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 : /* Pointers to shared-memory structures */
70 : PROC_HDR *ProcGlobal = NULL;
71 : NON_EXEC_STATIC PGPROC *AuxiliaryProcs = NULL;
72 : PGPROC *PreparedXactProcs = NULL;
73 :
74 : /* Is a deadlock check pending? */
75 : static volatile sig_atomic_t got_deadlock_timeout;
76 :
77 : static void RemoveProcFromArray(int code, Datum arg);
78 : static void ProcKill(int code, Datum arg);
79 : static void AuxiliaryProcKill(int code, Datum arg);
80 : static DeadLockState CheckDeadLock(void);
81 :
82 :
83 : /*
84 : * Report shared-memory space needed by PGPROC.
85 : */
86 : static Size
87 3267 : PGProcShmemSize(void)
88 : {
89 3267 : Size size = 0;
90 : Size TotalProcs =
91 3267 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
92 :
93 3267 : size = add_size(size, mul_size(TotalProcs, sizeof(PGPROC)));
94 3267 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->xids)));
95 3267 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->subxidStates)));
96 3267 : size = add_size(size, mul_size(TotalProcs, sizeof(*ProcGlobal->statusFlags)));
97 :
98 3267 : return size;
99 : }
100 :
101 : /*
102 : * Report shared-memory space needed by Fast-Path locks.
103 : */
104 : static Size
105 3267 : FastPathLockShmemSize(void)
106 : {
107 3267 : Size size = 0;
108 : Size TotalProcs =
109 3267 : add_size(MaxBackends, add_size(NUM_AUXILIARY_PROCS, max_prepared_xacts));
110 : Size fpLockBitsSize,
111 : fpRelIdSize;
112 :
113 : /*
114 : * Memory needed for PGPROC fast-path lock arrays. Make sure the sizes are
115 : * nicely aligned in each backend.
116 : */
117 3267 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
118 3267 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
119 :
120 3267 : size = add_size(size, mul_size(TotalProcs, (fpLockBitsSize + fpRelIdSize)));
121 :
122 3267 : return size;
123 : }
124 :
125 : /*
126 : * Report shared-memory space needed by InitProcGlobal.
127 : */
128 : Size
129 2127 : ProcGlobalShmemSize(void)
130 : {
131 2127 : Size size = 0;
132 :
133 : /* ProcGlobal */
134 2127 : size = add_size(size, sizeof(PROC_HDR));
135 2127 : size = add_size(size, sizeof(slock_t));
136 :
137 2127 : size = add_size(size, PGSemaphoreShmemSize(ProcGlobalSemas()));
138 2127 : size = add_size(size, PGProcShmemSize());
139 2127 : size = add_size(size, FastPathLockShmemSize());
140 :
141 2127 : return size;
142 : }
143 :
144 : /*
145 : * Report number of semaphores needed by InitProcGlobal.
146 : */
147 : int
148 4252 : ProcGlobalSemas(void)
149 : {
150 : /*
151 : * We need a sema per backend (including autovacuum), plus one for each
152 : * auxiliary process.
153 : */
154 4252 : return MaxBackends + NUM_AUXILIARY_PROCS;
155 : }
156 :
157 : /*
158 : * InitProcGlobal -
159 : * Initialize the global process table during postmaster or standalone
160 : * backend startup.
161 : *
162 : * We also create all the per-process semaphores we will need to support
163 : * the requested number of backends. We used to allocate semaphores
164 : * only when backends were actually started up, but that is bad because
165 : * it lets Postgres fail under load --- a lot of Unix systems are
166 : * (mis)configured with small limits on the number of semaphores, and
167 : * running out when trying to start another backend is a common failure.
168 : * So, now we grab enough semaphores to support the desired max number
169 : * of backends immediately at initialization --- if the sysadmin has set
170 : * MaxConnections, max_worker_processes, max_wal_senders, or
171 : * autovacuum_worker_slots higher than his kernel will support, he'll
172 : * find out sooner rather than later.
173 : *
174 : * Another reason for creating semaphores here is that the semaphore
175 : * implementation typically requires us to create semaphores in the
176 : * postmaster, not in backends.
177 : *
178 : * Note: this is NOT called by individual backends under a postmaster,
179 : * not even in the EXEC_BACKEND case. The ProcGlobal and AuxiliaryProcs
180 : * pointers must be propagated specially for EXEC_BACKEND operation.
181 : */
182 : void
183 1140 : InitProcGlobal(void)
184 : {
185 : PGPROC *procs;
186 : int i,
187 : j;
188 : bool found;
189 1140 : uint32 TotalProcs = MaxBackends + NUM_AUXILIARY_PROCS + max_prepared_xacts;
190 :
191 : /* Used for setup of per-backend fast-path slots. */
192 : char *fpPtr,
193 : *fpEndPtr PG_USED_FOR_ASSERTS_ONLY;
194 : Size fpLockBitsSize,
195 : fpRelIdSize;
196 : Size requestSize;
197 : char *ptr;
198 :
199 : /* Create the ProcGlobal shared structure */
200 1140 : ProcGlobal = (PROC_HDR *)
201 1140 : ShmemInitStruct("Proc Header", sizeof(PROC_HDR), &found);
202 : Assert(!found);
203 :
204 : /*
205 : * Initialize the data structures.
206 : */
207 1140 : ProcGlobal->spins_per_delay = DEFAULT_SPINS_PER_DELAY;
208 1140 : SpinLockInit(&ProcGlobal->freeProcsLock);
209 1140 : dlist_init(&ProcGlobal->freeProcs);
210 1140 : dlist_init(&ProcGlobal->autovacFreeProcs);
211 1140 : dlist_init(&ProcGlobal->bgworkerFreeProcs);
212 1140 : dlist_init(&ProcGlobal->walsenderFreeProcs);
213 1140 : ProcGlobal->startupBufferPinWaitBufId = -1;
214 1140 : ProcGlobal->walwriterProc = INVALID_PROC_NUMBER;
215 1140 : ProcGlobal->checkpointerProc = INVALID_PROC_NUMBER;
216 1140 : pg_atomic_init_u32(&ProcGlobal->procArrayGroupFirst, INVALID_PROC_NUMBER);
217 1140 : pg_atomic_init_u32(&ProcGlobal->clogGroupFirst, INVALID_PROC_NUMBER);
218 :
219 : /*
220 : * Create and initialize all the PGPROC structures we'll need. There are
221 : * six separate consumers: (1) normal backends, (2) autovacuum workers and
222 : * special workers, (3) background workers, (4) walsenders, (5) auxiliary
223 : * processes, and (6) prepared transactions. (For largely-historical
224 : * reasons, we combine autovacuum and special workers into one category
225 : * with a single freelist.) Each PGPROC structure is dedicated to exactly
226 : * one of these purposes, and they do not move between groups.
227 : */
228 1140 : requestSize = PGProcShmemSize();
229 :
230 1140 : ptr = ShmemInitStruct("PGPROC structures",
231 : requestSize,
232 : &found);
233 :
234 1140 : MemSet(ptr, 0, requestSize);
235 :
236 1140 : procs = (PGPROC *) ptr;
237 1140 : ptr = ptr + TotalProcs * sizeof(PGPROC);
238 :
239 1140 : ProcGlobal->allProcs = procs;
240 : /* XXX allProcCount isn't really all of them; it excludes prepared xacts */
241 1140 : ProcGlobal->allProcCount = MaxBackends + NUM_AUXILIARY_PROCS;
242 :
243 : /*
244 : * Allocate arrays mirroring PGPROC fields in a dense manner. See
245 : * PROC_HDR.
246 : *
247 : * XXX: It might make sense to increase padding for these arrays, given
248 : * how hotly they are accessed.
249 : */
250 1140 : ProcGlobal->xids = (TransactionId *) ptr;
251 1140 : ptr = ptr + (TotalProcs * sizeof(*ProcGlobal->xids));
252 :
253 1140 : ProcGlobal->subxidStates = (XidCacheStatus *) ptr;
254 1140 : ptr = ptr + (TotalProcs * sizeof(*ProcGlobal->subxidStates));
255 :
256 1140 : ProcGlobal->statusFlags = (uint8 *) ptr;
257 1140 : ptr = ptr + (TotalProcs * sizeof(*ProcGlobal->statusFlags));
258 :
259 : /* make sure wer didn't overflow */
260 : Assert((ptr > (char *) procs) && (ptr <= (char *) procs + requestSize));
261 :
262 : /*
263 : * Allocate arrays for fast-path locks. Those are variable-length, so
264 : * can't be included in PGPROC directly. We allocate a separate piece of
265 : * shared memory and then divide that between backends.
266 : */
267 1140 : fpLockBitsSize = MAXALIGN(FastPathLockGroupsPerBackend * sizeof(uint64));
268 1140 : fpRelIdSize = MAXALIGN(FastPathLockSlotsPerBackend() * sizeof(Oid));
269 :
270 1140 : requestSize = FastPathLockShmemSize();
271 :
272 1140 : fpPtr = ShmemInitStruct("Fast-Path Lock Array",
273 : requestSize,
274 : &found);
275 :
276 1140 : MemSet(fpPtr, 0, requestSize);
277 :
278 : /* For asserts checking we did not overflow. */
279 1140 : fpEndPtr = fpPtr + requestSize;
280 :
281 : /* Reserve space for semaphores. */
282 1140 : PGReserveSemaphores(ProcGlobalSemas());
283 :
284 149926 : for (i = 0; i < TotalProcs; i++)
285 : {
286 148786 : PGPROC *proc = &procs[i];
287 :
288 : /* Common initialization for all PGPROCs, regardless of type. */
289 :
290 : /*
291 : * Set the fast-path lock arrays, and move the pointer. We interleave
292 : * the two arrays, to (hopefully) get some locality for each backend.
293 : */
294 148786 : proc->fpLockBits = (uint64 *) fpPtr;
295 148786 : fpPtr += fpLockBitsSize;
296 :
297 148786 : proc->fpRelId = (Oid *) fpPtr;
298 148786 : fpPtr += fpRelIdSize;
299 :
300 : Assert(fpPtr <= fpEndPtr);
301 :
302 : /*
303 : * Set up per-PGPROC semaphore, latch, and fpInfoLock. Prepared xact
304 : * dummy PGPROCs don't need these though - they're never associated
305 : * with a real process
306 : */
307 148786 : if (i < MaxBackends + NUM_AUXILIARY_PROCS)
308 : {
309 147920 : proc->sem = PGSemaphoreCreate();
310 147920 : InitSharedLatch(&(proc->procLatch));
311 147920 : LWLockInitialize(&(proc->fpInfoLock), LWTRANCHE_LOCK_FASTPATH);
312 : }
313 :
314 : /*
315 : * Newly created PGPROCs for normal backends, autovacuum workers,
316 : * special workers, bgworkers, and walsenders must be queued up on the
317 : * appropriate free list. Because there can only ever be a small,
318 : * fixed number of auxiliary processes, no free list is used in that
319 : * case; InitAuxiliaryProcess() instead uses a linear search. PGPROCs
320 : * for prepared transactions are added to a free list by
321 : * TwoPhaseShmemInit().
