Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * proc.c
4 : * routines to manage per-process shared memory data structure
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/storage/lmgr/proc.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : /*
16 : * Interface (a):
17 : * JoinWaitQueue(), ProcSleep(), ProcWakeup()
18 : *
19 : * Waiting for a lock causes the backend to be put to sleep. Whoever releases
20 : * the lock wakes the process up again (and gives it an error code so it knows
21 : * whether it was awoken on an error condition).
22 : *
23 : * Interface (b):
24 : *
25 : * ProcReleaseLocks -- frees the locks associated with current transaction
26 : *
27 : * ProcKill -- destroys the shared memory state (and locks)
28 : * associated with the process.
29 : */
30 : #include "postgres.h"
31 :
32 : #include <signal.h>
33 : #include <unistd.h>
34 : #include <sys/time.h>
35 :
36 : #include "access/transam.h"
37 : #include "access/twophase.h"
38 : #include "access/xlogutils.h"
39 : #include "access/xlogwait.h"
40 : #include "miscadmin.h"
41 : #include "pgstat.h"
42 : #include "postmaster/autovacuum.h"
43 : #include "replication/slotsync.h"
44 : #include "replication/syncrep.h"
45 : #include "storage/condition_variable.h"
46 : #include "storage/ipc.h"
47 : #include "storage/lmgr.h"
48 : #include "storage/pmsignal.h"
49 : #include "storage/proc.h"
50 : #include "storage/procarray.h"
51 : #include "storage/procsignal.h"
52 : #include "storage/spin.h"
53 : #include "storage/standby.h"
54 : #include "utils/timeout.h"
55 : #include "utils/timestamp.h"
56 :
57 : /* GUC variables */
58 : int DeadlockTimeout = 1000;
59 : int StatementTimeout = 0;
60 : int LockTimeout = 0;
61 : int IdleInTransactionSessionTimeout = 0;
62 : int TransactionTimeout = 0;
63 : int IdleSessionTimeout = 0;
64 : bool log_lock_waits = true;
65 :
66 : /* Pointer to this process's PGPROC struct, if any */
67 : PGPROC *MyProc = NULL;
68 :
69 : /*
70 : * This spinlock protects the freelist of recycled PGPROC structures.
71 : * We cannot use an LWLock because the LWLock manager depends on already
72 : * having a PGPROC and a wait semaphore! But these structures are touched
73 : * relatively infrequently (only at backend startup or shutdown) and not for
74 : * very long, so a spinlock is okay.
75 : */
76 : NON_EXEC_STATIC slock_t *ProcStructLock = NULL;
77 :
78 : /* Pointers to shared-memory structures */
79 : PROC_HDR *ProcGlobal = NULL;
80 : NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
81 : PGPROC *PreparedXactProcs = NULL;
82 :
83 : static DeadLockState deadlock_state = DS_NOT_YET_CHECKED;
84 :
85 : /* Is a deadlock check pending? */
86 : static volatile sig_atomic_t got_deadlock_timeout;
87 :
88 : static void RemoveProcFromArray(int code, Datum arg);
89 : static void ProcKill(int code, Datum arg);
90 : static void AuxiliaryProcKill(int code, Datum arg);
91 : static void CheckDeadLock(void);
92 :
93 :
94 : /*
95 : * Report shared-memory space needed by PGPROC.
96 : */
97 : static Size
98 6534 : PGProcShmemSize(void)
99 : {
100 6534 : Size size = 0;
101 : Size TotalProcs =
102 6534 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
103 :
104 6534 : size = add_size(size, mul_size(TotalProcs, sizeof(PGPROC)));
105 6534 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->xids)));
106 6534 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->subxidStates)));
107 6534 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->statusFlags)));
108 :
109 6534 : return size;
110 : }
111 :
112 : /*
113 : * Report shared-memory space needed by Fast-Path locks.
114 : */
115 : static Size
116 6534 : FastPathLockShmemSize(void)
117 : {
118 6534 : Size size = 0;
119 : Size TotalProcs =
120 6534 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
121 : Size fpLockBitsSize,
122 : fpRelIdSize;
123 :
124 : /*
125 : * Memory needed for PGPROC fast-path lock arrays. Make sure the sizes are
126 : * nicely aligned in each backend.
127 : */
128 6534 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
129 6534 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
130 :
131 6534 : size = add_size(size, mul_size(TotalProcs, (fpLockBitsSize + fpRelIdSize)));
132 :
133 6534 : return size;
134 : }
135 :
136 : /*
137 : * Report shared-memory space needed by InitProcGlobal.
138 : */
139 : Size
140 4254 : ProcGlobalShmemSize(void)
141 : {
142 4254 : Size size = 0;
143 :
144 : /* ProcGlobal */
145 4254 : size = add_size(size, sizeof(PROC_HDR));
146 4254 : size = add_size(size, sizeof(slock_t));
147 :
148 4254 : size = add_size(size, PGSemaphoreShmemSize(ProcGlobalSemas()));
149 4254 : size = add_size(size, PGProcShmemSize());
150 4254 : size = add_size(size, FastPathLockShmemSize());
151 :
152 4254 : return size;
153 : }
154 :
155 : /*
156 : * Report number of semaphores needed by InitProcGlobal.
157 : */
158 : int
159 8504 : ProcGlobalSemas(void)
160 : {
161 : /*
162 : * We need a sema per backend (including autovacuum), plus one for each
163 : * auxiliary process.
164 : */
165 8504 : return MaxBackends + NUM_AUXILIARY_PROCS;
166 : }
167 :
168 : /*
169 : * InitProcGlobal -
170 : * Initialize the global process table during postmaster or standalone
171 : * backend startup.
172 : *
173 : * We also create all the per-process semaphores we will need to support
174 : * the requested number of backends. We used to allocate semaphores
175 : * only when backends were actually started up, but that is bad because
176 : * it lets Postgres fail under load --- a lot of Unix systems are
177 : * (mis)configured with small limits on the number of semaphores, and
178 : * running out when trying to start another backend is a common failure.
179 : * So, now we grab enough semaphores to support the desired max number
180 : * of backends immediately at initialization --- if the sysadmin has set
181 : * MaxConnections, max_worker_processes, max_wal_senders, or
182 : * autovacuum_worker_slots higher than his kernel will support, he'll
183 : * find out sooner rather than later.
184 : *
185 : * Another reason for creating semaphores here is that the semaphore
186 : * implementation typically requires us to create semaphores in the
187 : * postmaster, not in backends.
188 : *
189 : * Note: this is NOT called by individual backends under a postmaster,
190 : * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
191 : * pointers must be propagated specially for EXEC_BACKEND operation.
192 : */
193 : void
194 2280 : InitProcGlobal(void)
195 : {
196 : PGPROC *procs;
197 : int i,
198 : j;
199 : bool found;
200 2280 : uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
201 :
202 : /* Used for setup of per-backend fast-path slots. */
203 : char *fpPtr,
204 : *fpEndPtr PG_USED_FOR_ASSERTS_ONLY;
205 : Size fpLockBitsSize,
206 : fpRelIdSize;
207 : Size requestSize;
208 : char *ptr;
209 :
210 : /* Create the ProcGlobal shared structure */
211 2280 : ProcGlobal = (PROC_HDR *)
212 2280 : ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
213 : Assert(!found);
214 :
215 : /*
216 : * Initialize the data structures.
217 : */
218 2280 : ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
219 2280 : dlist_init(&ProcGlobal->freeProcs);
220 2280 : dlist_init(&ProcGlobal->autovacFreeProcs);
221 2280 : dlist_init(&ProcGlobal->bgworkerFreeProcs);
222 2280 : dlist_init(&ProcGlobal->walsenderFreeProcs);
223 2280 : ProcGlobal->startupBufferPinWaitBufId = -1;
224 2280 : ProcGlobal->walwriterProc = INVALID_PROC_NUMBER;
225 2280 : ProcGlobal->checkpointerProc = INVALID_PROC_NUMBER;
226 2280 : pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PROC_NUMBER);
227 2280 : pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PROC_NUMBER);
228 :
229 : /*
230 : * Create and initialize all the PGPROC structures we'll need. There are
231 : * six separate consumers: (1) normal backends, (2) autovacuum workers and
232 : * special workers, (3) background workers, (4) walsenders, (5) auxiliary
233 : * processes, and (6) prepared transactions. (For largely-historical
234 : * reasons, we combine autovacuum and special workers into one category
235 : * with a single freelist.) Each PGPROC structure is dedicated to exactly
236 : * one of these purposes, and they do not move between groups.
237 : */
238 2280 : requestSize = PGProcShmemSize();
239 :
240 2280 : ptr = ShmemInitStruct("PGPROC structures",
241 : requestSize,
242 : &found);
243 :
244 2280 : MemSet(ptr, 0, requestSize);
245 :
246 2280 : procs = (PGPROC *) ptr;
247 2280 : ptr = ptr + TotalProcs * sizeof(PGPROC);
248 :
249 2280 : ProcGlobal->allProcs = procs;
250 : /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
251 2280 : ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
252 :
253 : /*
254 : * Allocate arrays mirroring PGPROC fields in a dense manner. See
255 : * PROC_HDR.
256 : *
257 : * XXX: It might make sense to increase padding for these arrays, given
258 : * how hotly they are accessed.
259 : */
260 2280 : ProcGlobal->xids = (TransactionId *) ptr;
261 2280 : ptr = ptr + (TotalProcs * sizeof(*ProcGlobal->xids));
262 :
263 2280 : ProcGlobal->subxidStates = (XidCacheStatus *) ptr;
264 2280 : ptr = ptr + (TotalProcs * sizeof(*ProcGlobal->subxidStates));
265 :
266 2280 : ProcGlobal->statusFlags = (uint8 *) ptr;
267 2280 : ptr = ptr + (TotalProcs * sizeof(*ProcGlobal->statusFlags));
268 :
269 : /* make sure wer didn't overflow */
270 : Assert((ptr > (char *) procs) && (ptr <= (char *) procs + requestSize));
271 :
272 : /*
273 : * Allocate arrays for fast-path locks. Those are variable-length, so
274 : * can't be included in PGPROC directly. We allocate a separate piece of
275 : * shared memory and then divide that between backends.
