Thanks to visit codestin.com
Credit goes to doxygen.postgresql.org

PostgreSQL Source Code git master
procarray.c
Go to the documentation of this file.
1/*-------------------------------------------------------------------------
2 *
3 * procarray.c
4 * POSTGRES process array code.
5 *
6 *
7 * This module maintains arrays of PGPROC substructures, as well as associated
8 * arrays in ProcGlobal, for all active backends. Although there are several
9 * uses for this, the principal one is as a means of determining the set of
10 * currently running transactions.
11 *
12 * Because of various subtle race conditions it is critical that a backend
13 * hold the correct locks while setting or clearing its xid (in
14 * ProcGlobal->xids[]/MyProc->xid). See notes in
15 * src/backend/access/transam/README.
16 *
17 * The process arrays now also include structures representing prepared
18 * transactions. The xid and subxids fields of these are valid, as are the
19 * myProcLocks lists. They can be distinguished from regular backend PGPROCs
20 * at need by checking for pid == 0.
21 *
22 * During hot standby, we also keep a list of XIDs representing transactions
23 * that are known to be running on the primary (or more precisely, were running
24 * as of the current point in the WAL stream). This list is kept in the
25 * KnownAssignedXids array, and is updated by watching the sequence of
26 * arriving XIDs. This is necessary because if we leave those XIDs out of
27 * snapshots taken for standby queries, then they will appear to be already
28 * complete, leading to MVCC failures. Note that in hot standby, the PGPROC
29 * array represents standby processes, which by definition are not running
30 * transactions that have XIDs.
31 *
32 * It is perhaps possible for a backend on the primary to terminate without
33 * writing an abort record for its transaction. While that shouldn't really
34 * happen, it would tie up KnownAssignedXids indefinitely, so we protect
35 * ourselves by pruning the array when a valid list of running XIDs arrives.
36 *
37 * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
38 * Portions Copyright (c) 1994, Regents of the University of California
39 *
40 *
41 * IDENTIFICATION
42 * src/backend/storage/ipc/procarray.c
43 *
44 *-------------------------------------------------------------------------
45 */
46#include "postgres.h"
47
48#include <signal.h>
49
50#include "access/subtrans.h"
51#include "access/transam.h"
52#include "access/twophase.h"
53#include "access/xact.h"
54#include "access/xlogutils.h"
55#include "catalog/catalog.h"
56#include "catalog/pg_authid.h"
57#include "miscadmin.h"
58#include "pgstat.h"
59#include "port/pg_lfind.h"
60#include "storage/proc.h"
61#include "storage/procarray.h"
62#include "utils/acl.h"
63#include "utils/builtins.h"
64#include "utils/lsyscache.h"
65#include "utils/rel.h"
66#include "utils/snapmgr.h"
67
68#define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
69
70/* Our shared memory area */
71typedef struct ProcArrayStruct
72{
73 int numProcs; /* number of valid procs entries */
74 int maxProcs; /* allocated size of procs array */
75
76 /*
77 * Known assigned XIDs handling
78 */
79 int maxKnownAssignedXids; /* allocated size of array */
80 int numKnownAssignedXids; /* current # of valid entries */
81 int tailKnownAssignedXids; /* index of oldest valid element */
82 int headKnownAssignedXids; /* index of newest element, + 1 */
83
84 /*
85 * Highest subxid that has been removed from KnownAssignedXids array to
86 * prevent overflow; or InvalidTransactionId if none. We track this for
87 * similar reasons to tracking overflowing cached subxids in PGPROC
88 * entries. Must hold exclusive ProcArrayLock to change this, and shared
89 * lock to read it.
90 */
92
93 /* oldest xmin of any replication slot */
95 /* oldest catalog xmin of any replication slot */
97
98 /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
101
102/*
103 * State for the GlobalVisTest* family of functions. Those functions can
104 * e.g. be used to decide if a deleted row can be removed without violating
105 * MVCC semantics: If the deleted row's xmax is not considered to be running
106 * by anyone, the row can be removed.
107 *
108 * To avoid slowing down GetSnapshotData(), we don't calculate a precise
109 * cutoff XID while building a snapshot (looking at the frequently changing
110 * xmins scales badly). Instead we compute two boundaries while building the
111 * snapshot:
112 *
113 * 1) definitely_needed, indicating that rows deleted by XIDs >=
114 * definitely_needed are definitely still visible.
