock (the
1405 : * grantor did it all). We do have to worry about canceling the deadlock
1406 : * timeout and updating the locallock table, but if we lose control to an
1407 : * error, LockErrorCleanup will fix that up.
1408 : */
1409 : do
1410 : {
1411 6712 : if (InHotStandby)
1412 : {
1413 8 : bool maybe_log_conflict =
1414 8 : (standbyWaitStart != 0 && !logged_recovery_conflict);
1415 :
1416 : /* Set a timer and wait for that or for the lock to be granted */
1417 8 : ResolveRecoveryConflictWithLock(locallock->tag.lock,
1418 : maybe_log_conflict);
1419 :
1420 : /*
1421 : * Emit the log message if the startup process is waiting longer
1422 : * than deadlock_timeout for recovery conflict on lock.
1423 : */
1424 8 : if (maybe_log_conflict)
1425 : {
1426 4 : TimestampTz now = GetCurrentTimestamp();
1427 :
1428 4 : if (TimestampDifferenceExceeds(standbyWaitStart, now,
1429 : DeadlockTimeout))
1430 : {
1431 : VirtualTransactionId *vxids;
1432 : int cnt;
1433 :
1434 2 : vxids = GetLockConflicts(&locallock->tag.lock,
1435 : AccessExclusiveLock, &cnt);
1436 :
1437 : /*
1438 : * Log the recovery conflict and the list of PIDs of
1439 : * backends holding the conflicting lock. Note that we do
1440 : * logging even if there are no such backends right now
1441 : * because the startup process here has already waited
1442 : * longer than deadlock_timeout.
1443 : */
1444 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1445 : standbyWaitStart, now,
1446 2 : cnt > 0 ? vxids : NULL, true);
1447 2 : logged_recovery_conflict = true;
1448 : }
1449 : }
1450 : }
1451 : else
1452 : {
1453 6704 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1454 6704 : PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1455 6704 : ResetLatch(MyLatch);
1456 : /* check for deadlocks first, as that's probably log-worthy */
1457 6704 : if (got_deadlock_timeout)
1458 : {
1459 60 : CheckDeadLock();
1460 60 : got_deadlock_timeout = false;
1461 : }
1462 6704 : CHECK_FOR_INTERRUPTS();
1463 : }
1464 :
1465 : /*
1466 : * waitStatus could change from PROC_WAIT_STATUS_WAITING to something
1467 : * else asynchronously. Read it just once per loop to prevent
1468 : * surprising behavior (such as missing log messages).
1469 : */
1470 6624 : myWaitStatus = *((volatile ProcWaitStatus *) &MyProc->waitStatus);
1471 :
1472 : /*
1473 : * If we are not deadlocked, but are waiting on an autovacuum-induced
1474 : * task, send a signal to interrupt it.
1475 : */
1476 6624 : if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1477 : {
1478 0 : PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1479 : uint8 statusFlags;
1480 : uint8 lockmethod_copy;
1481 : LOCKTAG locktag_copy;
1482 :
1483 : /*
1484 : * Grab info we need, then release lock immediately. Note this
1485 : * coding means that there is a tiny chance that the process
1486 : * terminates its current transaction and starts a different one
1487 : * before we have a change to send the signal; the worst possible
1488 : * consequence is that a for-wraparound vacuum is canceled. But
1489 : * that could happen in any case unless we were to do kill() with
1490 : * the lock held, which is much more undesirable.
1491 : */
1492 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1493 0 : statusFlags = ProcGlobal->statusFlags[autovac->pgxactoff];
1494 0 : lockmethod_copy = lock->tag.locktag_lockmethodid;
1495 0 : locktag_copy = lock->tag;
1496 0 : LWLockRelease(ProcArrayLock);
1497 :
1498 : /*
1499 : * Only do it if the worker is not working to protect against Xid
1500 : * wraparound.