276 : */
277 2280 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
278 2280 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
279 :
280 2280 : requestSize = FastPathLockShmemSize();
281 :
282 2280 : fpPtr = ShmemInitStruct("Fast-Path Lock Array",
283 : requestSize,
284 : &found);
285 :
286 2280 : MemSet(fpPtr, 0, requestSize);
287 :
288 : /* For asserts checking we did not overflow. */
289 2280 : fpEndPtr = fpPtr + requestSize;
290 :
291 : /* Reserve space for semaphores. */
292 2280 : PGReserveSemaphores(ProcGlobalSemas());
293 :
294 299852 : for (i = 0; i < TotalProcs; i++)
295 : {
296 297572 : PGPROC *proc = &procs[i];
297 :
298 : /* Common initialization for all PGPROCs, regardless of type. */
299 :
300 : /*
301 : * Set the fast-path lock arrays, and move the pointer. We interleave
302 : * the two arrays, to (hopefully) get some locality for each backend.
303 : */
304 297572 : proc->fpLockBits = (uint64 *) fpPtr;
305 297572 : fpPtr += fpLockBitsSize;
306 :
307 297572 : proc->fpRelId = (Oid *) fpPtr;
308 297572 : fpPtr += fpRelIdSize;
309 :
310 : Assert(fpPtr <= fpEndPtr);
311 :
312 : /*
313 : * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
314 : * dummy PGPROCs don't need these though - they're never associated
315 : * with a real process
316 : */
317 297572 : if (i < MaxBackends + NUM_AUXILIARY_PROCS)
318 : {
319 295840 : proc->sem = PGSemaphoreCreate();
320 295840 : InitSharedLatch(&(proc->procLatch));
321 295840 : LWLockInitialize(&(proc->fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
322 : }
323 :
324 : /*
325 : * Newly created PGPROCs for normal backends, autovacuum workers,
326 : * special workers, bgworkers, and walsenders must be queued up on the
327 : * appropriate free list. Because there can only ever be a small,
328 : * fixed number of auxiliary processes, no free list is used in that
329 : * case; InitAuxiliaryProcess() instead uses a linear search. PGPROCs
330 : * for prepared transactions are added to a free list by
331 : * TwoPhaseShmemInit().
332 : */
333 297572 : if (i < MaxConnections)
334 : {
335 : /* PGPROC for normal backend, add to freeProcs list */
336 146684 : dlist_push_tail(&ProcGlobal->freeProcs, &proc->links);
337 146684 : proc->procgloballist = &ProcGlobal->freeProcs;
338 : }
339 150888 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS)
340 : {
341 : /* PGPROC for AV or special worker, add to autovacFreeProcs list */
342 29418 : dlist_push_tail(&ProcGlobal->autovacFreeProcs, &proc->links);
343 29418 : proc->procgloballist = &ProcGlobal->autovacFreeProcs;
344 : }
345 121470 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS + max_worker_processes)
346 : {
347 : /* PGPROC for bgworker, add to bgworkerFreeProcs list */
348 18222 : dlist_push_tail(&ProcGlobal->bgworkerFreeProcs, &proc->links);
349 18222 : proc->procgloballist = &ProcGlobal->bgworkerFreeProcs;
350 : }
351 103248 : else if (i < MaxBackends)
352 : {
353 : /* PGPROC for walsender, add to walsenderFreeProcs list */
354 14876 : dlist_push_tail(&ProcGlobal->walsenderFreeProcs, &proc->links);
355 14876 : proc->procgloballist = &ProcGlobal->walsenderFreeProcs;
356 : }
357 :
358 : /* Initialize myProcLocks[] shared memory queues. */
359 5058724 : for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
360 4761152 : dlist_init(&(proc->myProcLocks[j]));
361 :
362 : /* Initialize lockGroupMembers list. */
363 297572 : dlist_init(&proc->lockGroupMembers);
364 :
365 : /*
366 : * Initialize the atomic variables, otherwise, it won't be safe to
367 : * access them for backends that aren't currently in use.
368 : */
369 297572 : pg_atomic_init_u32(&(proc->procArrayGroupNext), INVALID_PROC_NUMBER);
370 297572 : pg_atomic_init_u32(&(proc->clogGroupNext), INVALID_PROC_NUMBER);
371 297572 : pg_atomic_init_u64(&(proc->waitStart), 0);
372 : }
373 :
374 : /* Should have consumed exactly the expected amount of fast-path memory. */
375 : Assert(fpPtr == fpEndPtr);
376 :
377 : /*
378 : * Save pointers to the blocks of PGPROC structures reserved for auxiliary
379 : * processes and prepared transactions.
380 : */
381 2280 : AuxiliaryProcs = &procs[MaxBackends];
382 2280 : PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
383 :
384 : /* Create ProcStructLock spinlock, too */
385 2280 : ProcStructLock = (slock_t *) ShmemInitStruct("ProcStructLock spinlock",
386 : sizeof(slock_t),
387 : &found);
388 2280 : SpinLockInit(ProcStructLock);
389 2280 : }
390 :
391 : /*
392 : * InitProcess -- initialize a per-process PGPROC entry for this backend
393 : */
394 : void
395 35780 : InitProcess(void)
396 : {
397 : dlist_head *procgloballist;
398 :
399 : /*
400 : * ProcGlobal should be set up already (if we are a backend, we inherit
401 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
402 : */
403 35780 : if (ProcGlobal == NULL)
404 0 : elog(PANIC, "proc header uninitialized");
405 :
406 35780 : if (MyProc != NULL)
407 0 : elog(ERROR, "you already exist");
408 :
409 : /*
410 : * Before we start accessing the shared memory in a serious way, mark
411 : * ourselves as an active postmaster child; this is so that the postmaster
412 : * can detect it if we exit without cleaning up.
413 : */
414 35780 : if (IsUnderPostmaster)
415 35538 : RegisterPostmasterChildActive();
416 :
417 : /*
418 : * Decide which list should supply our PGPROC. This logic must match the
419 : * way the freelists were constructed in InitProcGlobal().
420 : */
421 35780 : if (AmAutoVacuumWorkerProcess() || AmSpecialWorkerProcess())
422 3246 : procgloballist = &ProcGlobal->autovacFreeProcs;
423 32534 : else if (AmBackgroundWorkerProcess())
424 5042 : procgloballist = &ProcGlobal->bgworkerFreeProcs;
425 27492 : else if (AmWalSenderProcess())
426 2422 : procgloballist = &ProcGlobal->walsenderFreeProcs;
427 : else
428 25070 : procgloballist = &ProcGlobal->freeProcs;
429 :
430 : /*
431 : * Try to get a proc struct from the appropriate free list. If this
432 : * fails, we must be out of PGPROC structures (not to mention semaphores).
433 : *
434 : * While we are holding the ProcStructLock, also copy the current shared
435 : * estimate of spins_per_delay to local storage.
436 : */
437 35780 : SpinLockAcquire(ProcStructLock);
438 :
439 35780 : set_spins_per_delay(ProcGlobal->spins_per_delay);
440 :
441 35780 : if (!dlist_is_empty(procgloballist))
442 : {
443 35774 : MyProc = dlist_container(PGPROC, links, dlist_pop_head_node(procgloballist));
444 35774 : SpinLockRelease(ProcStructLock);
445 : }
446 : else
447 : {
448 : /*
449 : * If we reach here, all the PGPROCs are in use. This is one of the
450 : * possible places to detect "too many backends", so give the standard
451 : * error message. XXX do we need to give a different failure message
452 : * in the autovacuum case?
453 : */
454 6 : SpinLockRelease(ProcStructLock);
455 6 : if (AmWalSenderProcess())
456 4 : ereport(FATAL,
457 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
458 : errmsg("number of requested standby connections exceeds \"max_wal_senders\" (currently %d)",
459 : max_wal_senders)));
460 2 : ereport(FATAL,
461 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
462 : errmsg("sorry, too many clients already")));
463 : }
464 35774 : MyProcNumber = GetNumberFromPGProc(MyProc);
465 :
466 : /*
467 : * Cross-check that the PGPROC is of the type we expect; if this were not
468 : * the case, it would get returned to the wrong list.
469 : */
470 : Assert(MyProc->procgloballist == procgloballist);
471 :
472 : /*
473 : * Initialize all fields of MyProc, except for those previously
474 : * initialized by InitProcGlobal.
475 : */
476 35774 : dlist_node_init(&MyProc->links);
477 35774 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
478 35774 : MyProc->fpVXIDLock = false;
479 35774 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
480 35774 : MyProc->xid = InvalidTransactionId;
481 35774 : MyProc->xmin = InvalidTransactionId;
482 35774 : MyProc->pid = MyProcPid;
483 35774 : MyProc->vxid.procNumber = MyProcNumber;
484 35774 : MyProc->vxid.lxid = InvalidLocalTransactionId;
485 : /* databaseId and roleId will be filled in later */
486 35774 : MyProc->databaseId = InvalidOid;
487 35774 : MyProc->roleId = InvalidOid;
488 35774 : MyProc->tempNamespaceId = InvalidOid;
489 35774 : MyProc->isRegularBackend = AmRegularBackendProcess();
490 35774 : MyProc->delayChkptFlags = 0;
491 35774 : MyProc->statusFlags = 0;
492 : /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
493 35774 : if (AmAutoVacuumWorkerProcess())
494 2400 : MyProc->statusFlags |= PROC_IS_AUTOVACUUM;
495 35774 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
496 35774 : MyProc->lwWaitMode = 0;
497 35774 : MyProc->waitLock = NULL;
498 35774 : MyProc->waitProcLock = NULL;
499 35774 : pg_atomic_write_u64(&MyProc->waitStart, 0);
500 : #ifdef USE_ASSERT_CHECKING
501 : {
502 : int i;
503 :
504 : /* Last process should have released all locks. */
505 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
506 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
507 : }
508 : #endif
509 35774 : MyProc->recoveryConflictPending = false;
510 :
511 : /* Initialize fields for sync rep */
512 35774 : MyProc->waitLSN = InvalidXLogRecPtr;
513 35774 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
514 35774 : dlist_node_init(&MyProc->syncRepLinks);
515 :
516 : /* Initialize fields for group XID clearing. */
517 35774 : MyProc->procArrayGroupMember = false;
518 35774 : MyProc->procArrayGroupMemberXid = InvalidTransactionId;
519 : Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PROC_NUMBER);
520 :
521 : /* Check that group locking fields are in a proper initial state. */
522 : Assert(MyProc->lockGroupLeader == NULL);
523 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
524 :
525 : /* Initialize wait event information. */
526 35774 : MyProc->wait_event_info = 0;
527 :
528 : /* Initialize fields for group transaction status update. */
529 35774 : MyProc->clogGroupMember = false;
530 35774 : MyProc->clogGroupMemberXid = InvalidTransactionId;
531 35774 : MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
532 35774 : MyProc->clogGroupMemberPage = -1;
533 35774 : MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
534 : Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PROC_NUMBER);
535 :
536 : /*
537 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
538 : * on it. That allows us to repoint the process latch, which so far
539 : * points to process local one, to the shared one.