322 : */
323 148786 : if (i < MaxConnections)
324 : {
325 : /* PGPROC for normal backend, add to freeProcs list */
326 73342 : dlist_push_tail(&ProcGlobal->freeProcs, &proc->links);
327 73342 : proc->procgloballist = &ProcGlobal->freeProcs;
328 : }
329 75444 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS)
330 : {
331 : /* PGPROC for AV or special worker, add to autovacFreeProcs list */
332 14709 : dlist_push_tail(&ProcGlobal->autovacFreeProcs, &proc->links);
333 14709 : proc->procgloballist = &ProcGlobal->autovacFreeProcs;
334 : }
335 60735 : else if (i < MaxConnections + autovacuum_worker_slots + NUM_SPECIAL_WORKER_PROCS + max_worker_processes)
336 : {
337 : /* PGPROC for bgworker, add to bgworkerFreeProcs list */
338 9111 : dlist_push_tail(&ProcGlobal->bgworkerFreeProcs, &proc->links);
339 9111 : proc->procgloballist = &ProcGlobal->bgworkerFreeProcs;
340 : }
341 51624 : else if (i < MaxBackends)
342 : {
343 : /* PGPROC for walsender, add to walsenderFreeProcs list */
344 7438 : dlist_push_tail(&ProcGlobal->walsenderFreeProcs, &proc->links);
345 7438 : proc->procgloballist = &ProcGlobal->walsenderFreeProcs;
346 : }
347 :
348 : /* Initialize myProcLocks[] shared memory queues. */
349 2529362 : for (j = 0; j < NUM_LOCK_PARTITIONS; j++)
350 2380576 : dlist_init(&(proc->myProcLocks[j]));
351 :
352 : /* Initialize lockGroupMembers list. */
353 148786 : dlist_init(&proc->lockGroupMembers);
354 :
355 : /*
356 : * Initialize the atomic variables, otherwise, it won't be safe to
357 : * access them for backends that aren't currently in use.
358 : */
359 148786 : pg_atomic_init_u32(&(proc->procArrayGroupNext), INVALID_PROC_NUMBER);
360 148786 : pg_atomic_init_u32(&(proc->clogGroupNext), INVALID_PROC_NUMBER);
361 148786 : pg_atomic_init_u64(&(proc->waitStart), 0);
362 : }
363 :
364 : /* Should have consumed exactly the expected amount of fast-path memory. */
365 : Assert(fpPtr == fpEndPtr);
366 :
367 : /*
368 : * Save pointers to the blocks of PGPROC structures reserved for auxiliary
369 : * processes and prepared transactions.
370 : */
371 1140 : AuxiliaryProcs = &procs[MaxBackends];
372 1140 : PreparedXactProcs = &procs[MaxBackends + NUM_AUXILIARY_PROCS];
373 1140 : }
374 :
375 : /*
376 : * InitProcess -- initialize a per-process PGPROC entry for this backend
377 : */
378 : void
379 18460 : InitProcess(void)
380 : {
381 : dlist_head *procgloballist;
382 :
383 : /*
384 : * ProcGlobal should be set up already (if we are a backend, we inherit
385 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
386 : */
387 18460 : if (ProcGlobal == NULL)
388 0 : elog(PANIC, "proc header uninitialized");
389 :
390 18460 : if (MyProc != NULL)
391 0 : elog(ERROR, "you already exist");
392 :
393 : /*
394 : * Before we start accessing the shared memory in a serious way, mark
395 : * ourselves as an active postmaster child; this is so that the postmaster
396 : * can detect it if we exit without cleaning up.
397 : */
398 18460 : if (IsUnderPostmaster)
399 18339 : RegisterPostmasterChildActive();
400 :
401 : /*
402 : * Decide which list should supply our PGPROC. This logic must match the
403 : * way the freelists were constructed in InitProcGlobal().
404 : */
405 18460 : if (AmAutoVacuumWorkerProcess() || AmSpecialWorkerProcess())
406 2144 : procgloballist = &ProcGlobal->autovacFreeProcs;
407 16316 : else if (AmBackgroundWorkerProcess())
408 2501 : procgloballist = &ProcGlobal->bgworkerFreeProcs;
409 13815 : else if (AmWalSenderProcess())
410 1207 : procgloballist = &ProcGlobal->walsenderFreeProcs;
411 : else
412 12608 : procgloballist = &ProcGlobal->freeProcs;
413 :
414 : /*
415 : * Try to get a proc struct from the appropriate free list. If this
416 : * fails, we must be out of PGPROC structures (not to mention semaphores).
417 : *
418 : * While we are holding the spinlock, also copy the current shared
419 : * estimate of spins_per_delay to local storage.
420 : */
421 18460 : SpinLockAcquire(&ProcGlobal->freeProcsLock);
422 :
423 18460 : set_spins_per_delay(ProcGlobal->spins_per_delay);
424 :
425 18460 : if (!dlist_is_empty(procgloballist))
426 : {
427 18457 : MyProc = dlist_container(PGPROC, links, dlist_pop_head_node(procgloballist));
428 18457 : SpinLockRelease(&ProcGlobal->freeProcsLock);
429 : }
430 : else
431 : {
432 : /*
433 : * If we reach here, all the PGPROCs are in use. This is one of the
434 : * possible places to detect "too many backends", so give the standard
435 : * error message. XXX do we need to give a different failure message
436 : * in the autovacuum case?
437 : */
438 3 : SpinLockRelease(&ProcGlobal->freeProcsLock);
439 3 : if (AmWalSenderProcess())
440 2 : ereport(FATAL,
441 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
442 : errmsg("number of requested standby connections exceeds \"max_wal_senders\" (currently %d)",
443 : max_wal_senders)));
444 1 : ereport(FATAL,
445 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
446 : errmsg("sorry, too many clients already")));
447 : }
448 18457 : MyProcNumber = GetNumberFromPGProc(MyProc);
449 :
450 : /*
451 : * Cross-check that the PGPROC is of the type we expect; if this were not
452 : * the case, it would get returned to the wrong list.
453 : */
454 : Assert(MyProc->procgloballist == procgloballist);
455 :
456 : /*
457 : * Initialize all fields of MyProc, except for those previously
458 : * initialized by InitProcGlobal.
459 : */
460 18457 : dlist_node_init(&MyProc->links);
461 18457 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
462 18457 : MyProc->fpVXIDLock = false;
463 18457 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
464 18457 : MyProc->xid = InvalidTransactionId;
465 18457 : MyProc->xmin = InvalidTransactionId;
466 18457 : MyProc->pid = MyProcPid;
467 18457 : MyProc->vxid.procNumber = MyProcNumber;
468 18457 : MyProc->vxid.lxid = InvalidLocalTransactionId;
469 : /* databaseId and roleId will be filled in later */
470 18457 : MyProc->databaseId = InvalidOid;
471 18457 : MyProc->roleId = InvalidOid;
472 18457 : MyProc->tempNamespaceId = InvalidOid;
473 18457 : MyProc->backendType = MyBackendType;
474 18457 : MyProc->delayChkptFlags = 0;
475 18457 : MyProc->statusFlags = 0;
476 : /* NB -- autovac launcher intentionally does not set IS_AUTOVACUUM */
477 18457 : if (AmAutoVacuumWorkerProcess())
478 1723 : MyProc->statusFlags |= PROC_IS_AUTOVACUUM;
479 18457 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
480 18457 : MyProc->lwWaitMode = 0;
481 18457 : MyProc->waitLock = NULL;
482 18457 : MyProc->waitProcLock = NULL;
483 18457 : pg_atomic_write_u64(&MyProc->waitStart, 0);
484 : #ifdef USE_ASSERT_CHECKING
485 : {
486 : int i;
487 :
488 : /* Last process should have released all locks. */
489 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
490 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
491 : }
492 : #endif
493 18457 : pg_atomic_write_u32(&MyProc->pendingRecoveryConflicts, 0);
494 :
495 : /* Initialize fields for sync rep */
496 18457 : MyProc->waitLSN = InvalidXLogRecPtr;
497 18457 : MyProc->syncRepState = SYNC_REP_NOT_WAITING;
498 18457 : dlist_node_init(&MyProc->syncRepLinks);
499 :
500 : /* Initialize fields for group XID clearing. */
501 18457 : MyProc->procArrayGroupMember = false;
502 18457 : MyProc->procArrayGroupMemberXid = InvalidTransactionId;
503 : Assert(pg_atomic_read_u32(&MyProc->procArrayGroupNext) == INVALID_PROC_NUMBER);
504 :
505 : /* Check that group locking fields are in a proper initial state. */
506 : Assert(MyProc->lockGroupLeader == NULL);
507 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
508 :
509 : /* Initialize wait event information. */
510 18457 : MyProc->wait_event_info = 0;
511 :
512 : /* Initialize fields for group transaction status update. */
513 18457 : MyProc->clogGroupMember = false;
514 18457 : MyProc->clogGroupMemberXid = InvalidTransactionId;
515 18457 : MyProc->clogGroupMemberXidStatus = TRANSACTION_STATUS_IN_PROGRESS;
516 18457 : MyProc->clogGroupMemberPage = -1;
517 18457 : MyProc->clogGroupMemberLsn = InvalidXLogRecPtr;
518 : Assert(pg_atomic_read_u32(&MyProc->clogGroupNext) == INVALID_PROC_NUMBER);
519 :
520 : /*
521 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
522 : * on it. That allows us to repoint the process latch, which so far
523 : * points to process local one, to the shared one.
524 : */
525 18457 : OwnLatch(&MyProc->procLatch);
526 18457 : SwitchToSharedLatch();
527 :
528 : /* now that we have a proc, report wait events to shared memory */
529 18457 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
530 :
531 : /*
532 : * We might be reusing a semaphore that belonged to a failed process. So
533 : * be careful and reinitialize its value here. (This is not strictly
534 : * necessary anymore, but seems like a good idea for cleanliness.)
535 : */
536 18457 : PGSemaphoreReset(MyProc->sem);
537 :
538 : /*
539 : * Arrange to clean up at backend exit.
540 : */
541 18457 : on_shmem_exit(ProcKill, 0);
542 :
543 : /*
544 : * Now that we have a PGPROC, we could try to acquire locks, so initialize
545 : * local state needed for LWLocks, and the deadlock checker.