1501 : */
1502 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
1503 0 : !(statusFlags & PROC_VACUUM_FOR_WRAPAROUND))
1504 : {
1505 0 : int pid = autovac->pid;
1506 :
1507 : /* report the case, if configured to do so */
1508 0 : if (message_level_is_interesting(DEBUG1))
1509 : {
1510 : StringInfoData locktagbuf;
1511 : StringInfoData logbuf; /* errdetail for server log */
1512 :
1513 0 : initStringInfo(&locktagbuf);
1514 0 : initStringInfo(&logbuf);
1515 0 : DescribeLockTag(&locktagbuf, &locktag_copy);
1516 0 : appendStringInfo(&logbuf,
1517 : "Process %d waits for %s on %s.",
1518 : MyProcPid,
1519 : GetLockmodeName(lockmethod_copy, lockmode),
1520 : locktagbuf.data);
1521 :
1522 0 : ereport(DEBUG1,
1523 : (errmsg_internal("sending cancel to blocking autovacuum PID %d",
1524 : pid),
1525 : errdetail_log("%s", logbuf.data)));
1526 :
1527 0 : pfree(locktagbuf.data);
1528 0 : pfree(logbuf.data);
1529 : }
1530 :
1531 : /* send the autovacuum worker Back to Old Kent Road */
1532 0 : if (kill(pid, SIGINT) < 0)
1533 : {
1534 : /*
1535 : * There's a race condition here: once we release the
1536 : * ProcArrayLock, it's possible for the autovac worker to
1537 : * close up shop and exit before we can do the kill().
1538 : * Therefore, we do not whinge about no-such-process.
1539 : * Other errors such as EPERM could conceivably happen if
1540 : * the kernel recycles the PID fast enough, but such cases
1541 : * seem improbable enough that it's probably best to issue
1542 : * a warning if we see some other errno.
1543 : */
1544 0 : if (errno != ESRCH)
1545 0 : ereport(WARNING,
1546 : (errmsg("could not send signal to process %d: %m",
1547 : pid)));
1548 : }
1549 : }
1550 :
1551 : /* prevent signal from being sent again more than once */
1552 0 : allow_autovacuum_cancel = false;
1553 : }
1554 :
1555 : /*
1556 : * If awoken after the deadlock check interrupt has run, and
1557 : * log_lock_waits is on, then report about the wait.
1558 : */
1559 6624 : if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1560 : {
1561 : StringInfoData buf,
1562 : lock_waiters_sbuf,
1563 : lock_holders_sbuf;
1564 : const char *modename;
1565 : long secs;
1566 : int usecs;
1567 : long msecs;
1568 42 : int lockHoldersNum = 0;
1569 :
1570 42 : initStringInfo(&buf);
1571 42 : initStringInfo(&lock_waiters_sbuf);
1572 42 : initStringInfo(&lock_holders_sbuf);
1573 :
1574 42 : DescribeLockTag(&buf, &locallock->tag.lock);
1575 42 : modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1576 : lockmode);
1577 42 : TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1578 : GetCurrentTimestamp(),
1579 : &secs, &usecs);
1580 42 : msecs = secs * 1000 + usecs / 1000;
1581 42 : usecs = usecs % 1000;
1582 :
1583 : /* Gather a list of all lock holders and waiters */
1584 42 : LWLockAcquire(partitionLock, LW_SHARED);
1585 42 : GetLockHoldersAndWaiters(locallock, &lock_holders_sbuf,
1586 : &lock_waiters_sbuf, &lockHoldersNum);
1587 42 : LWLockRelease(partitionLock);
1588 :
1589 42 : if (deadlock_state == DS_SOFT_DEADLOCK)
1590 6 : ereport(LOG,
1591 : (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1592 : MyProcPid, modename, buf.data, msecs, usecs),
1593 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1594 : "Processes holding the lock: %s. Wait queue: %s.",
1595 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1596 36 : else if (deadlock_state == DS_HARD_DEADLOCK)
1597 : {
1598 : /*
1599 : * This message is a bit redundant with the error that will be
1600 : * reported subsequently, but in some cases the error report
1601 : * might not make it to the log (eg, if it's caught by an
1602 : * exception handler), and we want to ensure all long-wait
1603 : * events get logged.