115 *
116 * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
117 * definitely be removed
118 *
119 * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
120 * && XID < definitely_needed), the boundaries can be recomputed (using
121 * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
122 * maintaining an accurate value all the time.
123 *
124 * As it is not cheap to compute accurate boundaries, we limit the number of
125 * times that happens in short succession. See GlobalVisTestShouldUpdate().
126 *
127 *
128 * There are three backend lifetime instances of this struct, optimized for
129 * different types of relations. As e.g. a normal user defined table in one
130 * database is inaccessible to backends connected to another database, a test
131 * specific to a relation can be more aggressive than a test for a shared
132 * relation. Currently we track four different states:
133 *
134 * 1) GlobalVisSharedRels, which only considers an XID's
135 * effects visible-to-everyone if neither snapshots in any database, nor a
136 * replication slot's xmin, nor a replication slot's catalog_xmin might
137 * still consider XID as running.
138 *
139 * 2) GlobalVisCatalogRels, which only considers an XID's
140 * effects visible-to-everyone if neither snapshots in the current
141 * database, nor a replication slot's xmin, nor a replication slot's
142 * catalog_xmin might still consider XID as running.
143 *
144 * I.e. the difference to GlobalVisSharedRels is that
145 * snapshot in other databases are ignored.
146 *
147 * 3) GlobalVisDataRels, which only considers an XID's
148 * effects visible-to-everyone if neither snapshots in the current
149 * database, nor a replication slot's xmin consider XID as running.
150 *
151 * I.e. the difference to GlobalVisCatalogRels is that
152 * replication slot's catalog_xmin is not taken into account.
153 *
154 * 4) GlobalVisTempRels, which only considers the current session, as temp
155 * tables are not visible to other sessions.
156 *
157 * GlobalVisTestFor(relation) returns the appropriate state
158 * for the relation.
159 *
160 * The boundaries are FullTransactionIds instead of TransactionIds to avoid
161 * wraparound dangers. There e.g. would otherwise exist no procarray state to
162 * prevent maybe_needed to become old enough after the GetSnapshotData()
163 * call.
164 *
165 * The typedef is in the header.
166 */
168{
169 /* XIDs >= are considered running by some backend */
171
172 /* XIDs < are not considered to be running by any backend */
174};
175
176/*
177 * Result of ComputeXidHorizons().
178 */
180{
181 /*
182 * The value of TransamVariables->latestCompletedXid when
183 * ComputeXidHorizons() held ProcArrayLock.
184 */
186
187 /*
188 * The same for procArray->replication_slot_xmin and
189 * procArray->replication_slot_catalog_xmin.
190 */
193
194 /*
195 * Oldest xid that any backend might still consider running. This needs to
196 * include processes running VACUUM, in contrast to the normal visibility
197 * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
198 * determining visibility, but doesn't care about rows above its xmin to
199 * be removed.
200 *
201 * This likely should only be needed to determine whether pg_subtrans can
202 * be truncated. It currently includes the effects of replication slots,
203 * for historical reasons. But that could likely be changed.
204 */
206
207 /*
208 * Oldest xid for which deleted tuples need to be retained in shared
209 * tables.
210 *
211 * This includes the effects of replication slots. If that's not desired,
212 * look at shared_oldest_nonremovable_raw;
213 */
215
216 /*
217 * Oldest xid that may be necessary to retain in shared tables. This is
218 * the same as shared_oldest_nonremovable, except that is not affected by
219 * replication slot's catalog_xmin.
220 *
221 * This is mainly useful to be able to send the catalog_xmin to upstream
222 * streaming replication servers via hot_standby_feedback, so they can
223 * apply the limit only when accessing catalog tables.
224 */
226
227 /*
228 * Oldest xid for which deleted tuples need to be retained in non-shared
229 * catalog tables.
230 */
232
233 /*
234 * Oldest xid for which deleted tuples need to be retained in normal user
235 * defined tables.
236 */
238
239 /*
240 * Oldest xid for which deleted tuples need to be retained in this
241 * session's temporary tables.
242 */
245
246/*
247 * Return value for GlobalVisHorizonKindForRel().
248 */
250{
256
257/*
258 * Reason codes for KnownAssignedXidsCompress().
259 */
261{
262 KAX_NO_SPACE, /* need to free up space at array end */
263 KAX_PRUNE, /* we just pruned old entries */
264 KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
265 KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
267
268
270
272
273/*
274 * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
275 */
277
278/*
279 * Bookkeeping for tracking emulated transactions in recovery
280 */
284
285/*
286 * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
287 * the highest xid that might still be running that we don't have in
288 * KnownAssignedXids.