546 : */
547 18457 : InitLWLockAccess();
548 18457 : InitDeadLockChecking();
549 :
550 : #ifdef EXEC_BACKEND
551 :
552 : /*
553 : * Initialize backend-local pointers to all the shared data structures.
554 : * (We couldn't do this until now because it needs LWLocks.)
555 : */
556 : if (IsUnderPostmaster)
557 : AttachSharedMemoryStructs();
558 : #endif
559 18457 : }
560 :
561 : /*
562 : * InitProcessPhase2 -- make MyProc visible in the shared ProcArray.
563 : *
564 : * This is separate from InitProcess because we can't acquire LWLocks until
565 : * we've created a PGPROC, but in the EXEC_BACKEND case ProcArrayAdd won't
566 : * work until after we've done AttachSharedMemoryStructs.
567 : */
568 : void
569 18448 : InitProcessPhase2(void)
570 : {
571 : Assert(MyProc != NULL);
572 :
573 : /*
574 : * Add our PGPROC to the PGPROC array in shared memory.
575 : */
576 18448 : ProcArrayAdd(MyProc);
577 :
578 : /*
579 : * Arrange to clean that up at backend exit.
580 : */
581 18448 : on_shmem_exit(RemoveProcFromArray, 0);
582 18448 : }
583 :
584 : /*
585 : * InitAuxiliaryProcess -- create a PGPROC entry for an auxiliary process
586 : *
587 : * This is called by bgwriter and similar processes so that they will have a
588 : * MyProc value that's real enough to let them wait for LWLocks. The PGPROC
589 : * and sema that are assigned are one of the extra ones created during
590 : * InitProcGlobal.
591 : *
592 : * Auxiliary processes are presently not expected to wait for real (lockmgr)
593 : * locks, so we need not set up the deadlock checker. They are never added
594 : * to the ProcArray or the sinval messaging mechanism, either. They also
595 : * don't get a VXID assigned, since this is only useful when we actually
596 : * hold lockmgr locks.
597 : *
598 : * Startup process however uses locks but never waits for them in the
599 : * normal backend sense. Startup process also takes part in sinval messaging
600 : * as a sendOnly process, so never reads messages from sinval queue. So
601 : * Startup process does have a VXID and does show up in pg_locks.
602 : */
603 : void
604 4533 : InitAuxiliaryProcess(void)
605 : {
606 : PGPROC *auxproc;
607 : int proctype;
608 :
609 : /*
610 : * ProcGlobal should be set up already (if we are a backend, we inherit
611 : * this by fork() or EXEC_BACKEND mechanism from the postmaster).
612 : */
613 4533 : if (ProcGlobal == NULL || AuxiliaryProcs == NULL)
614 0 : elog(PANIC, "proc header uninitialized");
615 :
616 4533 : if (MyProc != NULL)
617 0 : elog(ERROR, "you already exist");
618 :
619 4533 : if (IsUnderPostmaster)
620 4533 : RegisterPostmasterChildActive();
621 :
622 : /*
623 : * We use the freeProcsLock to protect assignment and releasing of
624 : * AuxiliaryProcs entries.
625 : *
626 : * While we are holding the spinlock, also copy the current shared
627 : * estimate of spins_per_delay to local storage.
628 : */
629 4533 : SpinLockAcquire(&ProcGlobal->freeProcsLock);
630 :
631 4533 : set_spins_per_delay(ProcGlobal->spins_per_delay);
632 :
633 : /*
634 : * Find a free auxproc ... *big* trouble if there isn't one ...
635 : */
636 18049 : for (proctype = 0; proctype < NUM_AUXILIARY_PROCS; proctype++)
637 : {
638 18049 : auxproc = &AuxiliaryProcs[proctype];
639 18049 : if (auxproc->pid == 0)
640 4533 : break;
641 : }
642 4533 : if (proctype >= NUM_AUXILIARY_PROCS)
643 : {
644 0 : SpinLockRelease(&ProcGlobal->freeProcsLock);
645 0 : elog(FATAL, "all AuxiliaryProcs are in use");
646 : }
647 :
648 : /* Mark auxiliary proc as in use by me */
649 : /* use volatile pointer to prevent code rearrangement */
650 4533 : ((volatile PGPROC *) auxproc)->pid = MyProcPid;
651 :
652 4533 : SpinLockRelease(&ProcGlobal->freeProcsLock);
653 :
654 4533 : MyProc = auxproc;
655 4533 : MyProcNumber = GetNumberFromPGProc(MyProc);
656 :
657 : /*
658 : * Initialize all fields of MyProc, except for those previously
659 : * initialized by InitProcGlobal.
660 : */
661 4533 : dlist_node_init(&MyProc->links);
662 4533 : MyProc->waitStatus = PROC_WAIT_STATUS_OK;
663 4533 : MyProc->fpVXIDLock = false;
664 4533 : MyProc->fpLocalTransactionId = InvalidLocalTransactionId;
665 4533 : MyProc->xid = InvalidTransactionId;
666 4533 : MyProc->xmin = InvalidTransactionId;
667 4533 : MyProc->vxid.procNumber = INVALID_PROC_NUMBER;
668 4533 : MyProc->vxid.lxid = InvalidLocalTransactionId;
669 4533 : MyProc->databaseId = InvalidOid;
670 4533 : MyProc->roleId = InvalidOid;
671 4533 : MyProc->tempNamespaceId = InvalidOid;
672 4533 : MyProc->backendType = MyBackendType;
673 4533 : MyProc->delayChkptFlags = 0;
674 4533 : MyProc->statusFlags = 0;
675 4533 : MyProc->lwWaiting = LW_WS_NOT_WAITING;
676 4533 : MyProc->lwWaitMode = 0;
677 4533 : MyProc->waitLock = NULL;
678 4533 : MyProc->waitProcLock = NULL;
679 4533 : pg_atomic_write_u64(&MyProc->waitStart, 0);
680 : #ifdef USE_ASSERT_CHECKING
681 : {
682 : int i;
683 :
684 : /* Last process should have released all locks. */
685 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
686 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
687 : }
688 : #endif
689 :
690 : /*
691 : * Acquire ownership of the PGPROC's latch, so that we can use WaitLatch
692 : * on it. That allows us to repoint the process latch, which so far
693 : * points to process local one, to the shared one.
694 : */
695 4533 : OwnLatch(&MyProc->procLatch);
696 4533 : SwitchToSharedLatch();
697 :
698 : /* now that we have a proc, report wait events to shared memory */
699 4533 : pgstat_set_wait_event_storage(&MyProc->wait_event_info);
700 :
701 : /* Check that group locking fields are in a proper initial state. */
702 : Assert(MyProc->lockGroupLeader == NULL);
703 : Assert(dlist_is_empty(&MyProc->lockGroupMembers));
704 :
705 : /*
706 : * We might be reusing a semaphore that belonged to a failed process. So
707 : * be careful and reinitialize its value here. (This is not strictly
708 : * necessary anymore, but seems like a good idea for cleanliness.)
709 : */
710 4533 : PGSemaphoreReset(MyProc->sem);
711 :
712 : /*
713 : * Arrange to clean up at process exit.
714 : */
715 4533 : on_shmem_exit(AuxiliaryProcKill, Int32GetDatum(proctype));
716 :
717 : /*
718 : * Now that we have a PGPROC, we could try to acquire lightweight locks.
719 : * Initialize local state needed for them. (Heavyweight locks cannot be
720 : * acquired in aux processes.)
721 : */
722 4533 : InitLWLockAccess();
723 :
724 : #ifdef EXEC_BACKEND
725 :
726 : /*
727 : * Initialize backend-local pointers to all the shared data structures.
728 : * (We couldn't do this until now because it needs LWLocks.)
729 : */
730 : if (IsUnderPostmaster)
731 : AttachSharedMemoryStructs();
732 : #endif
733 4533 : }
734 :
735 : /*
736 : * Used from bufmgr to share the value of the buffer that Startup waits on,
737 : * or to reset the value to "not waiting" (-1). This allows processing
738 : * of recovery conflicts for buffer pins. Set is made before backends look
739 : * at this value, so locking not required, especially since the set is
740 : * an atomic integer set operation.
741 : */
742 : void
743 18 : SetStartupBufferPinWaitBufId(int bufid)
744 : {
745 : /* use volatile pointer to prevent code rearrangement */
746 18 : volatile PROC_HDR *procglobal = ProcGlobal;
747 :
748 18 : procglobal->startupBufferPinWaitBufId = bufid;
749 18 : }
750 :
751 : /*
752 : * Used by backends when they receive a request to check for buffer pin waits.
753 : */
754 : int
755 3 : GetStartupBufferPinWaitBufId(void)
756 : {
757 : /* use volatile pointer to prevent code rearrangement */
758 3 : volatile PROC_HDR *procglobal = ProcGlobal;
759 :
760 3 : return procglobal->startupBufferPinWaitBufId;
761 : }
762 :
763 : /*
764 : * Check whether there are at least N free PGPROC objects. If false is
765 : * returned, *nfree will be set to the number of free PGPROC objects.
766 : * Otherwise, *nfree will be set to n.
767 : *
768 : * Note: this is designed on the assumption that N will generally be small.
769 : */
770 : bool
771 246 : HaveNFreeProcs(int n, int *nfree)
772 : {
773 : dlist_iter iter;
774 :
775 : Assert(n > 0);
776 : Assert(nfree);
777 :
778 246 : SpinLockAcquire(&ProcGlobal->freeProcsLock);
779 :
780 246 : *nfree = 0;
781 735 : dlist_foreach(iter, &ProcGlobal->freeProcs)
782 : {
783 732 : (*nfree)++;
784 732 : if (*nfree == n)
785 243 : break;
786 : }
787 :
788 246 : SpinLockRelease(&ProcGlobal->freeProcsLock);
789 :
790 246 : return (*nfree == n);
791 : }
792 :
793 : /*
794 : * Cancel any pending wait for lock, when aborting a transaction, and revert
795 : * any strong lock count acquisition for a lock being acquired.
796 : *
797 : * (Normally, this would only happen if we accept a cancel/die
798 : * interrupt while waiting; but an ereport(ERROR) before or during the lock
799 : * wait is within the realm of possibility, too.)