1604 : */
1605 4 : ereport(LOG,
1606 : (errmsg("process %d detected deadlock while waiting for %s on %s after %ld.%03d ms",
1607 : MyProcPid, modename, buf.data, msecs, usecs),
1608 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1609 : "Processes holding the lock: %s. Wait queue: %s.",
1610 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1611 : }
1612 :
1613 42 : if (myWaitStatus == PROC_WAIT_STATUS_WAITING)
1614 20 : ereport(LOG,
1615 : (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1616 : MyProcPid, modename, buf.data, msecs, usecs),
1617 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1618 : "Processes holding the lock: %s. Wait queue: %s.",
1619 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1620 22 : else if (myWaitStatus == PROC_WAIT_STATUS_OK)
1621 18 : ereport(LOG,
1622 : (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1623 : MyProcPid, modename, buf.data, msecs, usecs)));
1624 : else
1625 : {
1626 : Assert(myWaitStatus == PROC_WAIT_STATUS_ERROR);
1627 :
1628 : /*
1629 : * Currently, the deadlock checker always kicks its own
1630 : * process, which means that we'll only see
1631 : * PROC_WAIT_STATUS_ERROR when deadlock_state ==
1632 : * DS_HARD_DEADLOCK, and there's no need to print redundant
1633 : * messages. But for completeness and future-proofing, print
1634 : * a message if it looks like someone else kicked us off the
1635 : * lock.
1636 : */
1637 4 : if (deadlock_state != DS_HARD_DEADLOCK)
1638 0 : ereport(LOG,
1639 : (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1640 : MyProcPid, modename, buf.data, msecs, usecs),
1641 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1642 : "Processes holding the lock: %s. Wait queue: %s.",
1643 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1644 : }
1645 :
1646 : /*
1647 : * At this point we might still need to wait for the lock. Reset
1648 : * state so we don't print the above messages again.
1649 : */
1650 42 : deadlock_state = DS_NO_DEADLOCK;
1651 :
1652 42 : pfree(buf.data);
1653 42 : pfree(lock_holders_sbuf.data);
1654 42 : pfree(lock_waiters_sbuf.data);
1655 : }
1656 6624 : } while (myWaitStatus == PROC_WAIT_STATUS_WAITING);
1657 :
1658 : /*
1659 : * Disable the timers, if they are still running. As in LockErrorCleanup,
1660 : * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1661 : * already caused QueryCancelPending to become set, we want the cancel to
1662 : * be reported as a lock timeout, not a user cancel.
1663 : */
1664 2480 : if (!InHotStandby)
1665 : {
1666 2478 : if (LockTimeout > 0)
1667 : {
1668 : DisableTimeoutParams timeouts[2];
1669 :
1670 218 : timeouts[0].id = DEADLOCK_TIMEOUT;
1671 218 : timeouts[0].keep_indicator = false;
1672 218 : timeouts[1].id = LOCK_TIMEOUT;
1673 218 : timeouts[1].keep_indicator = true;
1674 218 : disable_timeouts(timeouts, 2);
1675 : }
1676 : else
1677 2260 : disable_timeout(DEADLOCK_TIMEOUT, false);
1678 : }
1679 :
1680 : /*
1681 : * Emit the log message if recovery conflict on lock was resolved but the
1682 : * startup process waited longer than deadlock_timeout for it.
1683 : */
1684 2480 : if (InHotStandby && logged_recovery_conflict)
1685 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1686 : standbyWaitStart, GetCurrentTimestamp(),
1687 : NULL, false);
1688 :
1689 : /*
1690 : * We don't have to do anything else, because the awaker did all the
1691 : * necessary updates of the lock table and MyProc. (The caller is
1692 : * responsible for updating the local lock table.)
1693 : */
1694 2480 : return myWaitStatus;
1695 : }
1696 :
1697 :
1698 : /*
1699 : * ProcWakeup -- wake up a process by setting its latch.
1700 : *
1701 : * Also remove the process from the wait queue and set its links invalid.
1702 : *
1703 : * The appropriate lock partition lock must be held by caller.
1704 : *
1705 : * XXX: presently, this code is only used for the "success" case, and only
1706 : * works correctly for that case. To clean up in failure case, would need
1707 : * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1708 : * Hence, in practice the waitStatus parameter must be PROC_WAIT_STATUS_OK.
1709 : */
1710 : void
1711 2500 : ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus)
1712 : {
1713 2500 : if (dlist_node_is_detached(&proc->links))
1714 0 : return;
1715 :
1716 : Assert(proc->waitStatus == PROC_WAIT_STATUS_WAITING);
1717 :
1718 : /* Remove process from wait queue */
1719 2500 : dclist_delete_from_thoroughly(&proc->waitLock->waitProcs, &proc->links);
1720 :
1721 : /* Clean up process' state and pass it the ok/fail signal */
1722 2500 : proc->waitLock = NULL;
1723 2500 : proc->waitProcLock = NULL;
1724 2500 : proc->waitStatus = waitStatus;
1725 2500 : pg_atomic_write_u64(&MyProc->waitStart, 0);
1726 :
1727 : /* And awaken it */
1728 2500 : SetLatch(&proc->procLatch);
1729 : }
1730 :
1731 : /*
1732 : * ProcLockWakeup -- routine for waking up processes when a lock is
1733 : * released (or a prior waiter is aborted). Scan all waiters
1734 : * for lock, waken any that are no longer blocked.