289 */
291
292/*
293 * State for visibility checks on different types of relations. See struct
294 * GlobalVisState for details. As shared, catalog, normal and temporary
295 * relations can have different horizons, one such state exists for each.
296 */
301
302/*
303 * This backend's RecentXmin at the last time the accurate xmin horizon was
304 * recomputed, or InvalidTransactionId if it has not. Used to limit how many
305 * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
306 */
308
309#ifdef XIDCACHE_DEBUG
310
311/* counters for XidCache measurement */
312static long xc_by_recent_xmin = 0;
313static long xc_by_known_xact = 0;
314static long xc_by_my_xact = 0;
315static long xc_by_latest_xid = 0;
316static long xc_by_main_xid = 0;
317static long xc_by_child_xid = 0;
318static long xc_by_known_assigned = 0;
319static long xc_no_overflow = 0;
320static long xc_slow_answer = 0;
321
322#define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
323#define xc_by_known_xact_inc() (xc_by_known_xact++)
324#define xc_by_my_xact_inc() (xc_by_my_xact++)
325#define xc_by_latest_xid_inc() (xc_by_latest_xid++)
326#define xc_by_main_xid_inc() (xc_by_main_xid++)
327#define xc_by_child_xid_inc() (xc_by_child_xid++)
328#define xc_by_known_assigned_inc() (xc_by_known_assigned++)
329#define xc_no_overflow_inc() (xc_no_overflow++)
330#define xc_slow_answer_inc() (xc_slow_answer++)
331
332static void DisplayXidCache(void);
333#else /* !XIDCACHE_DEBUG */
334
335#define xc_by_recent_xmin_inc() ((void) 0)
336#define xc_by_known_xact_inc() ((void) 0)
337#define xc_by_my_xact_inc() ((void) 0)
338#define xc_by_latest_xid_inc() ((void) 0)
339#define xc_by_main_xid_inc() ((void) 0)
340#define xc_by_child_xid_inc() ((void) 0)
341#define xc_by_known_assigned_inc() ((void) 0)
342#define xc_no_overflow_inc() ((void) 0)
343#define xc_slow_answer_inc() ((void) 0)
344#endif /* XIDCACHE_DEBUG */
345
346/* Primitives for KnownAssignedXids array handling for standby */
347static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
348static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
349 bool exclusive_lock);
350static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
353static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
354 TransactionId *subxids);
356static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
358 TransactionId *xmin,
359 TransactionId xmax);
361static void KnownAssignedXidsDisplay(int trace_level);
362static void KnownAssignedXidsReset(void);
363static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
364static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
365static void MaintainLatestCompletedXid(TransactionId latestXid);
367
369 TransactionId xid);
371
372/*
373 * Report shared-memory space needed by ProcArrayShmemInit
374 */
375Size
377{
378 Size size;
379
380 /* Size of the ProcArray structure itself */
381#define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
382
383 size = offsetof(ProcArrayStruct, pgprocnos);
384 size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
385
386 /*
387 * During Hot Standby processing we have a data structure called
388 * KnownAssignedXids, created in shared memory. Local data structures are
389 * also created in various backends during GetSnapshotData(),
390 * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
391 * main structures created in those functions must be identically sized,
392 * since we may at times copy the whole of the data structures around. We
393 * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
394 *
395 * Ideally we'd only create this structure if we were actually doing hot
396 * standby in the current run, but we don't know that yet at the time
397 * shared memory is being set up.
398 */
399#define TOTAL_MAX_CACHED_SUBXIDS \
400 ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
401
403 {
404 size = add_size(size,
405 mul_size(sizeof(TransactionId),
407 size = add_size(size,
408 mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
409 }
410
411 return size;
412}
413
414/*
415 * Initialize the shared PGPROC array during postmaster startup.
416 */
417void
419{
420 bool found;
421
422 /* Create or attach to the ProcArray shared structure */
424 ShmemInitStruct("Proc Array",
425 add_size(offsetof(ProcArrayStruct, pgprocnos),
426 mul_size(sizeof(int),
428 &found);
429
430 if (!found)
431 {
432 /*
433 * We're the first - initialize.