800 : */
801 : void
802 584482 : LockErrorCleanup(void)
803 : {
804 : LOCALLOCK *lockAwaited;
805 : LWLock *partitionLock;
806 : DisableTimeoutParams timeouts[2];
807 :
808 584482 : HOLD_INTERRUPTS();
809 :
810 584482 : AbortStrongLockAcquire();
811 :
812 : /* Nothing to do if we weren't waiting for a lock */
813 584482 : lockAwaited = GetAwaitedLock();
814 584482 : if (lockAwaited == NULL)
815 : {
816 584442 : RESUME_INTERRUPTS();
817 584442 : return;
818 : }
819 :
820 : /*
821 : * Turn off the deadlock and lock timeout timers, if they are still
822 : * running (see ProcSleep). Note we must preserve the LOCK_TIMEOUT
823 : * indicator flag, since this function is executed before
824 : * ProcessInterrupts when responding to SIGINT; else we'd lose the
825 : * knowledge that the SIGINT came from a lock timeout and not an external
826 : * source.
827 : */
828 40 : timeouts[0].id = DEADLOCK_TIMEOUT;
829 40 : timeouts[0].keep_indicator = false;
830 40 : timeouts[1].id = LOCK_TIMEOUT;
831 40 : timeouts[1].keep_indicator = true;
832 40 : disable_timeouts(timeouts, 2);
833 :
834 : /* Unlink myself from the wait queue, if on it (might not be anymore!) */
835 40 : partitionLock = LockHashPartitionLock(lockAwaited->hashcode);
836 40 : LWLockAcquire(partitionLock, LW_EXCLUSIVE);
837 :
838 40 : if (!dlist_node_is_detached(&MyProc->links))
839 : {
840 : /* We could not have been granted the lock yet */
841 40 : RemoveFromWaitQueue(MyProc, lockAwaited->hashcode);
842 : }
843 : else
844 : {
845 : /*
846 : * Somebody kicked us off the lock queue already. Perhaps they
847 : * granted us the lock, or perhaps they detected a deadlock. If they
848 : * did grant us the lock, we'd better remember it in our local lock
849 : * table.
850 : */
851 0 : if (MyProc->waitStatus == PROC_WAIT_STATUS_OK)
852 0 : GrantAwaitedLock();
853 : }
854 :
855 40 : ResetAwaitedLock();
856 :
857 40 : LWLockRelease(partitionLock);
858 :
859 40 : RESUME_INTERRUPTS();
860 : }
861 :
862 :
863 : /*
864 : * ProcReleaseLocks() -- release locks associated with current transaction
865 : * at main transaction commit or abort
866 : *
867 : * At main transaction commit, we release standard locks except session locks.
868 : * At main transaction abort, we release all locks including session locks.
869 : *
870 : * Advisory locks are released only if they are transaction-level;
871 : * session-level holds remain, whether this is a commit or not.
872 : *
873 : * At subtransaction commit, we don't release any locks (so this func is not
874 : * needed at all); we will defer the releasing to the parent transaction.
875 : * At subtransaction abort, we release all locks held by the subtransaction;
876 : * this is implemented by retail releasing of the locks under control of
877 : * the ResourceOwner mechanism.
878 : */
879 : void
880 552582 : ProcReleaseLocks(bool isCommit)
881 : {
882 552582 : if (!MyProc)
883 0 : return;
884 : /* If waiting, get off wait queue (should only be needed after error) */
885 552582 : LockErrorCleanup();
886 : /* Release standard locks, including session-level if aborting */
887 552582 : LockReleaseAll(DEFAULT_LOCKMETHOD, !isCommit);
888 : /* Release transaction-level advisory locks */
889 552582 : LockReleaseAll(USER_LOCKMETHOD, false);
890 : }
891 :
892 :
893 : /*
894 : * RemoveProcFromArray() -- Remove this process from the shared ProcArray.
895 : */
896 : static void
897 18448 : RemoveProcFromArray(int code, Datum arg)
898 : {
899 : Assert(MyProc != NULL);
900 18448 : ProcArrayRemove(MyProc, InvalidTransactionId);
901 18448 : }
902 :
903 : /*
904 : * ProcKill() -- Destroy the per-proc data structure for
905 : * this process. Release any of its held LW locks.
906 : */
907 : static void
908 18457 : ProcKill(int code, Datum arg)
909 : {
910 : PGPROC *proc;
911 : dlist_head *procgloballist;
912 :
913 : Assert(MyProc != NULL);
914 :
915 : /* not safe if forked by system(), etc. */
916 18457 : if (MyProc->pid != (int) getpid())
917 0 : elog(PANIC, "ProcKill() called in child process");
918 :
919 : /* Make sure we're out of the sync rep lists */
920 18457 : SyncRepCleanupAtProcExit();
921 :
922 : #ifdef USE_ASSERT_CHECKING
923 : {
924 : int i;
925 :
926 : /* Last process should have released all locks. */
927 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
928 : Assert(dlist_is_empty(&(MyProc->myProcLocks[i])));
929 : }
930 : #endif
931 :
932 : /*
933 : * Release any LW locks I am holding. There really shouldn't be any, but
934 : * it's cheap to check again before we cut the knees off the LWLock
935 : * facility by releasing our PGPROC ...
936 : */
937 18457 : LWLockReleaseAll();
938 :
939 : /*
940 : * Cleanup waiting for LSN if any.
941 : */
942 18457 : WaitLSNCleanup();
943 :
944 : /* Cancel any pending condition variable sleep, too */
945 18457 : ConditionVariableCancelSleep();
946 :
947 : /*
948 : * Detach from any lock group of which we are a member. If the leader
949 : * exits before all other group members, its PGPROC will remain allocated
950 : * until the last group process exits; that process must return the
951 : * leader's PGPROC to the appropriate list.
952 : */
953 18457 : if (MyProc->lockGroupLeader != NULL)
954 : {
955 1566 : PGPROC *leader = MyProc->lockGroupLeader;
956 1566 : LWLock *leader_lwlock = LockHashPartitionLockByProc(leader);
957 :
958 1566 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
959 : Assert(!dlist_is_empty(&leader->lockGroupMembers));
960 1566 : dlist_delete(&MyProc->lockGroupLink);
961 1566 : if (dlist_is_empty(&leader->lockGroupMembers))
962 : {
963 82 : leader->lockGroupLeader = NULL;
964 82 : if (leader != MyProc)
965 : {
966 0 : procgloballist = leader->procgloballist;
967 :
968 : /* Leader exited first; return its PGPROC. */
969 0 : SpinLockAcquire(&ProcGlobal->freeProcsLock);
970 0 : dlist_push_head(procgloballist, &leader->links);
971 0 : SpinLockRelease(&ProcGlobal->freeProcsLock);
972 : }
973 : }
974 1484 : else if (leader != MyProc)
975 1484 : MyProc->lockGroupLeader = NULL;
976 1566 : LWLockRelease(leader_lwlock);
977 : }
978 :
979 : /*
980 : * Reset MyLatch to the process local one. This is so that signal
981 : * handlers et al can continue using the latch after the shared latch
982 : * isn't ours anymore.
983 : *
984 : * Similarly, stop reporting wait events to MyProc->wait_event_info.
985 : *
986 : * After that clear MyProc and disown the shared latch.
987 : */
988 18457 : SwitchBackToLocalLatch();
989 18457 : pgstat_reset_wait_event_storage();
990 :
991 18457 : proc = MyProc;
992 18457 : MyProc = NULL;
993 18457 : MyProcNumber = INVALID_PROC_NUMBER;
994 18457 : DisownLatch(&proc->procLatch);
995 :
996 : /* Mark the proc no longer in use */
997 18457 : proc->pid = 0;
998 18457 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
999 18457 : proc->vxid.lxid = InvalidTransactionId;
1000 :
1001 18457 : procgloballist = proc->procgloballist;
1002 18457 : SpinLockAcquire(&ProcGlobal->freeProcsLock);
1003 :
1004 : /*
1005 : * If we're still a member of a locking group, that means we're a leader
1006 : * which has somehow exited before its children. The last remaining child
1007 : * will release our PGPROC. Otherwise, release it now.
1008 : */
1009 18457 : if (proc->lockGroupLeader == NULL)
1010 : {
1011 : /* Since lockGroupLeader is NULL, lockGroupMembers should be empty. */
1012 : Assert(dlist_is_empty(&proc->lockGroupMembers));
1013 :
1014 : /* Return PGPROC structure (and semaphore) to appropriate freelist */
1015 18457 : dlist_push_tail(procgloballist, &proc->links);
1016 : }
1017 :
1018 : /* Update shared estimate of spins_per_delay */
1019 18457 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1020 :
1021 18457 : SpinLockRelease(&ProcGlobal->freeProcsLock);
1022 18457 : }
1023 :
1024 : /*
1025 : * AuxiliaryProcKill() -- Cut-down version of ProcKill for auxiliary
1026 : * processes (bgwriter, etc). The PGPROC and sema are not released, only
1027 : * marked as not-in-use.
1028 : */
1029 : static void
1030 4533 : AuxiliaryProcKill(int code, Datum arg)
1031 : {
1032 4533 : int proctype = DatumGetInt32(arg);
1033 : PGPROC *auxproc PG_USED_FOR_ASSERTS_ONLY;
1034 : PGPROC *proc;
1035 :
1036 : Assert(proctype >= 0 && proctype < NUM_AUXILIARY_PROCS);
1037 :
1038 : /* not safe if forked by system(), etc. */
1039 4533 : if (MyProc->pid != (int) getpid())
1040 0 : elog(PANIC, "AuxiliaryProcKill() called in child process");
1041 :
1042 4533 : auxproc = &AuxiliaryProcs[proctype];
1043 :
1044 : Assert(MyProc == auxproc);
1045 :
1046 : /* Release any LW locks I am holding (see notes above) */
1047 4533 : LWLockReleaseAll();
1048 :
1049 : /* Cancel any pending condition variable sleep, too */
1050 4533 : ConditionVariableCancelSleep();
1051 :
1052 : /* look at the equivalent ProcKill() code for comments */
1053 4533 : SwitchBackToLocalLatch();
1054 4533 : pgstat_reset_wait_event_storage();
1055 :
1056 4533 : proc = MyProc;
1057 4533 : MyProc = NULL;
1058 4533 : MyProcNumber = INVALID_PROC_NUMBER;
1059 4533 : DisownLatch(&proc->procLatch);
1060 :
1061 4533 : SpinLockAcquire(&ProcGlobal->freeProcsLock);
1062 :
1063 : /* Mark auxiliary proc no longer in use */
1064 4533 : proc->pid = 0;
1065 4533 : proc->vxid.procNumber = INVALID_PROC_NUMBER;
1066 4533 : proc->vxid.lxid = InvalidTransactionId;
1067 :
1068 : /* Update shared estimate of spins_per_delay */
1069 4533 : ProcGlobal->spins_per_delay = update_spins_per_delay(ProcGlobal->spins_per_delay);
1070 :
1071 4533 : SpinLockRelease(&ProcGlobal->freeProcsLock);
1072 4533 : }
1073 :
1074 : /*
1075 : * AuxiliaryPidGetProc -- get PGPROC for an auxiliary process
1076 : * given its PID
1077 : *
1078 : * Returns NULL if not found.