540 : */
541 35774 : OwnLatch(&MyProc->procLatch);
542 35774 : SwitchToSharedLatch();
543 :
544 : /* now that we have a proc, report wait events to shared memory */
545 35774 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
546 :
547 : /*
548 : * We might be reusing a semaphore that belonged to a failed process. So
549 : * be careful and reinitialize its value here. (This is not strictly
550 : * necessary anymore, but seems like a good idea for cleanliness.)
551 : */
552 35774 : PGSemaphoreReset(MyProc->sem);
553 :
554 : /*
555 : * Arrange to clean up at backend exit.
556 : */
557 35774 : on_shmem_exit(ProcKill, 0);
558 :
559 : /*
560 : * Now that we have a PGPROC, we could try to acquire locks, so initialize
561 : * local state needed for LWLocks, and the deadlock checker.
562 : */
563 35774 : InitLWLockAccess();
564 35774 : InitDeadLockChecking();
565 :
566 : #ifdef EXEC_BACKEND
567 :
568 : /*
569 : * Initialize backend-local pointers to all the shared data structures.
570 : * (We couldn't do this until now because it needs LWLocks.)
571 : */
572 : if (IsUnderPostmaster)
573 : AttachSharedMemoryStructs();
574 : #endif
575 35774 : }
576 :
577 : /*
578 : * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
579 : *
580 : * This is separate from InitProcess because we can't acquire LWLocks until
581 : * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
582 : * work until after we've done AttachSharedMemoryStructs.
583 : */
584 : void
585 35756 : InitProcessPhase2(void)
586 : {
587 : Assert(MyProc != NULL);
588 :
589 : /*
590 : * Add our PGPROC to the PGPROC array in shared memory.
591 : */
592 35756 : ProcArrayAdd(MyProc);
593 :
594 : /*
595 : * Arrange to clean that up at backend exit.
596 : */
597 35756 : on_shmem_exit(RemoveProcFromArray, 0);
598 35756 : }
599 :
600 : /*
601 : * InitAuxiliaryProcess -- create a PGPROC entry for an auxiliary process
602 : *
603 : * This is called by bgwriter and similar processes so that they will have a
604 : * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
605 : * and sema that are assigned are one of the extra ones created during
606 : * InitProcGlobal.
607 : *
608 : * Auxiliary processes are presently not expected to wait for real (lockmgr)
609 : * locks, so we need not set up the deadlock checker. They are never added
610 : * to the ProcArray or the sinval messaging mechanism, either. They also
611 : * don't get a VXID assigned, since this is only useful when we actually
612 : * hold lockmgr locks.
613 : *
614 : * Startup process however uses locks but never waits for them in the
615 : * normal backend sense. Startup process also takes part in sinval messaging
616 : * as a sendOnly process, so never reads messages from sinval queue. So
617 : * Startup process does have a VXID and does show up in pg_locks.
618 : */
619 : void
620 9078 : InitAuxiliaryProcess(void)
621 : {
622 : PGPROC *auxproc;
623 : int proctype;
624 :
625 : /*
626 : * ProcGlobal should be set up already (if we are a backend, we inherit
627 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
628 : */
629 9078 : if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
630 0 : elog(PANIC, "proc header uninitialized");
631 :
632 9078 : if (MyProc != NULL)
633 0 : elog(ERROR, "you already exist");
634 :
635 9078 : if (IsUnderPostmaster)
636 9078 : RegisterPostmasterChildActive();
637 :
638 : /*
639 : * We use the ProcStructLock to protect assignment and releasing of
640 : * AuxiliaryProcs entries.
641 : *
642 : * While we are holding the ProcStructLock, also copy the current shared
643 : * estimate of spins_per_delay to local storage.
644 : */
645 9078 : SpinLockAcquire(ProcStructLock);
646 :
647 9078 : set_spins_per_delay(ProcGlobal->spins_per_delay);
648 :
649 : /*
650 : * Find a free auxproc ... *big* trouble if there isn't one ...
651 : */
652 35656 : for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
653 : {
654 35656 : auxproc = &AuxiliaryProcs[proctype];
655 35656 : if (auxproc->pid == 0)
656 9078 : break;
657 : }
658 9078 : if (proctype >= NUM_AUXILIARY_PROCS)
659 : {
660 0 : SpinLockRelease(ProcStructLock);
661 0 : elog(FATAL, "all AuxiliaryProcs are in use");
662 : }
663 :
664 : /* Mark auxiliary proc as in use by me */
665 : /* use volatile pointer to prevent code rearrangement */
666 9078 : ((volatile PGPROC *) auxproc)->pid = MyProcPid;
667 :
668 9078 : SpinLockRelease(ProcStructLock);
669 :
670 9078 : MyProc = auxproc;
671 9078 : MyProcNumber = GetNumberFromPGProc(MyProc);
672 :
673 : /*
674 : * Initialize all fields of MyProc, except for those previously
675 : * initialized by InitProcGlobal.
676 : */
677 9078 : dlist_node_init(&MyProc->links);
678 9078 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
679 9078 : MyProc->fpVXIDLock = false;
680 9078 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
681 9078 : MyProc->xid = InvalidTransactionId;
682 9078 : MyProc->xmin = InvalidTransactionId;
683 9078 : MyProc->vxid.procNumber = INVALID_PROC_NUMBER;
684 9078 : MyProc->vxid.lxid = InvalidLocalTransactionId;
685 9078 : MyProc->databaseId = InvalidOid;
686 9078 : MyProc->roleId = InvalidOid;
687 9078 : MyProc->tempNamespaceId = InvalidOid;
688 9078 : MyProc->isRegularBackend = false;
689 9078 : MyProc->delayChkptFlags = 0;
690 9078 : MyProc->statusFlags = 0;
691 9078 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
692 9078 : MyProc->lwWaitMode = 0;
693 9078 : MyProc->waitLock = NULL;
694 9078 : MyProc->waitProcLock = NULL;
695 9078 : pg_atomic_write_u64(&MyProc->waitStart, 0);
696 : #ifdef USE_ASSERT_CHECKING
697 : {
698 : int i;
699 :
700 : /* Last process should have released all locks. */
701 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
702 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
703 : }
704 : #endif
705 :
706 : /*
707 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
708 : * on it. That allows us to repoint the process latch, which so far
709 : * points to process local one, to the shared one.
710 : */
711 9078 : OwnLatch(&MyProc->procLatch);
712 9078 : SwitchToSharedLatch();
713 :
714 : /* now that we have a proc, report wait events to shared memory */
715 9078 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
716 :
717 : /* Check that group locking fields are in a proper initial state. */
718 : Assert(MyProc->lockGroupLeader == NULL);
719 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
720 :
721 : /*
722 : * We might be reusing a semaphore that belonged to a failed process. So
723 : * be careful and reinitialize its value here. (This is not strictly
724 : * necessary anymore, but seems like a good idea for cleanliness.)
725 : */
726 9078 : PGSemaphoreReset(MyProc->sem);
727 :
728 : /*
729 : * Arrange to clean up at process exit.
730 : */
731 9078 : on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
732 :
733 : /*
734 : * Now that we have a PGPROC, we could try to acquire lightweight locks.
735 : * Initialize local state needed for them. (Heavyweight locks cannot be
736 : * acquired in aux processes.)
737 : */
738 9078 : InitLWLockAccess();
739 :
740 : #ifdef EXEC_BACKEND
741 :
742 : /*
743 : * Initialize backend-local pointers to all the shared data structures.
744 : * (We couldn't do this until now because it needs LWLocks.)
745 : */
746 : if (IsUnderPostmaster)
747 : AttachSharedMemoryStructs();
748 : #endif
749 9078 : }
750 :
751 : /*
752 : * Used from bufmgr to share the value of the buffer that Startup waits on,
753 : * or to reset the value to "not waiting" (-1). This allows processing
754 : * of recovery conflicts for buffer pins. Set is made before backends look
755 : * at this value, so locking not required, especially since the set is
756 : * an atomic integer set operation.
757 : */
758 : void
759 44 : SetStartupBufferPinWaitBufId(int bufid)
760 : {
761 : /* use volatile pointer to prevent code rearrangement */
762 44 : volatile PROC_HDR *procglobal = ProcGlobal;
763 :
764 44 : procglobal->startupBufferPinWaitBufId = bufid;
765 44 : }
766 :
767 : /*
768 : * Used by backends when they receive a request to check for buffer pin waits.
769 : */
770 : int
771 10 : GetStartupBufferPinWaitBufId(void)
772 : {
773 : /* use volatile pointer to prevent code rearrangement */
774 10 : volatile PROC_HDR *procglobal = ProcGlobal;
775 :
776 10 : return procglobal->startupBufferPinWaitBufId;
777 : }
778 :
779 : /*
780 : * Check whether there are at least N free PGPROC objects. If false is
781 : * returned, *nfree will be set to the number of free PGPROC objects.
782 : * Otherwise, *nfree will be set to n.
783 : *
784 : * Note: this is designed on the assumption that N will generally be small.