434 */
435 procArray->numProcs = 0;
445 }
446
448
449 /* Create or attach to the KnownAssignedXids arrays too, if needed */
451 {
453 ShmemInitStruct("KnownAssignedXids",
454 mul_size(sizeof(TransactionId),
456 &found);
457 KnownAssignedXidsValid = (bool *)
458 ShmemInitStruct("KnownAssignedXidsValid",
459 mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
460 &found);
461 }
462}
463
464/*
465 * Add the specified PGPROC to the shared array.
466 */
467void
469{
470 int pgprocno = GetNumberFromPGProc(proc);
471 ProcArrayStruct *arrayP = procArray;
472 int index;
473 int movecount;
474
475 /* See ProcGlobal comment explaining why both locks are held */
476 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
477 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
478
479 if (arrayP->numProcs >= arrayP->maxProcs)
480 {
481 /*
482 * Oops, no room. (This really shouldn't happen, since there is a
483 * fixed supply of PGPROC structs too, and so we should have failed
484 * earlier.)
485 */
487 (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
488 errmsg("sorry, too many clients already")));
489 }
490
491 /*
492 * Keep the procs array sorted by (PGPROC *) so that we can utilize
493 * locality of references much better. This is useful while traversing the
494 * ProcArray because there is an increased likelihood of finding the next
495 * PGPROC structure in the cache.
496 *
497 * Since the occurrence of adding/removing a proc is much lower than the
498 * access to the ProcArray itself, the overhead should be marginal
499 */
500 for (index = 0; index < arrayP->numProcs; index++)
501 {
502 int this_procno = arrayP->pgprocnos[index];
503
504 Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
505 Assert(allProcs[this_procno].pgxactoff == index);
506
507 /* If we have found our right position in the array, break */
508 if (this_procno > pgprocno)
509 break;
510 }
511
512 movecount = arrayP->numProcs - index;
513 memmove(&arrayP->pgprocnos[index + 1],
514 &arrayP->pgprocnos[index],
515 movecount * sizeof(*arrayP->pgprocnos));
516 memmove(&ProcGlobal->xids[index + 1],
518 movecount * sizeof(*ProcGlobal->xids));
519 memmove(&ProcGlobal->subxidStates[index + 1],
521 movecount * sizeof(*ProcGlobal->subxidStates));
522 memmove(&ProcGlobal->statusFlags[index + 1],
524 movecount * sizeof(*ProcGlobal->statusFlags));
525
526 arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
527 proc->pgxactoff = index;
528 ProcGlobal->xids[index] = proc->xid;
531
532 arrayP->numProcs++;
533
534 /* adjust pgxactoff for all following PGPROCs */
535 index++;
536 for (; index < arrayP->numProcs; index++)
537 {
538 int procno = arrayP->pgprocnos[index];
539
540 Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
541 Assert(allProcs[procno].pgxactoff == index - 1);
542
543 allProcs[procno].pgxactoff = index;
544 }
545
546 /*
547 * Release in reversed acquisition order, to reduce frequency of having to
548 * wait for XidGenLock while holding ProcArrayLock.
549 */
550 LWLockRelease(XidGenLock);
551 LWLockRelease(ProcArrayLock);
552}
553
554/*
555 * Remove the specified PGPROC from the shared array.
556 *
557 * When latestXid is a valid XID, we are removing a live 2PC gxact from the
558 * array, and thus causing it to appear as "not running" anymore. In this
559 * case we must advance latestCompletedXid. (This is essentially the same
560 * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
561 * the ProcArrayLock only once, and don't damage the content of the PGPROC;
562 * twophase.c depends on the latter.)