1735 : *
1736 : * The appropriate lock partition lock must be held by caller.
1737 : */
1738 : void
1739 2556 : ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1740 : {
1741 2556 : dclist_head *waitQueue = &lock->waitProcs;
1742 2556 : LOCKMASK aheadRequests = 0;
1743 : dlist_mutable_iter miter;
1744 :
1745 2556 : if (dclist_is_empty(waitQueue))
1746 96 : return;
1747 :
1748 5726 : dclist_foreach_modify(miter, waitQueue)
1749 : {
1750 3266 : PGPROC *proc = dlist_container(PGPROC, links, miter.cur);
1751 3266 : LOCKMODE lockmode = proc->waitLockMode;
1752 :
1753 : /*
1754 : * Waken if (a) doesn't conflict with requests of earlier waiters, and
1755 : * (b) doesn't conflict with already-held locks.
1756 : */
1757 3266 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1758 2912 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1759 : proc->waitProcLock))
1760 : {
1761 : /* OK to waken */
1762 2500 : GrantLock(lock, proc->waitProcLock, lockmode);
1763 : /* removes proc from the lock's waiting process queue */
1764 2500 : ProcWakeup(proc, PROC_WAIT_STATUS_OK);
1765 : }
1766 : else
1767 : {
1768 : /*
1769 : * Lock conflicts: Don't wake, but remember requested mode for
1770 : * later checks.
1771 : */
1772 766 : aheadRequests |= LOCKBIT_ON(lockmode);
1773 : }
1774 : }
1775 : }
1776 :
1777 : /*
1778 : * CheckDeadLock
1779 : *
1780 : * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1781 : * lock to be released by some other process. Check if there's a deadlock; if
1782 : * not, just return. (But signal ProcSleep to log a message, if
1783 : * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1784 : * the lock's wait queue and signal an error to ProcSleep.
1785 : */
1786 : static void
1787 60 : CheckDeadLock(void)
1788 : {
1789 : int i;
1790 :
1791 : /*
1792 : * Acquire exclusive lock on the entire shared lock data structures. Must
1793 : * grab LWLocks in partition-number order to avoid LWLock deadlock.
1794 : *
1795 : * Note that the deadlock check interrupt had better not be enabled
1796 : * anywhere that this process itself holds lock partition locks, else this
1797 : * will wait forever. Also note that LWLockAcquire creates a critical
1798 : * section, so that this routine cannot be interrupted by cancel/die
1799 : * interrupts.
1800 : */
1801 1020 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1802 960 : LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1803 :
1804 : /*
1805 : * Check to see if we've been awoken by anyone in the interim.
1806 : *
1807 : * If we have, we can return and resume our transaction -- happy day.
1808 : * Before we are awoken the process releasing the lock grants it to us so
1809 : * we know that we don't have to wait anymore.
1810 : *
1811 : * We check by looking to see if we've been unlinked from the wait queue.
1812 : * This is safe because we hold the lock partition lock.
1813 : */
1814 60 : if (MyProc->links.prev == NULL ||
1815 60 : MyProc->links.next == NULL)
1816 0 : goto check_done;
1817 :
1818 : #ifdef LOCK_DEBUG
1819 : if (Debug_deadlocks)
1820 : DumpAllLocks();
1821 : #endif
1822 :
1823 : /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1824 60 : deadlock_state = DeadLockCheck(MyProc);
1825 :
1826 60 : if (deadlock_state == DS_HARD_DEADLOCK)
1827 : {
1828 : /*
1829 : * Oops. We have a deadlock.
1830 : *
1831 : * Get this process out of wait state. (Note: we could do this more
1832 : * efficiently by relying on lockAwaited, but use this coding to
1833 : * preserve the flexibility to kill some other transaction than the
1834 : * one detecting the deadlock.)
1835 : *
1836 : * RemoveFromWaitQueue sets MyProc->waitStatus to
1837 : * PROC_WAIT_STATUS_ERROR, so ProcSleep will report an error after we
1838 : * return from the signal handler.