785 : */
786 : bool
787 488 : HaveNFreeProcs(int n, int *nfree)
788 : {
789 : dlist_iter iter;
790 :
791 : Assert(n > 0);
792 : Assert(nfree);
793 :
794 488 : SpinLockAcquire(ProcStructLock);
795 :
796 488 : *nfree = 0;
797 1460 : dlist_foreach(iter, &ProcGlobal->freeProcs)
798 : {
799 1454 : (*nfree)++;
800 1454 : if (*nfree == n)
801 482 : break;
802 : }
803 :
804 488 : SpinLockRelease(ProcStructLock);
805 :
806 488 : return (*nfree == n);
807 : }
808 :
809 : /*
810 : * Cancel any pending wait for lock, when aborting a transaction, and revert
811 : * any strong lock count acquisition for a lock being acquired.
812 : *
813 : * (Normally, this would only happen if we accept a cancel/die
814 : * interrupt while waiting; but an ereport(ERROR) before or during the lock
815 : * wait is within the realm of possibility, too.)
816 : */
817 : void
818 1031494 : LockErrorCleanup(void)
819 : {
820 : LOCALLOCK *lockAwaited;
821 : LWLock *partitionLock;
822 : DisableTimeoutParams timeouts[2];
823 :
824 1031494 : HOLD_INTERRUPTS();
825 :
826 1031494 : AbortStrongLockAcquire();
827 :
828 : /* Nothing to do if we weren't waiting for a lock */
829 1031494 : lockAwaited = GetAwaitedLock();
830 1031494 : if (lockAwaited == NULL)
831 : {
832 1031414 : RESUME_INTERRUPTS();
833 1031414 : return;
834 : }
835 :
836 : /*
837 : * Turn off the deadlock and lock timeout timers, if they are still
838 : * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
839 : * indicator flag, since this function is executed before
840 : * ProcessInterrupts when responding to SIGINT; else we'd lose the
841 : * knowledge that the SIGINT came from a lock timeout and not an external
842 : * source.
843 : */
844 80 : timeouts[0].id = DEADLOCK_TIMEOUT;
845 80 : timeouts[0].keep_indicator = false;
846 80 : timeouts[1].id = LOCK_TIMEOUT;
847 80 : timeouts[1].keep_indicator = true;
848 80 : disable_timeouts(timeouts, 2);
849 :
850 : /* Unlink myself from the wait queue, if on it (might not be anymore!) */
851 80 : partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
852 80 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
853 :
854 80 : if (!dlist_node_is_detached(&MyProc->links))
855 : {
856 : /* We could not have been granted the lock yet */
857 78 : RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
858 : }
859 : else
860 : {
861 : /*
862 : * Somebody kicked us off the lock queue already. Perhaps they
863 : * granted us the lock, or perhaps they detected a deadlock. If they
864 : * did grant us the lock, we'd better remember it in our local lock
865 : * table.
866 : */
867 2 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
868 2 : GrantAwaitedLock();
869 : }
870 :
871 80 : ResetAwaitedLock();
872 :
873 80 : LWLockRelease(partitionLock);
874 :
875 80 : RESUME_INTERRUPTS();
876 : }
877 :
878 :
879 : /*
880 : * ProcReleaseLocks() -- release locks associated with current transaction
881 : * at main transaction commit or abort
882 : *
883 : * At main transaction commit, we release standard locks except session locks.
884 : * At main transaction abort, we release all locks including session locks.
885 : *
886 : * Advisory locks are released only if they are transaction-level;
887 : * session-level holds remain, whether this is a commit or not.
888 : *
889 : * At subtransaction commit, we don't release any locks (so this func is not
890 : * needed at all); we will defer the releasing to the parent transaction.
891 : * At subtransaction abort, we release all locks held by the subtransaction;
892 : * this is implemented by retail releasing of the locks under control of
893 : * the ResourceOwner mechanism.
894 : */
895 : void
896 968372 : ProcReleaseLocks(bool isCommit)
897 : {
898 968372 : if (!MyProc)
899 0 : return;
900 : /* If waiting, get off wait queue (should only be needed after error) */
901 968372 : LockErrorCleanup();
902 : /* Release standard locks, including session-level if aborting */
903 968372 : LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
904 : /* Release transaction-level advisory locks */
905 968372 : LockReleaseAll(USER_LOCKMETHOD, false);
906 : }
907 :
908 :
909 : /*
910 : * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
911 : */
912 : static void
913 35756 : RemoveProcFromArray(int code, Datum arg)
914 : {
915 : Assert(MyProc != NULL);
916 35756 : ProcArrayRemove(MyProc, InvalidTransactionId);
917 35756 : }
918 :
919 : /*
920 : * ProcKill() -- Destroy the per-proc data structure for
921 : * this process. Release any of its held LW locks.
922 : */
923 : static void
924 35774 : ProcKill(int code, Datum arg)
925 : {
926 : PGPROC *proc;
927 : dlist_head *procgloballist;
928 :
929 : Assert(MyProc != NULL);
930 :
931 : /* not safe if forked by system(), etc. */
932 35774 : if (MyProc->pid != (int) getpid())
933 0 : elog(PANIC, "ProcKill() called in child process");
934 :
935 : /* Make sure we're out of the sync rep lists */
936 35774 : SyncRepCleanupAtProcExit();
937 :
938 : #ifdef USE_ASSERT_CHECKING
939 : {
940 : int i;
941 :
942 : /* Last process should have released all locks. */
943 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
944 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
945 : }
946 : #endif
947 :
948 : /*
949 : * Release any LW locks I am holding. There really shouldn't be any, but
950 : * it's cheap to check again before we cut the knees off the LWLock
951 : * facility by releasing our PGPROC ...
952 : */
953 35774 : LWLockReleaseAll();
954 :
955 : /*
956 : * Cleanup waiting for LSN if any.
957 : */
958 35774 : WaitLSNCleanup();
959 :
960 : /* Cancel any pending condition variable sleep, too */
961 35774 : ConditionVariableCancelSleep();
962 :
963 : /*
964 : * Detach from any lock group of which we are a member. If the leader
965 : * exits before all other group members, its PGPROC will remain allocated
966 : * until the last group process exits; that process must return the
967 : * leader's PGPROC to the appropriate list.
968 : */
969 35774 : if (MyProc->lockGroupLeader != NULL)
970 : {
971 3132 : PGPROC *leader = MyProc->lockGroupLeader;
972 3132 : LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
973 :
974 3132 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
975 : Assert(!dlist_is_empty(&leader->lockGroupMembers));
976 3132 : dlist_delete(&MyProc->lockGroupLink);
977 3132 : if (dlist_is_empty(&leader->lockGroupMembers))
978 : {
979 164 : leader->lockGroupLeader = NULL;
980 164 : if (leader != MyProc)
981 : {
982 0 : procgloballist = leader->procgloballist;
983 :
984 : /* Leader exited first; return its PGPROC. */
985 0 : SpinLockAcquire(ProcStructLock);
986 0 : dlist_push_head(procgloballist, &leader->links);
987 0 : SpinLockRelease(ProcStructLock);
988 : }
989 : }
990 2968 : else if (leader != MyProc)
991 2968 : MyProc->lockGroupLeader = NULL;
992 3132 : LWLockRelease(leader_lwlock);
993 : }
994 :
995 : /*
996 : * Reset MyLatch to the process local one. This is so that signal
997 : * handlers et al can continue using the latch after the shared latch
998 : * isn't ours anymore.
999 : *
1000 : * Similarly, stop reporting wait events to MyProc->wait_event_info.
1001 : *
1002 : * After that clear MyProc and disown the shared latch.
1003 : */
1004 35774 : SwitchBackToLocalLatch();
1005 35774 : pgstat_reset_wait_event_storage();
1006 :
1007 35774 : proc = MyProc;
1008 35774 : MyProc = NULL;
1009 35774 : MyProcNumber = INVALID_PROC_NUMBER;
1010 35774 : DisownLatch(&proc->procLatch);
1011 :
1012 : /* Mark the proc no longer in use */
1013 35774 : proc->pid = 0;
1014 35774 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1015 35774 : proc->vxid.lxid = InvalidTransactionId;
1016 :
1017 35774 : procgloballist = proc->procgloballist;
1018 35774 : SpinLockAcquire(ProcStructLock);
1019 :
1020 : /*
1021 : * If we're still a member of a locking group, that means we're a leader
1022 : * which has somehow exited before its children. The last remaining child
1023 : * will release our PGPROC. Otherwise, release it now.
1024 : */
1025 35774 : if (proc->lockGroupLeader == NULL)
1026 : {
1027 : /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
1028 : Assert(dlist_is_empty(&proc->lockGroupMembers));
1029 :
1030 : /* Return PGPROC structure (and semaphore) to appropriate freelist */
1031 35774 : dlist_push_tail(procgloballist, &proc->links);
1032 : }
1033 :
1034 : /* Update shared estimate of spins_per_delay */
1035 35774 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1036 :
1037 35774 : SpinLockRelease(ProcStructLock);
1038 35774 : }
1039 :
1040 : /*
1041 : * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
1042 : * processes (bgwriter, etc). The PGPROC and sema are not released, only
1043 : * marked as not-in-use.
1044 : */
1045 : static void
1046 9078 : AuxiliaryProcKill(int code, Datum arg)
1047 : {
1048 9078 : int proctype = DatumGetInt32(arg);
1049 : PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
1050 : PGPROC *proc;
1051 :
1052 : Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
1053 :
1054 : /* not safe if forked by system(), etc. */
1055 9078 : if (MyProc->pid != (int) getpid())
1056 0 : elog(PANIC, "AuxiliaryProcKill() called in child process");
1057 :
1058 9078 : auxproc = &AuxiliaryProcs[proctype];
1059 :
1060 : Assert(MyProc == auxproc);
1061 :
1062 : /* Release any LW locks I am holding (see notes above) */
1063 9078 : LWLockReleaseAll();
1064 :
1065 : /* Cancel any pending condition variable sleep, too */
1066 9078 : ConditionVariableCancelSleep();
1067 :
1068 : /* look at the equivalent ProcKill() code for comments */
1069 9078 : SwitchBackToLocalLatch();
1070 9078 : pgstat_reset_wait_event_storage();
1071 :
1072 9078 : proc = MyProc;
1073 9078 : MyProc = NULL;
1074 9078 : MyProcNumber = INVALID_PROC_NUMBER;
1075 9078 : DisownLatch(&proc->procLatch);
1076 :
1077 9078 : SpinLockAcquire(ProcStructLock);
1078 :
1079 : /* Mark auxiliary proc no longer in use */
1080 9078 : proc->pid = 0;
1081 9078 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1082 9078 : proc->vxid.lxid = InvalidTransactionId;
1083 :
1084 : /* Update shared estimate of spins_per_delay */
1085 9078 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1086 :
1087 9078 : SpinLockRelease(ProcStructLock);
1088 9078 : }
1089 :
1090 : /*
1091 : * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
1092 : * given its PID
1093 : *
1094 : * Returns NULL if not found.