1079 : */
1080 : PGPROC *
1081 6038 : AuxiliaryPidGetProc(int pid)
1082 : {
1083 6038 : PGPROC *result = NULL;
1084 : int index;
1085 :
1086 6038 : if (pid == 0) /* never match dummy PGPROCs */
1087 3 : return NULL;
1088 :
1089 29032 : for (index = 0; index < NUM_AUXILIARY_PROCS; index++)
1090 : {
1091 29032 : PGPROC *proc = &AuxiliaryProcs[index];
1092 :
1093 29032 : if (proc->pid == pid)
1094 : {
1095 6035 : result = proc;
1096 6035 : break;
1097 : }
1098 : }
1099 6035 : return result;
1100 : }
1101 :
1102 :
1103 : /*
1104 : * JoinWaitQueue -- join the wait queue on the specified lock
1105 : *
1106 : * It's not actually guaranteed that we need to wait when this function is
1107 : * called, because it could be that when we try to find a position at which
1108 : * to insert ourself into the wait queue, we discover that we must be inserted
1109 : * ahead of everyone who wants a lock that conflict with ours. In that case,
1110 : * we get the lock immediately. Because of this, it's sensible for this function
1111 : * to have a dontWait argument, despite the name.
1112 : *
1113 : * On entry, the caller has already set up LOCK and PROCLOCK entries to
1114 : * reflect that we have "requested" the lock. The caller is responsible for
1115 : * cleaning that up, if we end up not joining the queue after all.
1116 : *
1117 : * The lock table's partition lock must be held at entry, and is still held
1118 : * at exit. The caller must release it before calling ProcSleep().
1119 : *
1120 : * Result is one of the following:
1121 : *
1122 : * PROC_WAIT_STATUS_OK - lock was immediately granted
1123 : * PROC_WAIT_STATUS_WAITING - joined the wait queue; call ProcSleep()
1124 : * PROC_WAIT_STATUS_ERROR - immediate deadlock was detected, or would
1125 : * need to wait and dontWait == true
1126 : *
1127 : * NOTES: The process queue is now a priority queue for locking.
1128 : */
1129 : ProcWaitStatus
1130 2195 : JoinWaitQueue(LOCALLOCK *locallock, LockMethod lockMethodTable, bool dontWait)
1131 : {
1132 2195 : LOCKMODE lockmode = locallock->tag.mode;
1133 2195 : LOCK *lock = locallock->lock;
1134 2195 : PROCLOCK *proclock = locallock->proclock;
1135 2195 : uint32 hashcode = locallock->hashcode;
1136 2195 : LWLock *partitionLock PG_USED_FOR_ASSERTS_ONLY = LockHashPartitionLock(hashcode);
1137 2195 : dclist_head *waitQueue = &lock->waitProcs;
1138 2195 : PGPROC *insert_before = NULL;
1139 : LOCKMASK myProcHeldLocks;
1140 : LOCKMASK myHeldLocks;
1141 2195 : bool early_deadlock = false;
1142 2195 : PGPROC *leader = MyProc->lockGroupLeader;
1143 :
1144 : Assert(LWLockHeldByMeInMode(partitionLock, LW_EXCLUSIVE));
1145 :
1146 : /*
1147 : * Set bitmask of locks this process already holds on this object.
1148 : */
1149 2195 : myHeldLocks = MyProc->heldLocks = proclock->holdMask;
1150 :
1151 : /*
1152 : * Determine which locks we're already holding.
1153 : *
1154 : * If group locking is in use, locks held by members of my locking group
1155 : * need to be included in myHeldLocks. This is not required for relation
1156 : * extension lock which conflict among group members. However, including
1157 : * them in myHeldLocks will give group members the priority to get those
1158 : * locks as compared to other backends which are also trying to acquire
1159 : * those locks. OTOH, we can avoid giving priority to group members for
1160 : * that kind of locks, but there doesn't appear to be a clear advantage of
1161 : * the same.
1162 : */
1163 2195 : myProcHeldLocks = proclock->holdMask;
1164 2195 : myHeldLocks = myProcHeldLocks;
1165 2195 : if (leader != NULL)
1166 : {
1167 : dlist_iter iter;
1168 :
1169 36 : dlist_foreach(iter, &lock->procLocks)
1170 : {
1171 : PROCLOCK *otherproclock;
1172 :
1173 27 : otherproclock = dlist_container(PROCLOCK, lockLink, iter.cur);
1174 :
1175 27 : if (otherproclock->groupLeader == leader)
1176 13 : myHeldLocks |= otherproclock->holdMask;
1177 : }
1178 : }
1179 :
1180 : /*
1181 : * Determine where to add myself in the wait queue.
1182 : *
1183 : * Normally I should go at the end of the queue. However, if I already
1184 : * hold locks that conflict with the request of any previous waiter, put
1185 : * myself in the queue just in front of the first such waiter. This is not
1186 : * a necessary step, since deadlock detection would move me to before that
1187 : * waiter anyway; but it's relatively cheap to detect such a conflict
1188 : * immediately, and avoid delaying till deadlock timeout.
1189 : *
1190 : * Special case: if I find I should go in front of some waiter, check to
1191 : * see if I conflict with already-held locks or the requests before that
1192 : * waiter. If not, then just grant myself the requested lock immediately.
1193 : * This is the same as the test for immediate grant in LockAcquire, except
1194 : * we are only considering the part of the wait queue before my insertion
1195 : * point.
1196 : */
1197 2195 : if (myHeldLocks != 0 && !dclist_is_empty(waitQueue))
1198 : {
1199 3 : LOCKMASK aheadRequests = 0;
1200 : dlist_iter iter;
1201 :
1202 3 : dclist_foreach(iter, waitQueue)
1203 : {
1204 3 : PGPROC *proc = dlist_container(PGPROC, links, iter.cur);
1205 :
1206 : /*
1207 : * If we're part of the same locking group as this waiter, its
1208 : * locks neither conflict with ours nor contribute to
1209 : * aheadRequests.
1210 : */
1211 3 : if (leader != NULL && leader == proc->lockGroupLeader)
1212 0 : continue;
1213 :
1214 : /* Must he wait for me? */
1215 3 : if (lockMethodTable->conflictTab[proc->waitLockMode] & myHeldLocks)
1216 : {
1217 : /* Must I wait for him ? */
1218 3 : if (lockMethodTable->conflictTab[lockmode] & proc->heldLocks)
1219 : {
1220 : /*
1221 : * Yes, so we have a deadlock. Easiest way to clean up
1222 : * correctly is to call RemoveFromWaitQueue(), but we
1223 : * can't do that until we are *on* the wait queue. So, set
1224 : * a flag to check below, and break out of loop. Also,
1225 : * record deadlock info for later message.
1226 : */
1227 1 : RememberSimpleDeadLock(MyProc, lockmode, lock, proc);
1228 1 : early_deadlock = true;
1229 1 : break;
1230 : }
1231 : /* I must go before this waiter. Check special case. */
1232 2 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1233 2 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1234 : proclock))
1235 : {
1236 : /* Skip the wait and just grant myself the lock. */
1237 2 : GrantLock(lock, proclock, lockmode);
1238 2 : return PROC_WAIT_STATUS_OK;
1239 : }
1240 :
1241 : /* Put myself into wait queue before conflicting process */
1242 0 : insert_before = proc;
1243 0 : break;
1244 : }
1245 : /* Nope, so advance to next waiter */
1246 0 : aheadRequests |= LOCKBIT_ON(proc->waitLockMode);
1247 : }
1248 : }
1249 :
1250 : /*
1251 : * If we detected deadlock, give up without waiting. This must agree with
1252 : * CheckDeadLock's recovery code.
1253 : */
1254 2193 : if (early_deadlock)
1255 1 : return PROC_WAIT_STATUS_ERROR;
1256 :
1257 : /*
1258 : * At this point we know that we'd really need to sleep. If we've been
1259 : * commanded not to do that, bail out.
1260 : */
1261 2192 : if (dontWait)
1262 741 : return PROC_WAIT_STATUS_ERROR;
1263 :
1264 : /*
1265 : * Insert self into queue, at the position determined above.
1266 : */
1267 1451 : if (insert_before)
1268 0 : dclist_insert_before(waitQueue, &insert_before->links, &MyProc->links);
1269 : else
1270 1451 : dclist_push_tail(waitQueue, &MyProc->links);
1271 :
1272 1451 : lock->waitMask |= LOCKBIT_ON(lockmode);
1273 :
1274 : /* Set up wait information in PGPROC object, too */
1275 1451 : MyProc->heldLocks = myProcHeldLocks;
1276 1451 : MyProc->waitLock = lock;
1277 1451 : MyProc->waitProcLock = proclock;
1278 1451 : MyProc->waitLockMode = lockmode;
1279 :
1280 1451 : MyProc->waitStatus = PROC_WAIT_STATUS_WAITING;
1281 :
1282 1451 : return PROC_WAIT_STATUS_WAITING;
1283 : }
1284 :
1285 : /*
1286 : * ProcSleep -- put process to sleep waiting on lock
1287 : *
1288 : * This must be called when JoinWaitQueue() returns PROC_WAIT_STATUS_WAITING.
1289 : * Returns after the lock has been granted, or if a deadlock is detected. Can
1290 : * also bail out with ereport(ERROR), if some other error condition, or a
1291 : * timeout or cancellation is triggered.
1292 : *
1293 : * Result is one of the following:
1294 : *
1295 : * PROC_WAIT_STATUS_OK - lock was granted
1296 : * PROC_WAIT_STATUS_ERROR - a deadlock was detected
1297 : */
1298 : ProcWaitStatus
1299 1451 : ProcSleep(LOCALLOCK *locallock)
1300 : {
1301 1451 : LOCKMODE lockmode = locallock->tag.mode;
1302 1451 : LOCK *lock = locallock->lock;
1303 1451 : uint32 hashcode = locallock->hashcode;
1304 1451 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1305 1451 : TimestampTz standbyWaitStart = 0;
1306 1451 : bool allow_autovacuum_cancel = true;
1307 1451 : bool logged_recovery_conflict = false;
1308 : ProcWaitStatus myWaitStatus;
1309 : DeadLockState deadlock_state;
1310 :
1311 : /* The caller must've armed the on-error cleanup mechanism */
1312 : Assert(GetAwaitedLock() == locallock);
1313 : Assert(!LWLockHeldByMe(partitionLock));
1314 :
1315 : /*
1316 : * Now that we will successfully clean up after an ereport, it's safe to
1317 : * check to see if there's a buffer pin deadlock against the Startup
1318 : * process. Of course, that's only necessary if we're doing Hot Standby
1319 : * and are not the Startup process ourselves.