563 */
564void
566{
567 ProcArrayStruct *arrayP = procArray;
568 int myoff;
569 int movecount;
570
571#ifdef XIDCACHE_DEBUG
572 /* dump stats at backend shutdown, but not prepared-xact end */
573 if (proc->pid != 0)
574 DisplayXidCache();
575#endif
576
577 /* See ProcGlobal comment explaining why both locks are held */
578 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
579 LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
580
581 myoff = proc->pgxactoff;
582
583 Assert(myoff >= 0 && myoff < arrayP->numProcs);
584 Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
585
586 if (TransactionIdIsValid(latestXid))
587 {
589
590 /* Advance global latestCompletedXid while holding the lock */
592
593 /* Same with xactCompletionCount */
595
597 ProcGlobal->subxidStates[myoff].overflowed = false;
598 ProcGlobal->subxidStates[myoff].count = 0;
599 }
600 else
601 {
602 /* Shouldn't be trying to remove a live transaction here */
604 }
605
607 Assert(ProcGlobal->subxidStates[myoff].count == 0);
608 Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
609
610 ProcGlobal->statusFlags[myoff] = 0;
611
612 /* Keep the PGPROC array sorted. See notes above */
613 movecount = arrayP->numProcs - myoff - 1;
614 memmove(&arrayP->pgprocnos[myoff],
615 &arrayP->pgprocnos[myoff + 1],
616 movecount * sizeof(*arrayP->pgprocnos));
617 memmove(&ProcGlobal->xids[myoff],
618 &ProcGlobal->xids[myoff + 1],
619 movecount * sizeof(*ProcGlobal->xids));
620 memmove(&ProcGlobal->subxidStates[myoff],
621 &ProcGlobal->subxidStates[myoff + 1],
622 movecount * sizeof(*ProcGlobal->subxidStates));
623 memmove(&ProcGlobal->statusFlags[myoff],
624 &ProcGlobal->statusFlags[myoff + 1],
625 movecount * sizeof(*ProcGlobal->statusFlags));
626
627 arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
628 arrayP->numProcs--;
629
630 /*
631 * Adjust pgxactoff of following procs for removed PGPROC (note that
632 * numProcs already has been decremented).
633 */
634 for (int index = myoff; index < arrayP->numProcs; index++)
635 {
636 int procno = arrayP->pgprocnos[index];
637
638 Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
639 Assert(allProcs[procno].pgxactoff - 1 == index);
640
641 allProcs[procno].pgxactoff = index;
642 }
643
644 /*
645 * Release in reversed acquisition order, to reduce frequency of having to
646 * wait for XidGenLock while holding ProcArrayLock.
647 */
648 LWLockRelease(XidGenLock);
649 LWLockRelease(ProcArrayLock);
650}
651
652
653/*
654 * ProcArrayEndTransaction -- mark a transaction as no longer running
655 *
656 * This is used interchangeably for commit and abort cases. The transaction
657 * commit/abort must already be reported to WAL and pg_xact.
658 *
659 * proc is currently always MyProc, but we pass it explicitly for flexibility.
660 * latestXid is the latest Xid among the transaction's main XID and
661 * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
662 * the caller to pass latestXid, instead of computing it from the PGPROC's
663 * contents, because the subxid information in the PGPROC might be
664 * incomplete.)
665 */
666void
668{
669 if (TransactionIdIsValid(latestXid))
670 {
671 /*
672 * We must lock ProcArrayLock while clearing our advertised XID, so
673 * that we do not exit the set of "running" transactions while someone
674 * else is taking a snapshot. See discussion in
675 * src/backend/access/transam/README.
676 */
678
679 /*
680 * If we can immediately acquire ProcArrayLock, we clear our own XID
681 * and release the lock. If not, use group XID clearing to improve
682 * efficiency.
683 */
684 if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
685 {
686 ProcArrayEndTransactionInternal(proc, latestXid);
687 LWLockRelease(ProcArrayLock);
688 }
689 else
690 ProcArrayGroupClearXid(proc, latestXid);
691 }
692 else
693 {
694 /*
695 * If we have no XID, we don't need to lock, since we won't affect
696 * anyone else's calculation of a snapshot. We might change their
697 * estimate of global xmin, but that's OK.
698 */
700 Assert(proc->subxidStatus.count == 0);
702
705
706 /* be sure this is cleared in abort */
707 proc->delayChkptFlags = 0;
708
709 proc->recoveryConflictPending = false;
710
711 /* must be cleared with xid/xmin: */
712 /* avoid unnecessarily dirtying shared cachelines */
714 {
715 Assert(!LWLockHeldByMe(ProcArrayLock));
716 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
718 proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
720 LWLockRelease(ProcArrayLock);
721 }
722 }
723}
724
725/*
726 * Mark a write transaction as no longer running.
727 *
728 * We don't do any locking here; caller must handle that.
729 */
730static inline void
732{
733 int pgxactoff = proc->pgxactoff;
734
735 /*
736 * Note: we need exclusive lock here because we're going to change other
737 * processes' PGPROC entries.