1095 : */
1096 : PGPROC *
1097 11336 : AuxiliaryPidGetProc(int pid)
1098 : {
1099 11336 : PGPROC *result = NULL;
1100 : int index;
1101 :
1102 11336 : if (pid == 0) /* never match dummy PGPROCs */
1103 6 : return NULL;
1104 :
1105 53040 : for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
1106 : {
1107 53040 : PGPROC *proc = &AuxiliaryProcs[index];
1108 :
1109 53040 : if (proc->pid == pid)
1110 : {
1111 11330 : result = proc;
1112 11330 : break;
1113 : }
1114 : }
1115 11330 : return result;
1116 : }
1117 :
1118 :
1119 : /*
1120 : * JoinWaitQueue -- join the wait queue on the specified lock
1121 : *
1122 : * It's not actually guaranteed that we need to wait when this function is
1123 : * called, because it could be that when we try to find a position at which
1124 : * to insert ourself into the wait queue, we discover that we must be inserted
1125 : * ahead of everyone who wants a lock that conflict with ours. In that case,
1126 : * we get the lock immediately. Because of this, it's sensible for this function
1127 : * to have a dontWait argument, despite the name.
1128 : *
1129 : * On entry, the caller has already set up LOCK and PROCLOCK entries to
1130 : * reflect that we have "requested" the lock. The caller is responsible for
1131 : * cleaning that up, if we end up not joining the queue after all.
1132 : *
1133 : * The lock table's partition lock must be held at entry, and is still held
1134 : * at exit. The caller must release it before calling ProcSleep().
1135 : *
1136 : * Result is one of the following:
1137 : *
1138 : * PROC_WAIT_STATUS_OK - lock was immediately granted
1139 : * PROC_WAIT_STATUS_WAITING - joined the wait queue; call ProcSleep()
1140 : * PROC_WAIT_STATUS_ERROR - immediate deadlock was detected, or would
1141 : * need to wait and dontWait == true
1142 : *
1143 : * NOTES: The process queue is now a priority queue for locking.
1144 : */
1145 : ProcWaitStatus
1146 4210 : JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
1147 : {
1148 4210 : LOCKMODE lockmode = locallock->tag.mode;
1149 4210 : LOCK *lock = locallock->lock;
1150 4210 : PROCLOCK *proclock = locallock->proclock;
1151 4210 : uint32 hashcode = locallock->hashcode;
1152 4210 : LWLock *partitionLock PG_USED_FOR_ASSERTS_ONLY = LockHashPartitionLock(hashcode);
1153 4210 : dclist_head *waitQueue = &lock->waitProcs;
1154 4210 : PGPROC *insert_before = NULL;
1155 : LOCKMASK myProcHeldLocks;
1156 : LOCKMASK myHeldLocks;
1157 4210 : bool early_deadlock = false;
1158 4210 : PGPROC *leader = MyProc->lockGroupLeader;
1159 :
1160 : Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
1161 :
1162 : /*
1163 : * Set bitmask of locks this process already holds on this object.
1164 : */
1165 4210 : myHeldLocks = MyProc->heldLocks = proclock->holdMask;
1166 :
1167 : /*
1168 : * Determine which locks we're already holding.
1169 : *
1170 : * If group locking is in use, locks held by members of my locking group
1171 : * need to be included in myHeldLocks. This is not required for relation
1172 : * extension lock which conflict among group members. However, including
1173 : * them in myHeldLocks will give group members the priority to get those
1174 : * locks as compared to other backends which are also trying to acquire
1175 : * those locks. OTOH, we can avoid giving priority to group members for
1176 : * that kind of locks, but there doesn't appear to be a clear advantage of
1177 : * the same.
1178 : */
1179 4210 : myProcHeldLocks = proclock->holdMask;
1180 4210 : myHeldLocks = myProcHeldLocks;
1181 4210 : if (leader != NULL)
1182 : {
1183 : dlist_iter iter;
1184 :
1185 66 : dlist_foreach(iter, &lock->procLocks)
1186 : {
1187 : PROCLOCK *otherproclock;
1188 :
1189 50 : otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
1190 :
1191 50 : if (otherproclock->groupLeader == leader)
1192 24 : myHeldLocks |= otherproclock->holdMask;
1193 : }
1194 : }
1195 :
1196 : /*
1197 : * Determine where to add myself in the wait queue.
1198 : *
1199 : * Normally I should go at the end of the queue. However, if I already
1200 : * hold locks that conflict with the request of any previous waiter, put
1201 : * myself in the queue just in front of the first such waiter. This is not
1202 : * a necessary step, since deadlock detection would move me to before that
1203 : * waiter anyway; but it's relatively cheap to detect such a conflict
1204 : * immediately, and avoid delaying till deadlock timeout.
1205 : *
1206 : * Special case: if I find I should go in front of some waiter, check to
1207 : * see if I conflict with already-held locks or the requests before that
1208 : * waiter. If not, then just grant myself the requested lock immediately.
1209 : * This is the same as the test for immediate grant in LockAcquire, except
1210 : * we are only considering the part of the wait queue before my insertion
1211 : * point.
1212 : */
1213 4210 : if (myHeldLocks != 0 && !dclist_is_empty(waitQueue))
1214 : {
1215 12 : LOCKMASK aheadRequests = 0;
1216 : dlist_iter iter;
1217 :
1218 12 : dclist_foreach(iter, waitQueue)
1219 : {
1220 12 : PGPROC *proc = dlist_container(PGPROC, links, iter.cur);
1221 :
1222 : /*
1223 : * If we're part of the same locking group as this waiter, its
1224 : * locks neither conflict with ours nor contribute to
1225 : * aheadRequests.
1226 : */
1227 12 : if (leader != NULL && leader == proc->lockGroupLeader)
1228 0 : continue;
1229 :
1230 : /* Must he wait for me? */
1231 12 : if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1232 : {
1233 : /* Must I wait for him ? */
1234 12 : if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1235 : {
1236 : /*
1237 : * Yes, so we have a deadlock. Easiest way to clean up
1238 : * correctly is to call RemoveFromWaitQueue(), but we
1239 : * can't do that until we are *on* the wait queue. So, set
1240 : * a flag to check below, and break out of loop. Also,
1241 : * record deadlock info for later message.
1242 : */
1243 2 : RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1244 2 : early_deadlock = true;
1245 2 : break;
1246 : }
1247 : /* I must go before this waiter. Check special case. */
1248 10 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1249 10 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1250 : proclock))
1251 : {
1252 : /* Skip the wait and just grant myself the lock. */
1253 10 : GrantLock(lock, proclock, lockmode);
1254 10 : return PROC_WAIT_STATUS_OK;
1255 : }
1256 :
1257 : /* Put myself into wait queue before conflicting process */
1258 0 : insert_before = proc;
1259 0 : break;
1260 : }
1261 : /* Nope, so advance to next waiter */
1262 0 : aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1263 : }
1264 : }
1265 :
1266 : /*
1267 : * If we detected deadlock, give up without waiting. This must agree with
1268 : * CheckDeadLock's recovery code.
1269 : */
1270 4200 : if (early_deadlock)
1271 2 : return PROC_WAIT_STATUS_ERROR;
1272 :
1273 : /*
1274 : * At this point we know that we'd really need to sleep. If we've been
1275 : * commanded not to do that, bail out.
1276 : */
1277 4198 : if (dontWait)
1278 1448 : return PROC_WAIT_STATUS_ERROR;
1279 :
1280 : /*
1281 : * Insert self into queue, at the position determined above.
1282 : */
1283 2750 : if (insert_before)
1284 0 : dclist_insert_before(waitQueue, &insert_before->links, &MyProc->links);
1285 : else
1286 2750 : dclist_push_tail(waitQueue, &MyProc->links);
1287 :
1288 2750 : lock->waitMask |= LOCKBIT_ON(lockmode);
1289 :
1290 : /* Set up wait information in PGPROC object, too */
1291 2750 : MyProc->heldLocks = myProcHeldLocks;
1292 2750 : MyProc->waitLock = lock;
1293 2750 : MyProc->waitProcLock = proclock;
1294 2750 : MyProc->waitLockMode = lockmode;
1295 :
1296 2750 : MyProc->waitStatus = PROC_WAIT_STATUS_WAITING;
1297 :
1298 2750 : return PROC_WAIT_STATUS_WAITING;
1299 : }
1300 :
1301 : /*
1302 : * ProcSleep -- put process to sleep waiting on lock
1303 : *
1304 : * This must be called when JoinWaitQueue() returns PROC_WAIT_STATUS_WAITING.
1305 : * Returns after the lock has been granted, or if a deadlock is detected. Can
1306 : * also bail out with ereport(ERROR), if some other error condition, or a
1307 : * timeout or cancellation is triggered.