1320 : */
1321 1451 : if (RecoveryInProgress() && !InRecovery)
1322 1 : CheckRecoveryConflictDeadlock();
1323 :
1324 : /* Reset deadlock_state before enabling the timeout handler */
1325 1451 : deadlock_state = DS_NOT_YET_CHECKED;
1326 1451 : got_deadlock_timeout = false;
1327 :
1328 : /*
1329 : * Set timer so we can wake up after awhile and check for a deadlock. If a
1330 : * deadlock is detected, the handler sets MyProc->waitStatus =
1331 : * PROC_WAIT_STATUS_ERROR, allowing us to know that we must report failure
1332 : * rather than success.
1333 : *
1334 : * By delaying the check until we've waited for a bit, we can avoid
1335 : * running the rather expensive deadlock-check code in most cases.
1336 : *
1337 : * If LockTimeout is set, also enable the timeout for that. We can save a
1338 : * few cycles by enabling both timeout sources in one call.
1339 : *
1340 : * If InHotStandby we set lock waits slightly later for clarity with other
1341 : * code.
1342 : */
1343 1451 : if (!InHotStandby)
1344 : {
1345 1450 : if (LockTimeout > 0)
1346 : {
1347 : EnableTimeoutParams timeouts[2];
1348 :
1349 103 : timeouts[0].id = DEADLOCK_TIMEOUT;
1350 103 : timeouts[0].type = TMPARAM_AFTER;
1351 103 : timeouts[0].delay_ms = DeadlockTimeout;
1352 103 : timeouts[1].id = LOCK_TIMEOUT;
1353 103 : timeouts[1].type = TMPARAM_AFTER;
1354 103 : timeouts[1].delay_ms = LockTimeout;
1355 103 : enable_timeouts(timeouts, 2);
1356 : }
1357 : else
1358 1347 : enable_timeout_after(DEADLOCK_TIMEOUT, DeadlockTimeout);
1359 :
1360 : /*
1361 : * Use the current time obtained for the deadlock timeout timer as
1362 : * waitStart (i.e., the time when this process started waiting for the
1363 : * lock). Since getting the current time newly can cause overhead, we
1364 : * reuse the already-obtained time to avoid that overhead.
1365 : *
1366 : * Note that waitStart is updated without holding the lock table's
1367 : * partition lock, to avoid the overhead by additional lock
1368 : * acquisition. This can cause "waitstart" in pg_locks to become NULL
1369 : * for a very short period of time after the wait started even though
1370 : * "granted" is false. This is OK in practice because we can assume
1371 : * that users are likely to look at "waitstart" when waiting for the
1372 : * lock for a long time.
1373 : */
1374 1450 : pg_atomic_write_u64(&MyProc->waitStart,
1375 1450 : get_timeout_start_time(DEADLOCK_TIMEOUT));
1376 : }
1377 1 : else if (log_recovery_conflict_waits)
1378 : {
1379 : /*
1380 : * Set the wait start timestamp if logging is enabled and in hot
1381 : * standby.
1382 : */
1383 1 : standbyWaitStart = GetCurrentTimestamp();
1384 : }
1385 :
1386 : /*
1387 : * If somebody wakes us between LWLockRelease and WaitLatch, the latch
1388 : * will not wait. But a set latch does not necessarily mean that the lock
1389 : * is free now, as there are many other sources for latch sets than
1390 : * somebody releasing the lock.
1391 : *
1392 : * We process interrupts whenever the latch has been set, so cancel/die
1393 : * interrupts are processed quickly. This means we must not mind losing
1394 : * control to a cancel/die interrupt here. We don't, because we have no
1395 : * shared-state-change work to do after being granted the lock (the
1396 : * grantor did it all). We do have to worry about canceling the deadlock
1397 : * timeout and updating the locallock table, but if we lose control to an
1398 : * error, LockErrorCleanup will fix that up.
1399 : */
1400 : do
1401 : {
1402 1843 : if (InHotStandby)
1403 : {
1404 4 : bool maybe_log_conflict =
1405 4 : (standbyWaitStart != 0 && !logged_recovery_conflict);
1406 :
1407 : /* Set a timer and wait for that or for the lock to be granted */
1408 4 : ResolveRecoveryConflictWithLock(locallock->tag.lock,
1409 : maybe_log_conflict);
1410 :
1411 : /*
1412 : * Emit the log message if the startup process is waiting longer
1413 : * than deadlock_timeout for recovery conflict on lock.
1414 : */
1415 4 : if (maybe_log_conflict)
1416 : {
1417 2 : TimestampTz now = GetCurrentTimestamp();
1418 :
1419 2 : if (TimestampDifferenceExceeds(standbyWaitStart, now,
1420 : DeadlockTimeout))
1421 : {
1422 : VirtualTransactionId *vxids;
1423 : int cnt;
1424 :
1425 1 : vxids = GetLockConflicts(&locallock->tag.lock,
1426 : AccessExclusiveLock, &cnt);
1427 :
1428 : /*
1429 : * Log the recovery conflict and the list of PIDs of
1430 : * backends holding the conflicting lock. Note that we do
1431 : * logging even if there are no such backends right now
1432 : * because the startup process here has already waited
1433 : * longer than deadlock_timeout.
1434 : */
1435 1 : LogRecoveryConflict(RECOVERY_CONFLICT_LOCK,
1436 : standbyWaitStart, now,
1437 1 : cnt > 0 ? vxids : NULL, true);
1438 1 : logged_recovery_conflict = true;
1439 : }
1440 : }
1441 : }
1442 : else
1443 : {
1444 1839 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1445 1839 : PG_WAIT_LOCK | locallock->tag.lock.locktag_type);
1446 1839 : ResetLatch(MyLatch);
1447 : /* check for deadlocks first, as that's probably log-worthy */
1448 1839 : if (got_deadlock_timeout)
1449 : {
1450 25 : deadlock_state = CheckDeadLock();
1451 25 : got_deadlock_timeout = false;
1452 : }
1453 1839 : CHECK_FOR_INTERRUPTS();
1454 : }
1455 :
1456 : /*
1457 : * waitStatus could change from PROC_WAIT_STATUS_WAITING to something
1458 : * else asynchronously. Read it just once per loop to prevent
1459 : * surprising behavior (such as missing log messages).
1460 : */
1461 1803 : myWaitStatus = *((volatile ProcWaitStatus *) &MyProc->waitStatus);
1462 :
1463 : /*
1464 : * If we are not deadlocked, but are waiting on an autovacuum-induced
1465 : * task, send a signal to interrupt it.
1466 : */
1467 1803 : if (deadlock_state == DS_BLOCKED_BY_AUTOVACUUM && allow_autovacuum_cancel)
1468 : {
1469 0 : PGPROC *autovac = GetBlockingAutoVacuumPgproc();
1470 : uint8 statusFlags;
1471 : uint8 lockmethod_copy;
1472 : LOCKTAG locktag_copy;
1473 :
1474 : /*
1475 : * Grab info we need, then release lock immediately. Note this
1476 : * coding means that there is a tiny chance that the process
1477 : * terminates its current transaction and starts a different one
1478 : * before we have a change to send the signal; the worst possible
1479 : * consequence is that a for-wraparound vacuum is canceled. But
1480 : * that could happen in any case unless we were to do kill() with
1481 : * the lock held, which is much more undesirable.
1482 : */
1483 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1484 0 : statusFlags = ProcGlobal->statusFlags[autovac->pgxactoff];
1485 0 : lockmethod_copy = lock->tag.locktag_lockmethodid;
1486 0 : locktag_copy = lock->tag;
1487 0 : LWLockRelease(ProcArrayLock);
1488 :
1489 : /*
1490 : * Only do it if the worker is not working to protect against Xid
1491 : * wraparound.
1492 : */
1493 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
1494 0 : !(statusFlags & PROC_VACUUM_FOR_WRAPAROUND))
1495 : {
1496 0 : int pid = autovac->pid;
1497 :
1498 : /* report the case, if configured to do so */
1499 0 : if (message_level_is_interesting(DEBUG1))
1500 : {
1501 : StringInfoData locktagbuf;
1502 : StringInfoData logbuf; /* errdetail for server log */
1503 :
1504 0 : initStringInfo(&locktagbuf);
1505 0 : initStringInfo(&logbuf);
1506 0 : DescribeLockTag(&locktagbuf, &locktag_copy);
1507 0 : appendStringInfo(&logbuf,
1508 : "Process %d waits for %s on %s.",
1509 : MyProcPid,
1510 : GetLockmodeName(lockmethod_copy, lockmode),
1511 : locktagbuf.data);
1512 :
1513 0 : ereport(DEBUG1,
1514 : (errmsg_internal("sending cancel to blocking autovacuum PID %d",
1515 : pid),
1516 : errdetail_log("%s", logbuf.data)));
1517 :
1518 0 : pfree(locktagbuf.data);
1519 0 : pfree(logbuf.data);
1520 : }
1521 :
1522 : /* send the autovacuum worker Back to Old Kent Road */
1523 0 : if (kill(pid, SIGINT) < 0)
1524 : {
1525 : /*
1526 : * There's a race condition here: once we release the
1527 : * ProcArrayLock, it's possible for the autovac worker to
1528 : * close up shop and exit before we can do the kill().
1529 : * Therefore, we do not whinge about no-such-process.
1530 : * Other errors such as EPERM could conceivably happen if
1531 : * the kernel recycles the PID fast enough, but such cases
1532 : * seem improbable enough that it's probably best to issue
1533 : * a warning if we see some other errno.
1534 : */
1535 0 : if (errno != ESRCH)
1536 0 : ereport(WARNING,
1537 : (errmsg("could not send signal to process %d: %m",
1538 : pid)));
1539 : }
1540 : }
1541 :
1542 : /* prevent signal from being sent again more than once */
1543 0 : allow_autovacuum_cancel = false;
1544 : }
1545 :
1546 : /*
1547 : * If awoken after the deadlock check interrupt has run, and
1548 : * log_lock_waits is on, then report about the wait.