738 */
741 Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
742
747
748 /* be sure this is cleared in abort */
749 proc->delayChkptFlags = 0;
750
751 proc->recoveryConflictPending = false;
752
753 /* must be cleared with xid/xmin: */
754 /* avoid unnecessarily dirtying shared cachelines */
756 {
757 proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
759 }
760
761 /* Clear the subtransaction-XID cache too while holding the lock */
762 Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
764 if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
765 {
766 ProcGlobal->subxidStates[pgxactoff].count = 0;
767 ProcGlobal->subxidStates[pgxactoff].overflowed = false;
768 proc->subxidStatus.count = 0;
769 proc->subxidStatus.overflowed = false;
770 }
771
772 /* Also advance global latestCompletedXid while holding the lock */
774
775 /* Same with xactCompletionCount */
777}
778
779/*
780 * ProcArrayGroupClearXid -- group XID clearing
781 *
782 * When we cannot immediately acquire ProcArrayLock in exclusive mode at
783 * commit time, add ourselves to a list of processes that need their XIDs
784 * cleared. The first process to add itself to the list will acquire
785 * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
786 * on behalf of all group members. This avoids a great deal of contention
787 * around ProcArrayLock when many processes are trying to commit at once,
788 * since the lock need not be repeatedly handed off from one committing
789 * process to the next.
790 */
791static void
793{
794 int pgprocno = GetNumberFromPGProc(proc);
795 PROC_HDR *procglobal = ProcGlobal;
796 uint32 nextidx;
797 uint32 wakeidx;
798
799 /* We should definitely have an XID to clear. */
801
802 /* Add ourselves to the list of processes needing a group XID clear. */
803 proc->procArrayGroupMember = true;
804 proc->procArrayGroupMemberXid = latestXid;
805 nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
806 while (true)
807 {
809
811 &nextidx,
812 (uint32) pgprocno))
813 break;
814 }
815
816 /*
817 * If the list was not empty, the leader will clear our XID. It is
818 * impossible to have followers without a leader because the first process
819 * that has added itself to the list will always have nextidx as
820 * INVALID_PROC_NUMBER.
821 */
822 if (nextidx != INVALID_PROC_NUMBER)
823 {
824 int extraWaits = 0;
825
826 /* Sleep until the leader clears our XID. */
827 pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
828 for (;;)
829 {
830 /* acts as a read barrier */
831 PGSemaphoreLock(proc->sem);
832 if (!proc->procArrayGroupMember)
833 break;
834 extraWaits++;
835 }
837
839
840 /* Fix semaphore count for any absorbed wakeups */
841 while (extraWaits-- > 0)
842 PGSemaphoreUnlock(proc->sem);
843 return;
844 }
845
846 /* We are the leader. Acquire the lock on behalf of everyone. */
847 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
848
849 /*
850 * Now that we've got the lock, clear the list of processes waiting for
851 * group XID clearing, saving a pointer to the head of the list. Trying
852 * to pop elements one at a time could lead to an ABA problem.
853 */
854 nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
856
857 /* Remember head of list so we can perform wakeups after dropping lock. */
858 wakeidx = nextidx;
859
860 /* Walk the list and clear all XIDs. */
861 while (nextidx != INVALID_PROC_NUMBER)
862 {
863 PGPROC *nextproc = &allProcs[nextidx];
864
866
867 /* Move to next proc in list. */
868 nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
869 }
870
871 /* We're done with the lock now. */
872 LWLockRelease(ProcArrayLock);
873
874 /*
875 * Now that we've released the lock, go back and wake everybody up. We
876 * don't do this under the lock so as to keep lock hold times to a
877 * minimum. The system calls we need to perform to wake other processes
878 * up are probably much slower than the simple memory writes we did while
879 * holding the lock.
880 */
881 while (wakeidx != INVALID_PROC_NUMBER)
882 {
883 PGPROC *nextproc = &allProcs[wakeidx];
884
885 wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
887
888 /* ensure all previous writes are visible before follower continues. */
890
891 nextproc->procArrayGroupMember = false;
892
893 if (nextproc != MyProc)
894 PGSemaphoreUnlock(nextproc->sem);
895 }
896}
897
898/*
899 * ProcArrayClearTransaction -- clear the transaction fields
900 *
901 * This is used after successfully preparing a 2-phase transaction. We are
902 * not actually reporting the transaction's XID as no longer running --- it
903 * will still appear as running because the 2PC's gxact is in the ProcArray
904 * too. We just have to clear out our own PGPROC.
905 */
906void
908{
909 int pgxactoff;
910
911 /*
912 * Currently we need to lock ProcArrayLock exclusively here, as we
913 * increment xactCompletionCount below. We also need it at least in shared
914 * mode for pgproc->pgxactoff to stay the same below.