1839 : */
1840 : Assert(MyProc->waitLock != NULL);
1841 12 : RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1842 :
1843 : /*
1844 : * We're done here. Transaction abort caused by the error that
1845 : * ProcSleep will raise will cause any other locks we hold to be
1846 : * released, thus allowing other processes to wake up; we don't need
1847 : * to do that here. NOTE: an exception is that releasing locks we
1848 : * hold doesn't consider the possibility of waiters that were blocked
1849 : * behind us on the lock we just failed to get, and might now be
1850 : * wakable because we're not in front of them anymore. However,
1851 : * RemoveFromWaitQueue took care of waking up any such processes.
1852 : */
1853 : }
1854 :
1855 : /*
1856 : * And release locks. We do this in reverse order for two reasons: (1)
1857 : * Anyone else who needs more than one of the locks will be trying to lock
1858 : * them in increasing order; we don't want to release the other process
1859 : * until it can get all the locks it needs. (2) This avoids O(N^2)
1860 : * behavior inside LWLockRelease.
1861 : */
1862 48 : check_done:
1863 1020 : for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1864 960 : LWLockRelease(LockHashPartitionLockByIndex(i));
1865 60 : }
1866 :
1867 : /*
1868 : * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1869 : *
1870 : * NB: Runs inside a signal handler, be careful.
1871 : */
1872 : void
1873 60 : CheckDeadLockAlert(void)
1874 : {
1875 60 : int save_errno = errno;
1876 :
1877 60 : got_deadlock_timeout = true;
1878 :
1879 : /*
1880 : * Have to set the latch again, even if handle_sig_alarm already did. Back
1881 : * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1882 : * ever would be a problem, but setting a set latch again is cheap.
1883 : *
1884 : * Note that, when this function runs inside procsignal_sigusr1_handler(),
1885 : * the handler function sets the latch again after the latch is set here.
1886 : */
1887 60 : SetLatch(MyLatch);
1888 60 : errno = save_errno;
1889 60 : }
1890 :
1891 : /*
1892 : * GetLockHoldersAndWaiters - get lock holders and waiters for a lock
1893 : *
1894 : * Fill lock_holders_sbuf and lock_waiters_sbuf with the PIDs of processes holding
1895 : * and waiting for the lock, and set lockHoldersNum to the number of lock holders.
1896 : *
1897 : * The lock table's partition lock must be held on entry and remains held on exit.
1898 : */
1899 : void
1900 42 : GetLockHoldersAndWaiters(LOCALLOCK *locallock, StringInfo lock_holders_sbuf,
1901 : StringInfo lock_waiters_sbuf, int *lockHoldersNum)
1902 : {
1903 : dlist_iter proc_iter;
1904 : PROCLOCK *curproclock;
1905 42 : LOCK *lock = locallock->lock;
1906 42 : bool first_holder = true,
1907 42 : first_waiter = true;
1908 :
1909 : #ifdef USE_ASSERT_CHECKING
1910 : {
1911 : uint32 hashcode = locallock->hashcode;
1912 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1913 :
1914 : Assert(LWLockHeldByMe(partitionLock));
1915 : }
1916 : #endif
1917 :
1918 42 : *lockHoldersNum = 0;
1919 :
1920 : /*
1921 : * Loop over the lock's procLocks to gather a list of all holders and
1922 : * waiters. Thus we will be able to provide more detailed information for
1923 : * lock debugging purposes.
1924 : *
1925 : * lock->procLocks contains all processes which hold or wait for this
1926 : * lock.
1927 : */
1928 130 : dlist_foreach(proc_iter, &lock->procLocks)
1929 : {
1930 88 : curproclock =
1931 88 : dlist_container(PROCLOCK, lockLink, proc_iter.cur);
1932 :
1933 : /*
1934 : * We are a waiter if myProc->waitProcLock == curproclock; we are a
1935 : * holder if it is NULL or something different.