1308 : *
1309 : * Result is one of the following:
1310 : *
1311 : * PROC_WAIT_STATUS_OK - lock was granted
1312 : * PROC_WAIT_STATUS_ERROR - a deadlock was detected
1313 : */
1314 : ProcWaitStatus
1315 2750 : ProcSleep(LOCALLOCK *locallock)
1316 : {
1317 2750 : LOCKMODE lockmode = locallock->tag.mode;
1318 2750 : LOCK *lock = locallock->lock;
1319 2750 : uint32 hashcode = locallock->hashcode;
1320 2750 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1321 2750 : TimestampTz standbyWaitStart = 0;
1322 2750 : bool allow_autovacuum_cancel = true;
1323 2750 : bool logged_recovery_conflict = false;
1324 : ProcWaitStatus myWaitStatus;
1325 :
1326 : /* The caller must've armed the on-error cleanup mechanism */
1327 : Assert(GetAwaitedLock() == locallock);
1328 : Assert(!LWLockHeldByMe(partitionLock));
1329 :
1330 : /*
1331 : * Now that we will successfully clean up after an ereport, it's safe to
1332 : * check to see if there's a buffer pin deadlock against the Startup
1333 : * process. Of course, that's only necessary if we're doing Hot Standby
1334 : * and are not the Startup process ourselves.
1335 : */
1336 2750 : if (RecoveryInProgress() && !InRecovery)
1337 2 : CheckRecoveryConflictDeadlock();
1338 :
1339 : /* Reset deadlock_state before enabling the timeout handler */
1340 2750 : deadlock_state = DS_NOT_YET_CHECKED;
1341 2750 : got_deadlock_timeout = false;
1342 :
1343 : /*
1344 : * Set timer so we can wake up after awhile and check for a deadlock. If a
1345 : * deadlock is detected, the handler sets MyProc->waitStatus =
1346 : * PROC_WAIT_STATUS_ERROR, allowing us to know that we must report failure
1347 : * rather than success.
1348 : *
1349 : * By delaying the check until we've waited for a bit, we can avoid
1350 : * running the rather expensive deadlock-check code in most cases.
1351 : *
1352 : * If LockTimeout is set, also enable the timeout for that. We can save a
1353 : * few cycles by enabling both timeout sources in one call.
1354 : *
1355 : * If InHotStandby we set lock waits slightly later for clarity with other
1356 : * code.
1357 : */
1358 2750 : if (!InHotStandby)
1359 : {
1360 2748 : if (LockTimeout > 0)
1361 : {
1362 : EnableTimeoutParams timeouts[2];
1363 :
1364 210 : timeouts[0].id = DEADLOCK_TIMEOUT;
1365 210 : timeouts[0].type = TMPARAM_AFTER;
1366 210 : timeouts[0].delay_ms = DeadlockTimeout;
1367 210 : timeouts[1].id = LOCK_TIMEOUT;
1368 210 : timeouts[1].type = TMPARAM_AFTER;
1369 210 : timeouts[1].delay_ms = LockTimeout;
1370 210 : enable_timeouts(timeouts, 2);
1371 : }
1372 : else
1373 2538 : enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1374 :
1375 : /*
1376 : * Use the current time obtained for the deadlock timeout timer as
1377 : * waitStart (i.e., the time when this process started waiting for the
1378 : * lock). Since getting the current time newly can cause overhead, we
1379 : * reuse the already-obtained time to avoid that overhead.
1380 : *
1381 : * Note that waitStart is updated without holding the lock table's
1382 : * partition lock, to avoid the overhead by additional lock
1383 : * acquisition. This can cause "waitstart" in pg_locks to become NULL
1384 : * for a very short period of time after the wait started even though
1385 : * "granted" is false. This is OK in practice because we can assume
1386 : * that users are likely to look at "waitstart" when waiting for the
1387 : * lock for a long time.
1388 : */
1389 2748 : pg_atomic_write_u64(&MyProc->waitStart,
1390 2748 : get_timeout_start_time(DEADLOCK_TIMEOUT));
1391 : }
1392 2 : else if (log_recovery_conflict_waits)
1393 : {
1394 : /*
1395 : * Set the wait start timestamp if logging is enabled and in hot
1396 : * standby.
1397 : */
1398 2 : standbyWaitStart = GetCurrentTimestamp();
1399 : }
1400 :
1401 : /*
1402 : * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1403 : * will not wait. But a set latch does not necessarily mean that the lock
1404 : * is free now, as there are many other sources for latch sets than
1405 : * somebody releasing the lock.
1406 : *
1407 : * We process interrupts whenever the latch has been set, so cancel/die
1408 : * interrupts are processed quickly. This means we must not mind losing
1409 : * control to a cancel/die interrupt here. We don't, because we have no
1410 : * shared-state-change work to do after being granted the lock (the
1411 : * grantor did it all). We do have to worry about canceling the deadlock
1412 : * timeout and updating the locallock table, but if we lose control to an
1413 : * error, LockErrorCleanup will fix that up.
1414 : */
1415 : do
1416 : {
1417 4852 : if (InHotStandby)
1418 : {
1419 6 : bool maybe_log_conflict =
1420 6 : (standbyWaitStart != 0 && !logged_recovery_conflict);
1421 :
1422 : /* Set a timer and wait for that or for the lock to be granted */
1423 6 : ResolveRecoveryConflictWithLock(locallock->tag.lock,
1424 : maybe_log_conflict);
1425 :
1426 : /*
1427 : * Emit the log message if the startup process is waiting longer
1428 : * than deadlock_timeout for recovery conflict on lock.
1429 : */
1430 6 : if (maybe_log_conflict)
1431 : {
1432 2 : TimestampTz now = GetCurrentTimestamp();
1433 :
1434 2 : if (TimestampDifferenceExceeds(standbyWaitStart, now,
1435 : DeadlockTimeout))
1436 : {
1437 : VirtualTransactionId *vxids;
1438 : int cnt;
1439 :
1440 2 : vxids = GetLockConflicts(&locallock->tag.lock,
1441 : AccessExclusiveLock, &cnt);
1442 :
1443 : /*
1444 : * Log the recovery conflict and the list of PIDs of
1445 : * backends holding the conflicting lock. Note that we do
1446 : * logging even if there are no such backends right now
1447 : * because the startup process here has already waited
1448 : * longer than deadlock_timeout.
1449 : */
1450 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1451 : standbyWaitStart, now,
1452 2 : cnt > 0 ? vxids : NULL, true);
1453 2 : logged_recovery_conflict = true;
1454 : }
1455 : }
1456 : }
1457 : else
1458 : {
1459 4846 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1460 4846 : PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1461 4846 : ResetLatch(MyLatch);
1462 : /* check for deadlocks first, as that's probably log-worthy */
1463 4846 : if (got_deadlock_timeout)
1464 : {
1465 52 : CheckDeadLock();
1466 52 : got_deadlock_timeout = false;
1467 : }
1468 4846 : CHECK_FOR_INTERRUPTS();
1469 : }
1470 :
1471 : /*
1472 : * waitStatus could change from PROC_WAIT_STATUS_WAITING to something
1473 : * else asynchronously. Read it just once per loop to prevent
1474 : * surprising behavior (such as missing log messages).
1475 : */
1476 4770 : myWaitStatus = *((volatile ProcWaitStatus *) &MyProc->waitStatus);
1477 :
1478 : /*
1479 : * If we are not deadlocked, but are waiting on an autovacuum-induced
1480 : * task, send a signal to interrupt it.
1481 : */
1482 4770 : if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1483 : {
1484 0 : PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1485 : uint8 statusFlags;
1486 : uint8 lockmethod_copy;
1487 : LOCKTAG locktag_copy;
1488 :
1489 : /*
1490 : * Grab info we need, then release lock immediately. Note this
1491 : * coding means that there is a tiny chance that the process
1492 : * terminates its current transaction and starts a different one
1493 : * before we have a change to send the signal; the worst possible
1494 : * consequence is that a for-wraparound vacuum is canceled. But
1495 : * that could happen in any case unless we were to do kill() with
1496 : * the lock held, which is much more undesirable.
1497 : */
1498 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1499 0 : statusFlags = ProcGlobal->statusFlags[autovac->pgxactoff];
1500 0 : lockmethod_copy = lock->tag.locktag_lockmethodid;
1501 0 : locktag_copy = lock->tag;
1502 0 : LWLockRelease(ProcArrayLock);
1503 :
1504 : /*
1505 : * Only do it if the worker is not working to protect against Xid
1506 : * wraparound.
1507 : */
1508 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
1509 0 : !(statusFlags & PROC_VACUUM_FOR_WRAPAROUND))
1510 : {
1511 0 : int pid = autovac->pid;
1512 :
1513 : /* report the case, if configured to do so */
1514 0 : if (message_level_is_interesting(DEBUG1))
1515 : {
1516 : StringInfoData locktagbuf;
1517 : StringInfoData logbuf; /* errdetail for server log */
1518 :
1519 0 : initStringInfo(&locktagbuf);
1520 0 : initStringInfo(&logbuf);
1521 0 : DescribeLockTag(&locktagbuf, &locktag_copy);
1522 0 : appendStringInfo(&logbuf,
1523 : "Process %d waits for %s on %s.",
1524 : MyProcPid,
1525 : GetLockmodeName(lockmethod_copy, lockmode),
1526 : locktagbuf.data);
1527 :
1528 0 : ereport(DEBUG1,
1529 : (errmsg_internal("sending cancel to blocking autovacuum PID %d",
1530 : pid),
1531 : errdetail_log("%s", logbuf.data)));
1532 :
1533 0 : pfree(locktagbuf.data);
1534 0 : pfree(logbuf.data);
1535 : }
1536 :
1537 : /* send the autovacuum worker Back to Old Kent Road */
1538 0 : if (kill(pid, SIGINT) < 0)
1539 : {
1540 : /*
1541 : * There's a race condition here: once we release the
1542 : * ProcArrayLock, it's possible for the autovac worker to
1543 : * close up shop and exit before we can do the kill().
1544 : * Therefore, we do not whinge about no-such-process.
1545 : * Other errors such as EPERM could conceivably happen if
1546 : * the kernel recycles the PID fast enough, but such cases
1547 : * seem improbable enough that it's probably best to issue
1548 : * a warning if we see some other errno.
1549 : */
1550 0 : if (errno != ESRCH)
1551 0 : ereport(WARNING,
1552 : (errmsg("could not send signal to process %d: %m",
1553 : pid)));
1554 : }
1555 : }
1556 :
1557 : /* prevent signal from being sent again more than once */
1558 0 : allow_autovacuum_cancel = false;
1559 : }
1560 :
1561 : /*
1562 : * If awoken after the deadlock check interrupt has run, and
1563 : * log_lock_waits is on, then report about the wait.