915 *
916 * We could however, as this action does not actually change anyone's view
917 * of the set of running XIDs (our entry is duplicate with the gxact that
918 * has already been inserted into the ProcArray), lower the lock level to
919 * shared if we were to make xactCompletionCount an atomic variable. But
920 * that doesn't seem worth it currently, as a 2PC commit is heavyweight
921 * enough for this not to be the bottleneck. If it ever becomes a
922 * bottleneck it may also be worth considering to combine this with the
923 * subsequent ProcArrayRemove()
924 */
925 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
926
927 pgxactoff = proc->pgxactoff;
928
931
934 proc->recoveryConflictPending = false;
935
937 Assert(!proc->delayChkptFlags);
938
939 /*
940 * Need to increment completion count even though transaction hasn't
941 * really committed yet. The reason for that is that GetSnapshotData()
942 * omits the xid of the current transaction, thus without the increment we
943 * otherwise could end up reusing the snapshot later. Which would be bad,
944 * because it might not count the prepared transaction as running.
945 */
947
948 /* Clear the subtransaction-XID cache too */
949 Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
951 if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
952 {
953 ProcGlobal->subxidStates[pgxactoff].count = 0;
954 ProcGlobal->subxidStates[pgxactoff].overflowed = false;
955 proc->subxidStatus.count = 0;
956 proc->subxidStatus.overflowed = false;
957 }
958
959 LWLockRelease(ProcArrayLock);
960}
961
962/*
963 * Update TransamVariables->latestCompletedXid to point to latestXid if
964 * currently older.
965 */
966static void
968{
970
971 Assert(FullTransactionIdIsValid(cur_latest));
973 Assert(LWLockHeldByMe(ProcArrayLock));
974
975 if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
976 {
978 FullXidRelativeTo(cur_latest, latestXid);
979 }
980
983}
984
985/*
986 * Same as MaintainLatestCompletedXid, except for use during WAL replay.
987 */
988static void
990{
993
995 Assert(LWLockHeldByMe(ProcArrayLock));
996
997 /*
998 * Need a FullTransactionId to compare latestXid with. Can't rely on
999 * latestCompletedXid to be initialized in recovery. But in recovery it's
1000 * safe to access nextXid without a lock for the startup process.
1001 */
1004
1005 if (!FullTransactionIdIsValid(cur_latest) ||
1006 TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1007 {
1009 FullXidRelativeTo(rel, latestXid);
1010 }
1011
1013}
1014
1015/*
1016 * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1017 *
1018 * Remember up to where the startup process initialized the CLOG and subtrans
1019 * so we can ensure it's initialized gaplessly up to the point where necessary
1020 * while in recovery.
1021 */
1022void
1024{
1026 Assert(TransactionIdIsNormal(initializedUptoXID));
1027
1028 /*
1029 * we set latestObservedXid to the xid SUBTRANS has been initialized up
1030 * to, so we can extend it from that point onwards in
1031 * RecordKnownAssignedTransactionIds, and when we get consistent in
1032 * ProcArrayApplyRecoveryInfo().
1033 */
1034 latestObservedXid = initializedUptoXID;
1036}
1037
1038/*
1039 * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1040 *
1041 * Takes us through 3 states: Initialized, Pending and Ready.
1042 * Normal case is to go all the way to Ready straight away, though there
1043 * are atypical cases where we need to take it in steps.
1044 *
1045 * Use the data about running transactions on the primary to create the initial
1046 * state of KnownAssignedXids. We also use these records to regularly prune
1047 * KnownAssignedXids because we know it is possible that some transactions
1048 * with FATAL errors fail to write abort records, which could cause eventual
1049 * overflow.
1050 *
1051 * See comments for LogStandbySnapshot().
1052 */
1053void
1055{
1056 TransactionId *xids;
1057 TransactionId advanceNextXid;
1058 int nxids;
1059 int i;
1060
1065
1066 /*
1067 * Remove stale transactions, if any.
1068 */
1070
1071 /*
1072 * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1073 * because we will need it up to date for accessing two-phase transactions
1074 * in StandbyReleaseOldLocks().
1075 */
1076 advanceNextXid = running->nextXid;
1077 TransactionIdRetreat(advanceNextXid);
1080
1081 /*
1082 * Remove stale locks, if any.
1083 */
1085
1086 /*
1087 * If our snapshot is already valid, nothing else to do...