1549 : */
1550 1803 : if (log_lock_waits && deadlock_state != DS_NOT_YET_CHECKED)
1551 : {
1552 : StringInfoData buf,
1553 : lock_waiters_sbuf,
1554 : lock_holders_sbuf;
1555 : const char *modename;
1556 : long secs;
1557 : int usecs;
1558 : long msecs;
1559 221 : int lockHoldersNum = 0;
1560 :
1561 221 : initStringInfo(&buf);
1562 221 : initStringInfo(&lock_waiters_sbuf);
1563 221 : initStringInfo(&lock_holders_sbuf);
1564 :
1565 221 : DescribeLockTag(&buf, &locallock->tag.lock);
1566 221 : modename = GetLockmodeName(locallock->tag.lock.locktag_lockmethodid,
1567 : lockmode);
1568 221 : TimestampDifference(get_timeout_start_time(DEADLOCK_TIMEOUT),
1569 : GetCurrentTimestamp(),
1570 : &secs, &usecs);
1571 221 : msecs = secs * 1000 + usecs / 1000;
1572 221 : usecs = usecs % 1000;
1573 :
1574 : /* Gather a list of all lock holders and waiters */
1575 221 : LWLockAcquire(partitionLock, LW_SHARED);
1576 221 : GetLockHoldersAndWaiters(locallock, &lock_holders_sbuf,
1577 : &lock_waiters_sbuf, &lockHoldersNum);
1578 221 : LWLockRelease(partitionLock);
1579 :
1580 221 : if (deadlock_state == DS_SOFT_DEADLOCK)
1581 3 : ereport(LOG,
1582 : (errmsg("process %d avoided deadlock for %s on %s by rearranging queue order after %ld.%03d ms",
1583 : MyProcPid, modename, buf.data, msecs, usecs),
1584 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1585 : "Processes holding the lock: %s. Wait queue: %s.",
1586 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1587 218 : else if (deadlock_state == DS_HARD_DEADLOCK)
1588 : {
1589 : /*
1590 : * This message is a bit redundant with the error that will be
1591 : * reported subsequently, but in some cases the error report
1592 : * might not make it to the log (eg, if it's caught by an
1593 : * exception handler), and we want to ensure all long-wait
1594 : * events get logged.
1595 : */
1596 5 : ereport(LOG,
1597 : (errmsg("process %d detected deadlock while waiting for %s on %s 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 : }
1603 :
1604 221 : if (myWaitStatus == PROC_WAIT_STATUS_WAITING)
1605 198 : ereport(LOG,
1606 : (errmsg("process %d still 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 23 : else if (myWaitStatus == PROC_WAIT_STATUS_OK)
1612 18 : ereport(LOG,
1613 : (errmsg("process %d acquired %s on %s after %ld.%03d ms",
1614 : MyProcPid, modename, buf.data, msecs, usecs)));
1615 : else
1616 : {
1617 : Assert(myWaitStatus == PROC_WAIT_STATUS_ERROR);
1618 :
1619 : /*
1620 : * Currently, the deadlock checker always kicks its own
1621 : * process, which means that we'll only see
1622 : * PROC_WAIT_STATUS_ERROR when deadlock_state ==
1623 : * DS_HARD_DEADLOCK, and there's no need to print redundant
1624 : * messages. But for completeness and future-proofing, print
1625 : * a message if it looks like someone else kicked us off the
1626 : * lock.
1627 : */
1628 5 : if (deadlock_state != DS_HARD_DEADLOCK)
1629 0 : ereport(LOG,
1630 : (errmsg("process %d failed to acquire %s on %s after %ld.%03d ms",
1631 : MyProcPid, modename, buf.data, msecs, usecs),
1632 : (errdetail_log_plural("Process holding the lock: %s. Wait queue: %s.",
1633 : "Processes holding the lock: %s. Wait queue: %s.",
1634 : lockHoldersNum, lock_holders_sbuf.data, lock_waiters_sbuf.data))));
1635 : }
1636 :
1637 : /*
1638 : * At this point we might still need to wait for the lock. Reset
1639 : * state so we don't print the above messages again.
1640 : */
1641 221 : deadlock_state = DS_NO_DEADLOCK;
1642 :
1643 221 : pfree(buf.data);
1644 221 : pfree(lock_holders_sbuf.data);
1645 221 : pfree(lock_waiters_sbuf.data);
1646 : }
1647 1803 : } while (myWaitStatus == PROC_WAIT_STATUS_WAITING);
1648 :
1649 : /*
1650 : * Disable the timers, if they are still running. As in LockErrorCleanup,
1651 : * we must preserve the LOCK_TIMEOUT indicator flag: if a lock timeout has
1652 : * already caused QueryCancelPending to become set, we want the cancel to
1653 : * be reported as a lock timeout, not a user cancel.
1654 : */
1655 1411 : if (!InHotStandby)
1656 : {
1657 1410 : if (LockTimeout > 0)
1658 : {
1659 : DisableTimeoutParams timeouts[2];
1660 :
1661 97 : timeouts[0].id = DEADLOCK_TIMEOUT;
1662 97 : timeouts[0].keep_indicator = false;
1663 97 : timeouts[1].id = LOCK_TIMEOUT;
1664 97 : timeouts[1].keep_indicator = true;
1665 97 : disable_timeouts(timeouts, 2);
1666 : }
1667 : else
1668 1313 : disable_timeout(DEADLOCK_TIMEOUT, false);
1669 : }
1670 :
1671 : /*
1672 : * Emit the log message if recovery conflict on lock was resolved but the
1673 : * startup process waited longer than deadlock_timeout for it.
1674 : */
1675 1411 : if (InHotStandby && logged_recovery_conflict)
1676 1 : LogRecoveryConflict(RECOVERY_CONFLICT_LOCK,
1677 : standbyWaitStart, GetCurrentTimestamp(),
1678 : NULL, false);
1679 :
1680 : /*
1681 : * We don't have to do anything else, because the awaker did all the
1682 : * necessary updates of the lock table and MyProc. (The caller is
1683 : * responsible for updating the local lock table.)
1684 : */
1685 1411 : return myWaitStatus;
1686 : }
1687 :
1688 :
1689 : /*
1690 : * ProcWakeup -- wake up a process by setting its latch.
1691 : *
1692 : * Also remove the process from the wait queue and set its links invalid.
1693 : *
1694 : * The appropriate lock partition lock must be held by caller.
1695 : *
1696 : * XXX: presently, this code is only used for the "success" case, and only
1697 : * works correctly for that case. To clean up in failure case, would need
1698 : * to twiddle the lock's request counts too --- see RemoveFromWaitQueue.
1699 : * Hence, in practice the waitStatus parameter must be PROC_WAIT_STATUS_OK.
1700 : */
1701 : void
1702 1407 : ProcWakeup(PGPROC *proc, ProcWaitStatus waitStatus)
1703 : {
1704 1407 : if (dlist_node_is_detached(&proc->links))
1705 0 : return;
1706 :
1707 : Assert(proc->waitStatus == PROC_WAIT_STATUS_WAITING);
1708 :
1709 : /* Remove process from wait queue */
1710 1407 : dclist_delete_from_thoroughly(&proc->waitLock->waitProcs, &proc->links);
1711 :
1712 : /* Clean up process' state and pass it the ok/fail signal */
1713 1407 : proc->waitLock = NULL;
1714 1407 : proc->waitProcLock = NULL;
1715 1407 : proc->waitStatus = waitStatus;
1716 1407 : pg_atomic_write_u64(&MyProc->waitStart, 0);
1717 :
1718 : /* And awaken it */
1719 1407 : SetLatch(&proc->procLatch);
1720 : }
1721 :
1722 : /*
1723 : * ProcLockWakeup -- routine for waking up processes when a lock is
1724 : * released (or a prior waiter is aborted). Scan all waiters
1725 : * for lock, waken any that are no longer blocked.
1726 : *
1727 : * The appropriate lock partition lock must be held by caller.
1728 : */
1729 : void
1730 1425 : ProcLockWakeup(LockMethod lockMethodTable, LOCK *lock)
1731 : {
1732 1425 : dclist_head *waitQueue = &lock->waitProcs;
1733 1425 : LOCKMASK aheadRequests = 0;
1734 : dlist_mutable_iter miter;
1735 :
1736 1425 : if (dclist_is_empty(waitQueue))
1737 45 : return;
1738 :
1739 3260 : dclist_foreach_modify(miter, waitQueue)
1740 : {
1741 1880 : PGPROC *proc = dlist_container(PGPROC, links, miter.cur);
1742 1880 : LOCKMODE lockmode = proc->waitLockMode;
1743 :
1744 : /*
1745 : * Waken if (a) doesn't conflict with requests of earlier waiters, and
1746 : * (b) doesn't conflict with already-held locks.
1747 : */
1748 1880 : if ((lockMethodTable->conflictTab[lockmode] & aheadRequests) == 0 &&
1749 1632 : !LockCheckConflicts(lockMethodTable, lockmode, lock,
1750 : proc->waitProcLock))
1751 : {
1752 : /* OK to waken */
1753 1407 : GrantLock(lock, proc->waitProcLock, lockmode);
1754 : /* removes proc from the lock's waiting process queue */
1755 1407 : ProcWakeup(proc, PROC_WAIT_STATUS_OK);
1756 : }
1757 : else
1758 : {
1759 : /*
1760 : * Lock conflicts: Don't wake, but remember requested mode for
1761 : * later checks.
1762 : */
1763 473 : aheadRequests |= LOCKBIT_ON(lockmode);
1764 : }
1765 : }
1766 : }
1767 :
1768 : /*
1769 : * CheckDeadLock
1770 : *
1771 : * We only get to this routine, if DEADLOCK_TIMEOUT fired while waiting for a
1772 : * lock to be released by some other process. Check if there's a deadlock; if
1773 : * not, just return. If we have a real deadlock, remove ourselves from the
1774 : * lock's wait queue.
1775 : */
1776 : static DeadLockState
1777 25 : CheckDeadLock(void)
1778 : {
1779 : int i;
1780 : DeadLockState result;
1781 :
1782 : /*
1783 : * Acquire exclusive lock on the entire shared lock data structures. Must
1784 : * grab LWLocks in partition-number order to avoid LWLock deadlock.
1785 : *
1786 : * Note that the deadlock check interrupt had better not be enabled
1787 : * anywhere that this process itself holds lock partition locks, else this
1788 : * will wait forever. Also note that LWLockAcquire creates a critical
1789 : * section, so that this routine cannot be interrupted by cancel/die
1790 : * interrupts.
1791 : */
1792 425 : for (i = 0; i < NUM_LOCK_PARTITIONS; i++)
1793 400 : LWLockAcquire(LockHashPartitionLockByIndex(i), LW_EXCLUSIVE);
1794 :
1795 : /*
1796 : * Check to see if we've been awoken by anyone in the interim.
1797 : *
1798 : * If we have, we can return and resume our transaction -- happy day.
1799 : * Before we are awoken the process releasing the lock grants it to us so
1800 : * we know that we don't have to wait anymore.