1564 : */
1565 4770 : if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1566 : {
1567 : StringInfoData buf,
1568 : lock_waiters_sbuf,
1569 : lock_holders_sbuf;
1570 : const char *modename;
1571 : long secs;
1572 : int usecs;
1573 : long msecs;
1574 1874 : int lockHoldersNum = 0;
1575 :
1576 1874 : initStringInfo(&buf);
1577 1874 : initStringInfo(&lock_waiters_sbuf);
1578 1874 : initStringInfo(&lock_holders_sbuf);
1579 :
1580 1874 : DescribeLockTag(&buf, &locallock->tag.lock);
1581 1874 : modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1582 : lockmode);
1583 1874 : TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1584 : GetCurrentTimestamp(),
1585 : &secs, &usecs);
1586 1874 : msecs = secs * 1000 + usecs / 1000;
1587 1874 : usecs = usecs % 1000;
1588 :
1589 : /* Gather a list of all lock holders and waiters */
1590 1874 : LWLockAcquire(partitionLock, LW_SHARED);
1591 1874 : GetLockHoldersAndWaiters(locallock, &lock_holders_sbuf,
1592 : &lock_waiters_sbuf, &lockHoldersNum);
1593 1874 : LWLockRelease(partitionLock);
1594 :
1595 1874 : if (deadlock_state == DS_SOFT_DEADLOCK)
1596 6 : ereport(LOG,
1597 : (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1598 : MyProcPid, modename, buf.data, msecs, usecs),
1599 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1600 : "Processes holding the lock: %s. Wait queue: %s.",
1601 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1602 1868 : else if (deadlock_state == DS_HARD_DEADLOCK)
1603 : {
1604 : /*
1605 : * This message is a bit redundant with the error that will be
1606 : * reported subsequently, but in some cases the error report
1607 : * might not make it to the log (eg, if it's caught by an
1608 : * exception handler), and we want to ensure all long-wait
1609 : * events get logged.
1610 : */
1611 10 : ereport(LOG,
1612 : (errmsg("process %d detected deadlock while 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 : }
1618 :
1619 1874 : if (myWaitStatus == PROC_WAIT_STATUS_WAITING)
1620 1830 : ereport(LOG,
1621 : (errmsg("process %d still waiting for %s on %s after %ld.%03d ms",
1622 : MyProcPid, modename, buf.data, msecs, usecs),
1623 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1624 : "Processes holding the lock: %s. Wait queue: %s.",
1625 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1626 44 : else if (myWaitStatus == PROC_WAIT_STATUS_OK)
1627 34 : ereport(LOG,
1628 : (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1629 : MyProcPid, modename, buf.data, msecs, usecs)));
1630 : else
1631 : {
1632 : Assert(myWaitStatus == PROC_WAIT_STATUS_ERROR);
1633 :
1634 : /*
1635 : * Currently, the deadlock checker always kicks its own
1636 : * process, which means that we'll only see
1637 : * PROC_WAIT_STATUS_ERROR when deadlock_state ==
1638 : * DS_HARD_DEADLOCK, and there's no need to print redundant
1639 : * messages. But for completeness and future-proofing, print
1640 : * a message if it looks like someone else kicked us off the
1641 : * lock.
1642 : */
1643 10 : if (deadlock_state != DS_HARD_DEADLOCK)
1644 0 : ereport(LOG,
1645 : (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1646 : MyProcPid, modename, buf.data, msecs, usecs),
1647 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1648 : "Processes holding the lock: %s. Wait queue: %s.",
1649 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1650 : }
1651 :
1652 : /*
1653 : * At this point we might still need to wait for the lock. Reset
1654 : * state so we don't print the above messages again.
1655 : */
1656 1874 : deadlock_state = DS_NO_DEADLOCK;
1657 :
1658 1874 : pfree(buf.data);
1659 1874 : pfree(lock_holders_sbuf.data);
1660 1874 : pfree(lock_waiters_sbuf.data);
1661 : }
1662 4770 : } while (myWaitStatus == PROC_WAIT_STATUS_WAITING);
1663 :
1664 : /*
1665 : * Disable the timers, if they are still running. As in LockErrorCleanup,
1666 : * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1667 : * already caused QueryCancelPending to become set, we want the cancel to
1668 : * be reported as a lock timeout, not a user cancel.
1669 : */
1670 2668 : if (!InHotStandby)
1671 : {
1672 2666 : if (LockTimeout > 0)
1673 : {
1674 : DisableTimeoutParams timeouts[2];
1675 :
1676 198 : timeouts[0].id = DEADLOCK_TIMEOUT;
1677 198 : timeouts[0].keep_indicator = false;
1678 198 : timeouts[1].id = LOCK_TIMEOUT;
1679 198 : timeouts[1].keep_indicator = true;
1680 198 : disable_timeouts(timeouts, 2);
1681 : }
1682 : else
1683 2468 : disable_timeout(DEADLOCK_TIMEOUT, false);
1684 : }
1685 :
1686 : /*
1687 : * Emit the log message if recovery conflict on lock was resolved but the
1688 : * startup process waited longer than deadlock_timeout for it.
1689 : */
1690 2668 : if (InHotStandby && logged_recovery_conflict)
1691 2 : LogRecoveryConflict(PROCSIG_RECOVERY_CONFLICT_LOCK,
1692 : standbyWaitStart, GetCurrentTimestamp(),
1693 : NULL, false);
1694 :
1695 : /*
1696 : * We don't have to do anything else, because the awaker did all the
1697 : * necessary updates of the lock table and MyProc. (The caller is
1698 : * responsible for updating the local lock table.)
1699 : */
1700 2668 : return myWaitStatus;
1701 : }
1702 :
1703 :
1704 : /*
1705 : * ProcWakeup -- wake up a process by setting its latch.
1706 : *
1707 : * Also remove the process from the wait queue and set its links invalid.
1708 : *
1709 : * The appropriate lock partition lock must be held by caller.
1710 : *
1711 : * XXX: presently, this code is only used for the "success" case, and only
1712 : * works correctly for that case. To clean up in failure case, would need
1713 : * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1714 : * Hence, in practice the waitStatus parameter must be PROC_WAIT_STATUS_OK.
1715 : */
1716 : void
1717 2662 : ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus)
1718 : {
1719 2662 : if (dlist_node_is_detached(&proc->links))
1720 0 : return;
1721 :
1722 : Assert(proc->waitStatus == PROC_WAIT_STATUS_WAITING);
1723 :
1724 : /* Remove process from wait queue */
1725 2662 : dclist_delete_from_thoroughly(&proc->waitLock->waitProcs, &proc->links);
1726 :
1727 : /* Clean up process' state and pass it the ok/fail signal */
1728 2662 : proc->waitLock = NULL;
1729 2662 : proc->waitProcLock = NULL;
1730 2662 : proc->waitStatus = waitStatus;
1731 2662 : pg_atomic_write_u64(&MyProc->waitStart, 0);
1732 :
1733 : /* And awaken it */
1734 2662 : SetLatch(&proc->procLatch);
1735 : }
1736 :
1737 : /*
1738 : * ProcLockWakeup -- routine for waking up processes when a lock is
1739 : * released (or a prior waiter is aborted). Scan all waiters
1740 : * for lock, waken any that are no longer blocked.
1741 : *
1742 : * The appropriate lock partition lock must be held by caller.
1743 : */
1744 : void
1745 2708 : ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1746 : {
1747 2708 : dclist_head *waitQueue = &lock->waitProcs;
1748 2708 : LOCKMASK aheadRequests = 0;
1749 : dlist_mutable_iter miter;
1750 :
1751 2708 : if (dclist_is_empty(waitQueue))
1752 88 : return;
1753 :
1754 6528 : dclist_foreach_modify(miter, waitQueue)
1755 : {
1756 3908 : PGPROC *proc = dlist_container(PGPROC, links, miter.cur);
1757 3908 : LOCKMODE lockmode = proc->waitLockMode;
1758 :
1759 : /*
1760 : * Waken if (a) doesn't conflict with requests of earlier waiters, and
1761 : * (b) doesn't conflict with already-held locks.
1762 : */
1763 3908 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1764 3174 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1765 : proc->waitProcLock))
1766 : {
1767 : /* OK to waken */
1768 2662 : GrantLock(lock, proc->waitProcLock, lockmode);
1769 : /* removes proc from the lock's waiting process queue */
1770 2662 : ProcWakeup(proc, PROC_WAIT_STATUS_OK);
1771 : }
1772 : else
1773 : {
1774 : /*
1775 : * Lock conflicts: Don't wake, but remember requested mode for
1776 : * later checks.
1777 : */
1778 1246 : aheadRequests |= LOCKBIT_ON(lockmode);
1779 : }
1780 : }
1781 : }
1782 :
1783 : /*
1784 : * CheckDeadLock
1785 : *
1786 : * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1787 : * lock to be released by some other process. Check if there's a deadlock; if
1788 : * not, just return. (But signal ProcSleep to log a message, if
1789 : * log_lock_waits is true.) If we have a real deadlock, remove ourselves from
1790 : * the lock's wait queue and signal an error to ProcSleep.
1791 : */
1792 : static void
1793 52 : CheckDeadLock(void)
1794 : {
1795 : int i;
1796 :
1797 : /*
1798 : * Acquire exclusive lock on the entire shared lock data structures. Must
1799 : * grab LWLocks in partition-number order to avoid LWLock deadlock.
1800 : *
1801 : * Note that the deadlock check interrupt had better not be enabled
1802 : * anywhere that this process itself holds lock partition locks, else this
1803 : * will wait forever. Also note that LWLockAcquire creates a critical
1804 : * section, so that this routine cannot be interrupted by cancel/die
1805 : * interrupts.
1806 : */
1807 884 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1808 832 : LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1809 :
1810 : /*
1811 : * Check to see if we've been awoken by anyone in the interim.