1936 : */
1937 88 : if (curproclock->tag.myProc->waitProcLock == curproclock)
1938 : {
1939 38 : if (first_waiter)
1940 : {
1941 22 : appendStringInfo(lock_waiters_sbuf, "%d",
1942 22 : curproclock->tag.myProc->pid);
1943 22 : first_waiter = false;
1944 : }
1945 : else
1946 16 : appendStringInfo(lock_waiters_sbuf, ", %d",
1947 16 : curproclock->tag.myProc->pid);
1948 : }
1949 : else
1950 : {
1951 50 : if (first_holder)
1952 : {
1953 42 : appendStringInfo(lock_holders_sbuf, "%d",
1954 42 : curproclock->tag.myProc->pid);
1955 42 : first_holder = false;
1956 : }
1957 : else
1958 8 : appendStringInfo(lock_holders_sbuf, ", %d",
1959 8 : curproclock->tag.myProc->pid);
1960 :
1961 50 : (*lockHoldersNum)++;
1962 : }
1963 : }
1964 42 : }
1965 :
1966 : /*
1967 : * ProcWaitForSignal - wait for a signal from another backend.
1968 : *
1969 : * As this uses the generic process latch the caller has to be robust against
1970 : * unrelated wakeups: Always check that the desired state has occurred, and
1971 : * wait again if not.
1972 : */
1973 : void
1974 42 : ProcWaitForSignal(uint32 wait_event_info)
1975 : {
1976 42 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1977 : wait_event_info);
1978 42 : ResetLatch(MyLatch);
1979 42 : CHECK_FOR_INTERRUPTS();
1980 42 : }
1981 :
1982 : /*
1983 : * ProcSendSignal - set the latch of a backend identified by ProcNumber
1984 : */
1985 : void
1986 14 : ProcSendSignal(ProcNumber procNumber)
1987 : {
1988 14 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
1989 0 : elog(ERROR, "procNumber out of range");
1990 :
1991 14 : SetLatch(&ProcGlobal->allProcs[procNumber].procLatch);
1992 14 : }
1993 :
1994 : /*
1995 : * BecomeLockGroupLeader - designate process as lock group leader
1996 : *
1997 : * Once this function has returned, other processes can join the lock group
1998 : * by calling BecomeLockGroupMember.
1999 : */
2000 : void
2001 1172 : BecomeLockGroupLeader(void)
2002 : {
2003 : LWLock *leader_lwlock;
2004 :
2005 : /* If we already did it, we don't need to do it again. */
2006 1172 : if (MyProc->lockGroupLeader == MyProc)
2007 1034 : return;
2008 :
2009 : /* We had better not be a follower. */
2010 : Assert(MyProc->lockGroupLeader == NULL);
2011 :
2012 : /* Create single-member group, containing only ourselves. */
2013 138 : leader_lwlock = LockHashPartitionLockByProc(MyProc);
2014 138 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2015 138 : MyProc->lockGroupLeader = MyProc;
2016 138 : dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
2017 138 : LWLockRelease(leader_lwlock);
2018 : }
2019 :
2020 : /*
2021 : * BecomeLockGroupMember - designate process as lock group member
2022 : *
2023 : * This is pretty straightforward except for the possibility that the leader
2024 : * whose group we're trying to join might exit before we manage to do so;
2025 : * and the PGPROC might get recycled for an unrelated process. To avoid
2026 : * that, we require the caller to pass the PID of the intended PGPROC as
2027 : * an interlock. Returns true if we successfully join the intended lock
2028 : * group, and false if not.
2029 : */
2030 : bool
2031 2734 : BecomeLockGroupMember(PGPROC *leader, int pid)
2032 : {
2033 : LWLock *leader_lwlock;
2034 2734 : bool ok = false;
2035 :
2036 : /* Group leader can't become member of group */
2037 : Assert(MyProc != leader);
2038 :
2039 : /* Can't already be a member of a group */
2040 : Assert(MyProc->lockGroupLeader == NULL);
2041 :
2042 : /* PID must be valid. */
2043 : Assert(pid != 0);
2044 :
2045 : /*
2046 : * Get lock protecting the group fields. Note LockHashPartitionLockByProc
2047 : * calculates the proc number based on the PGPROC slot without looking at
2048 : * its contents, so we will acquire the correct lock even if the leader
2049 : * PGPROC is in process of being recycled.
2050 : */
2051 2734 : leader_lwlock = LockHashPartitionLockByProc(leader);
2052 2734 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2053 :
2054 : /* Is this the leader we're looking for? */
2055 2734 : if (leader->pid == pid && leader->lockGroupLeader == leader)
2056 : {
2057 : /* OK, join the group */
2058 2734 : ok = true;
2059 2734 : MyProc->lockGroupLeader = leader;
2060 2734 : dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
2061 : }
2062 2734 : LWLockRelease(leader_lwlock);
2063 :
2064 2734 : return ok;
2065 : }
|