1801 : *
1802 : * We check by looking to see if we've been unlinked from the wait queue.
1803 : * This is safe because we hold the lock partition lock.
1804 : */
1805 25 : if (MyProc->links.prev == NULL ||
1806 23 : MyProc->links.next == NULL)
1807 : {
1808 2 : result = DS_NO_DEADLOCK;
1809 2 : goto check_done;
1810 : }
1811 :
1812 : #ifdef LOCK_DEBUG
1813 : if (Debug_deadlocks)
1814 : DumpAllLocks();
1815 : #endif
1816 :
1817 : /* Run the deadlock check */
1818 23 : result = DeadLockCheck(MyProc);
1819 :
1820 23 : if (result == DS_HARD_DEADLOCK)
1821 : {
1822 : /*
1823 : * Oops. We have a deadlock.
1824 : *
1825 : * Get this process out of wait state. (Note: we could do this more
1826 : * efficiently by relying on lockAwaited, but use this coding to
1827 : * preserve the flexibility to kill some other transaction than the
1828 : * one detecting the deadlock.)
1829 : *
1830 : * RemoveFromWaitQueue sets MyProc->waitStatus to
1831 : * PROC_WAIT_STATUS_ERROR, so ProcSleep will report an error after we
1832 : * return.
1833 : */
1834 : Assert(MyProc->waitLock != NULL);
1835 5 : RemoveFromWaitQueue(MyProc, LockTagHashCode(&(MyProc->waitLock->tag)));
1836 :
1837 : /*
1838 : * We're done here. Transaction abort caused by the error that
1839 : * ProcSleep will raise will cause any other locks we hold to be
1840 : * released, thus allowing other processes to wake up; we don't need
1841 : * to do that here. NOTE: an exception is that releasing locks we
1842 : * hold doesn't consider the possibility of waiters that were blocked
1843 : * behind us on the lock we just failed to get, and might now be
1844 : * wakable because we're not in front of them anymore. However,
1845 : * RemoveFromWaitQueue took care of waking up any such processes.
1846 : */
1847 : }
1848 :
1849 : /*
1850 : * And release locks. We do this in reverse order for two reasons: (1)
1851 : * Anyone else who needs more than one of the locks will be trying to lock
1852 : * them in increasing order; we don't want to release the other process
1853 : * until it can get all the locks it needs. (2) This avoids O(N^2)
1854 : * behavior inside LWLockRelease.
1855 : */
1856 18 : check_done:
1857 425 : for (i = NUM_LOCK_PARTITIONS; --i >= 0;)
1858 400 : LWLockRelease(LockHashPartitionLockByIndex(i));
1859 :
1860 25 : return result;
1861 : }
1862 :
1863 : /*
1864 : * CheckDeadLockAlert - Handle the expiry of deadlock_timeout.
1865 : *
1866 : * NB: Runs inside a signal handler, be careful.
1867 : */
1868 : void
1869 25 : CheckDeadLockAlert(void)
1870 : {
1871 25 : int save_errno = errno;
1872 :
1873 25 : got_deadlock_timeout = true;
1874 :
1875 : /*
1876 : * Have to set the latch again, even if handle_sig_alarm already did. Back
1877 : * then got_deadlock_timeout wasn't yet set... It's unlikely that this
1878 : * ever would be a problem, but setting a set latch again is cheap.
1879 : *
1880 : * Note that, when this function runs inside procsignal_sigusr1_handler(),
1881 : * the handler function sets the latch again after the latch is set here.
1882 : */
1883 25 : SetLatch(MyLatch);
1884 25 : errno = save_errno;
1885 25 : }
1886 :
1887 : /*
1888 : * GetLockHoldersAndWaiters - get lock holders and waiters for a lock
1889 : *
1890 : * Fill lock_holders_sbuf and lock_waiters_sbuf with the PIDs of processes holding
1891 : * and waiting for the lock, and set lockHoldersNum to the number of lock holders.
1892 : *
1893 : * The lock table's partition lock must be held on entry and remains held on exit.
1894 : */
1895 : void
1896 221 : GetLockHoldersAndWaiters(LOCALLOCK *locallock, StringInfo lock_holders_sbuf,
1897 : StringInfo lock_waiters_sbuf, int *lockHoldersNum)
1898 : {
1899 : dlist_iter proc_iter;
1900 : PROCLOCK *curproclock;
1901 221 : LOCK *lock = locallock->lock;
1902 221 : bool first_holder = true,
1903 221 : first_waiter = true;
1904 :
1905 : #ifdef USE_ASSERT_CHECKING
1906 : {
1907 : uint32 hashcode = locallock->hashcode;
1908 : LWLock *partitionLock = LockHashPartitionLock(hashcode);
1909 :
1910 : Assert(LWLockHeldByMe(partitionLock));
1911 : }
1912 : #endif
1913 :
1914 221 : *lockHoldersNum = 0;
1915 :
1916 : /*
1917 : * Loop over the lock's procLocks to gather a list of all holders and
1918 : * waiters. Thus we will be able to provide more detailed information for
1919 : * lock debugging purposes.
1920 : *
1921 : * lock->procLocks contains all processes which hold or wait for this
1922 : * lock.
1923 : */
1924 663 : dlist_foreach(proc_iter, &lock->procLocks)
1925 : {
1926 442 : curproclock =
1927 442 : dlist_container(PROCLOCK, lockLink, proc_iter.cur);
1928 :
1929 : /*
1930 : * We are a waiter if myProc->waitProcLock == curproclock; we are a
1931 : * holder if it is NULL or something different.
1932 : */
1933 442 : if (curproclock->tag.myProc->waitProcLock == curproclock)
1934 : {
1935 214 : if (first_waiter)
1936 : {
1937 199 : appendStringInfo(lock_waiters_sbuf, "%d",
1938 199 : curproclock->tag.myProc->pid);
1939 199 : first_waiter = false;
1940 : }
1941 : else
1942 15 : appendStringInfo(lock_waiters_sbuf, ", %d",
1943 15 : curproclock->tag.myProc->pid);
1944 : }
1945 : else
1946 : {
1947 228 : if (first_holder)
1948 : {
1949 221 : appendStringInfo(lock_holders_sbuf, "%d",
1950 221 : curproclock->tag.myProc->pid);
1951 221 : first_holder = false;
1952 : }
1953 : else
1954 7 : appendStringInfo(lock_holders_sbuf, ", %d",
1955 7 : curproclock->tag.myProc->pid);
1956 :
1957 228 : (*lockHoldersNum)++;
1958 : }
1959 : }
1960 221 : }
1961 :
1962 : /*
1963 : * ProcWaitForSignal - wait for a signal from another backend.
1964 : *
1965 : * As this uses the generic process latch the caller has to be robust against
1966 : * unrelated wakeups: Always check that the desired state has occurred, and
1967 : * wait again if not.
1968 : */
1969 : void
1970 22 : ProcWaitForSignal(uint32 wait_event_info)
1971 : {
1972 22 : (void) WaitLatch(MyLatch, WL_LATCH_SET | WL_EXIT_ON_PM_DEATH, 0,
1973 : wait_event_info);
1974 22 : ResetLatch(MyLatch);
1975 22 : CHECK_FOR_INTERRUPTS();
1976 22 : }
1977 :
1978 : /*
1979 : * ProcSendSignal - set the latch of a backend identified by ProcNumber
1980 : */
1981 : void
1982 10 : ProcSendSignal(ProcNumber procNumber)
1983 : {
1984 10 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
1985 0 : elog(ERROR, "procNumber out of range");
1986 :
1987 10 : SetLatch(&GetPGProcByNumber(procNumber)->procLatch);
1988 10 : }
1989 :
1990 : /*
1991 : * BecomeLockGroupLeader - designate process as lock group leader
1992 : *
1993 : * Once this function has returned, other processes can join the lock group
1994 : * by calling BecomeLockGroupMember.
1995 : */
1996 : void
1997 632 : BecomeLockGroupLeader(void)
1998 : {
1999 : LWLock *leader_lwlock;
2000 :
2001 : /* If we already did it, we don't need to do it again. */
2002 632 : if (MyProc->lockGroupLeader == MyProc)
2003 550 : return;
2004 :
2005 : /* We had better not be a follower. */
2006 : Assert(MyProc->lockGroupLeader == NULL);
2007 :
2008 : /* Create single-member group, containing only ourselves. */
2009 82 : leader_lwlock = LockHashPartitionLockByProc(MyProc);
2010 82 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2011 82 : MyProc->lockGroupLeader = MyProc;
2012 82 : dlist_push_head(&MyProc->lockGroupMembers, &MyProc->lockGroupLink);
2013 82 : LWLockRelease(leader_lwlock);
2014 : }
2015 :
2016 : /*
2017 : * BecomeLockGroupMember - designate process as lock group member
2018 : *
2019 : * This is pretty straightforward except for the possibility that the leader
2020 : * whose group we're trying to join might exit before we manage to do so;
2021 : * and the PGPROC might get recycled for an unrelated process. To avoid
2022 : * that, we require the caller to pass the PID of the intended PGPROC as
2023 : * an interlock. Returns true if we successfully join the intended lock
2024 : * group, and false if not.
2025 : */
2026 : bool
2027 1484 : BecomeLockGroupMember(PGPROC *leader, int pid)
2028 : {
2029 : LWLock *leader_lwlock;
2030 1484 : bool ok = false;
2031 :
2032 : /* Group leader can't become member of group */
2033 : Assert(MyProc != leader);
2034 :
2035 : /* Can't already be a member of a group */
2036 : Assert(MyProc->lockGroupLeader == NULL);
2037 :
2038 : /* PID must be valid. */
2039 : Assert(pid != 0);
2040 :
2041 : /*
2042 : * Get lock protecting the group fields. Note LockHashPartitionLockByProc
2043 : * calculates the proc number based on the PGPROC slot without looking at
2044 : * its contents, so we will acquire the correct lock even if the leader
2045 : * PGPROC is in process of being recycled.
2046 : */
2047 1484 : leader_lwlock = LockHashPartitionLockByProc(leader);
2048 1484 : LWLockAcquire(leader_lwlock, LW_EXCLUSIVE);
2049 :
2050 : /* Is this the leader we're looking for? */
2051 1484 : if (leader->pid == pid && leader->lockGroupLeader == leader)
2052 : {
2053 : /* OK, join the group */
2054 1484 : ok = true;
2055 1484 : MyProc->lockGroupLeader = leader;
2056 1484 : dlist_push_tail(&leader->lockGroupMembers, &MyProc->lockGroupLink);
2057 : }
2058 1484 : LWLockRelease(leader_lwlock);
2059 :
2060 1484 : return ok;
2061 : }
|