1812 : *
1813 : * If we have, we can return and resume our transaction -- happy day.
1814 : * Before we are awoken the process releasing the lock grants it to us so
1815 : * we know that we don't have to wait anymore.
1816 : *
1817 : * We check by looking to see if we've been unlinked from the wait queue.
1818 : * This is safe because we hold the lock partition lock.
1819 : */
1820 52 : if (MyProc->links.prev == NULL ||
1821 52 : MyProc->links.next == NULL)
1822 0 : goto check_done;
1823 :
1824 : #ifdef LOCK_DEBUG
1825 : if (Debug_deadlocks)
1826 : DumpAllLocks();
1827 : #endif
1828 :
1829 : /* Run the deadlock check, and set deadlock_state for use by ProcSleep */
1830 52 : deadlock_state = DeadLockCheck(MyProc);
1831 :
1832 52 : if (deadlock_state == DS_HARD_DEADLOCK)
1833 : {
1834 : /*
1835 : * Oops. We have a deadlock.
1836 : *
1837 : * Get this process out of wait state. (Note: we could do this more
1838 : * efficiently by relying on lockAwaited, but use this coding to
1839 : * preserve the flexibility to kill some other transaction than the
1840 : * one detecting the deadlock.)
1841 : *
1842 : * RemoveFromWaitQueue sets MyProc->waitStatus to
1843 : * PROC_WAIT_STATUS_ERROR, so ProcSleep will report an error after we
1844 : * return from the signal handler.
1845 : */
1846 : Assert(MyProc->waitLock != NULL);
1847 10 : RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1848 :
1849 : /*
1850 : * We're done here. Transaction abort caused by the error that
1851 : * ProcSleep will raise will cause any other locks we hold to be
1852 : * released, thus allowing other processes to wake up; we don't need
1853 : * to do that here. NOTE: an exception is that releasing locks we
1854 : * hold doesn't consider the possibility of waiters that were blocked
1855 : * behind us on the lock we just failed to get, and might now be
1856 : * wakable because we're not in front of them anymore. However,
1857 : * RemoveFromWaitQueue took care of waking up any such processes.
1858 : */
1859 : }
1860 :
1861 : /*
1862 : * And release locks. We do this in reverse order for two reasons: (1)
1863 : * Anyone else who needs more than one of the locks will be trying to lock
1864 : * them in increasing order; we don't want to release the other process
1865 : * until it can get all the locks it needs. (2) This avoids O(N^2)
1866 : * behavior inside LWLockRelease.
1867 : */
1868 42 : check_done:
1869 884 : for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1870 832 : LWLockRelease(LockHashPartitionLockByIndex(i));
1871 52 : }
1872 :
1873 : /*
1874 : * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1875 : *
1876 : * NB: Runs inside a signal handler, be careful.
1877 : */
1878 : void
1879 52 : CheckDeadLockAlert(void)
1880 : {
1881 52 : int save_errno = errno;
1882 :
1883 52 : got_deadlock_timeout = true;
1884 :
1885 : /*
1886 : * Have to set the latch again, even if handle_sig_alarm already did. Back
1887 : * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1888 : * ever would be a problem, but setting a set latch again is cheap.
1889 : *
1890 : * Note that, when this function runs inside procsignal_sigusr1_handler(),
1891 : * the handler function sets the latch again after the latch is set here.
1892 : */
1893 52 : SetLatch(MyLatch);
1894 52 : errno = save_errno;
1895 52 : }
1896 :
1897 : /*
1898 : * GetLockHoldersAndWaiters - get lock holders and waiters for a lock
1899 : *
1900 : * Fill lock_holders_sbuf and lock_waiters_sbuf with the PIDs of processes holding
1901 : * and waiting for the lock, and set lockHoldersNum to the number of lock holders.
1902 : *
1903 : * The lock table's partition lock must be held on entry and remains held on exit.
1904 : */
1905 : void
1906 1874 : GetLockHoldersAndWaiters(LOCALLOCK *locallock, StringInfo lock_holders_sbuf,
1907 : StringInfo lock_waiters_sbuf, int *lockHoldersNum)
1908 : {
1909 : dlist_iter proc_iter;
1910 : PROCLOCK *curproclock;
1911 1874 : LOCK *lock = locallock->lock;
1912 1874 : bool first_holder = true,
1913 1874 : first_waiter = true;
1914 :
1915 : #ifdef USE_ASSERT_CHECKING
1916 : {
1917 : uint32 hashcode = locallock->hashcode;
1918 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1919 :
1920 : Assert(LWLockHeldByMe(partitionLock));
1921 : }
1922 : #endif
1923 :
1924 1874 : *lockHoldersNum = 0;
1925 :
1926 : /*
1927 : * Loop over the lock's procLocks to gather a list of all holders and
1928 : * waiters. Thus we will be able to provide more detailed information for
1929 : * lock debugging purposes.
1930 : *
1931 : * lock->procLocks contains all processes which hold or wait for this
1932 : * lock.
1933 : */
1934 5622 : dlist_foreach(proc_iter, &lock->procLocks)
1935 : {
1936 3748 : curproclock =
1937 3748 : dlist_container(PROCLOCK, lockLink, proc_iter.cur);
1938 :
1939 : /*
1940 : * We are a waiter if myProc->waitProcLock == curproclock; we are a
1941 : * holder if it is NULL or something different.
1942 : */
1943 3748 : if (curproclock->tag.myProc->waitProcLock == curproclock)
1944 : {
1945 1860 : if (first_waiter)
1946 : {
1947 1832 : appendStringInfo(lock_waiters_sbuf, "%d",
1948 1832 : curproclock->tag.myProc->pid);
1949 1832 : first_waiter = false;
1950 : }
1951 : else
1952 28 : appendStringInfo(lock_waiters_sbuf, ", %d",
1953 28 : curproclock->tag.myProc->pid);
1954 : }
1955 : else
1956 : {
1957 1888 : if (first_holder)
1958 : {
1959 1874 : appendStringInfo(lock_holders_sbuf, "%d",
1960 1874 : curproclock->tag.myProc->pid);
1961 1874 : first_holder = false;
1962 : }
1963 : else
1964 14 : appendStringInfo(lock_holders_sbuf, ", %d",
1965 14 : curproclock->tag.myProc->pid);
1966 :
1967 1888 : (*lockHoldersNum)++;
1968 : }
1969 : }
1970 1874 : }
1971 :
1972 : /*
1973 : * ProcWaitForSignal - wait for a signal from another backend.
1974 : *
1975 : * As this uses the generic process latch the caller has to be robust against
1976 : * unrelated wakeups: Always check that the desired state has occurred, and
1977 : * wait again if not.
1978 : */
1979 : void
1980 176 : ProcWaitForSignal(uint32 wait_event_info)
1981 : {
1982 176 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1983 : wait_event_info);
1984 176 : ResetLatch(MyLatch);
1985 176 : CHECK_FOR_INTERRUPTS();
1986 176 : }
1987 :
1988 : /*
1989 : * ProcSendSignal - set the latch of a backend identified by ProcNumber
1990 : */
1991 : void
1992 146 : ProcSendSignal(ProcNumber procNumber)
1993 : {
1994 146 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
1995 0 : elog(ERROR, "procNumber out of range");
1996 :
1997 146 : SetLatch(&GetPGProcByNumber(procNumber)->procLatch);
1998 146 : }
1999 :
2000 : /*
2001 : * BecomeLockGroupLeader - designate process as lock group leader
2002 : *
2003 : * Once this function has returned, other processes can join the lock group
2004 : * by calling BecomeLockGroupMember.
2005 : */
2006 : void
2007 1262 : BecomeLockGroupLeader(void)
2008 : {
2009 : LWLock *leader_lwlock;
2010 :
2011 : /* If we already did it, we don't need to do it again. */
2012 1262 : if (MyProc->lockGroupLeader == MyProc)
2013 1098 : return;
2014 :
2015 : /* We had better not be a follower. */
2016 : Assert(MyProc->lockGroupLeader == NULL);
2017 :
2018 : /* Create single-member group, containing only ourselves. */
2019 164 : leader_lwlock = LockHashPartitionLockByProc(MyProc);
2020 164 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2021 164 : MyProc->lockGroupLeader = MyProc;
2022 164 : dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
2023 164 : LWLockRelease(leader_lwlock);
2024 : }
2025 :
2026 : /*
2027 : * BecomeLockGroupMember - designate process as lock group member
2028 : *
2029 : * This is pretty straightforward except for the possibility that the leader
2030 : * whose group we're trying to join might exit before we manage to do so;
2031 : * and the PGPROC might get recycled for an unrelated process. To avoid
2032 : * that, we require the caller to pass the PID of the intended PGPROC as
2033 : * an interlock. Returns true if we successfully join the intended lock
2034 : * group, and false if not.
2035 : */
2036 : bool
2037 2968 : BecomeLockGroupMember(PGPROC *leader, int pid)
2038 : {
2039 : LWLock *leader_lwlock;
2040 2968 : bool ok = false;
2041 :
2042 : /* Group leader can't become member of group */
2043 : Assert(MyProc != leader);
2044 :
2045 : /* Can't already be a member of a group */
2046 : Assert(MyProc->lockGroupLeader == NULL);
2047 :
2048 : /* PID must be valid. */
2049 : Assert(pid != 0);
2050 :
2051 : /*
2052 : * Get lock protecting the group fields. Note LockHashPartitionLockByProc
2053 : * calculates the proc number based on the PGPROC slot without looking at
2054 : * its contents, so we will acquire the correct lock even if the leader
2055 : * PGPROC is in process of being recycled.
2056 : */
2057 2968 : leader_lwlock = LockHashPartitionLockByProc(leader);
2058 2968 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2059 :
2060 : /* Is this the leader we're looking for? */
2061 2968 : if (leader->pid == pid && leader->lockGroupLeader == leader)
2062 : {
2063 : /* OK, join the group */
2064 2968 : ok = true;
2065 2968 : MyProc->lockGroupLeader = leader;
2066 2968 : dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
2067 : }
2068 2968 : LWLockRelease(leader_lwlock);
2069 :
2070 2968 : return ok;
2071 : }
|