1088 */
1090 return;
1091
1092 /*
1093 * If our initial RunningTransactionsData had an overflowed snapshot then
1094 * we knew we were missing some subxids from our snapshot. If we continue
1095 * to see overflowed snapshots then we might never be able to start up, so
1096 * we make another test to see if our snapshot is now valid. We know that
1097 * the missing subxids are equal to or earlier than nextXid. After we
1098 * initialise we continue to apply changes during recovery, so once the
1099 * oldestRunningXid is later than the nextXid from the initial snapshot we
1100 * know that we no longer have missing information and can mark the
1101 * snapshot as valid.
1102 */
1104 {
1105 /*
1106 * If the snapshot isn't overflowed or if its empty we can reset our
1107 * pending state and use this snapshot instead.
1108 */
1109 if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1110 {
1111 /*
1112 * If we have already collected known assigned xids, we need to
1113 * throw them away before we apply the recovery snapshot.
1114 */
1117 }
1118 else
1119 {
1121 running->oldestRunningXid))
1122 {
1124 elog(DEBUG1,
1125 "recovery snapshots are now enabled");
1126 }
1127 else
1128 elog(DEBUG1,
1129 "recovery snapshot waiting for non-overflowed snapshot or "
1130 "until oldest active xid on standby is at least %u (now %u)",
1132 running->oldestRunningXid);
1133 return;
1134 }
1135 }
1136
1138
1139 /*
1140 * NB: this can be reached at least twice, so make sure new code can deal
1141 * with that.
1142 */
1143
1144 /*
1145 * Nobody else is running yet, but take locks anyhow
1146 */
1147 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1148
1149 /*
1150 * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1151 * sort them first.
1152 *
1153 * Some of the new xids are top-level xids and some are subtransactions.
1154 * We don't call SubTransSetParent because it doesn't matter yet. If we
1155 * aren't overflowed then all xids will fit in snapshot and so we don't
1156 * need subtrans. If we later overflow, an xid assignment record will add
1157 * xids to subtrans. If RunningTransactionsData is overflowed then we
1158 * don't have enough information to correctly update subtrans anyway.
1159 */
1160
1161 /*
1162 * Allocate a temporary array to avoid modifying the array passed as
1163 * argument.
1164 */
1165 xids = palloc(sizeof(TransactionId) * (running->xcnt + running->subxcnt));
1166
1167 /*
1168 * Add to the temp array any xids which have not already completed.
1169 */
1170 nxids = 0;
1171 for (i = 0; i < running->xcnt + running->subxcnt; i++)
1172 {
1173 TransactionId xid = running->xids[i];
1174
1175 /*
1176 * The running-xacts snapshot can contain xids that were still visible
1177 * in the procarray when the snapshot was taken, but were already
1178 * WAL-logged as completed. They're not running anymore, so ignore
1179 * them.
1180 */
1182 continue;
1183
1184 xids[nxids++] = xid;
1185 }
1186
1187 if (nxids > 0)
1188 {
1190 {
1191 LWLockRelease(ProcArrayLock);
1192 elog(ERROR, "KnownAssignedXids is not empty");
1193 }
1194
1195 /*
1196 * Sort the array so that we can add them safely into
1197 * KnownAssignedXids.
1198 *
1199 * We have to sort them logically, because in KnownAssignedXidsAdd we
1200 * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1201 * come from RUNNING_XACTS, which means there are only normal XIDs
1202 * from the same epoch, so this is safe.
1203 */
1204 qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1205
1206 /*
1207 * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1208 * snapshot may include duplicated xids because of prepared
1209 * transactions, so ignore them.
1210 */
1211 for (i = 0; i < nxids; i++)
1212 {
1213 if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1214 {
1215 elog(DEBUG1,
1216 "found duplicated transaction %u for KnownAssignedXids insertion",
1217 xids[i]);
1218 continue;
1219 }
1220 KnownAssignedXidsAdd(xids[i], xids[i], true);
1221 }
1222
1224 }
1225
1226 pfree(xids);
1227
1228 /*
1229 * latestObservedXid is at least set to the point where SUBTRANS was
1230 * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1231 * RecordKnownAssignedTransactionIds() was called for. Initialize
1232 * subtrans from thereon, up to nextXid - 1.
1233 *
1234 * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1235 * because we've just added xids to the known assigned xids machinery that
1236 * haven't gone through RecordKnownAssignedTransactionId().
1237 */
1241 {
1244 }
1245 TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1246
1247 /* ----------
1248 * Now we've got the running xids we need to set the global values that
1249 * are used to track snapshots as they evolve further.
1250 *
1251 * - latestCompletedXid which will be the xmax for snapshots
1252 * - lastOverflowedXid which shows whether snapshots overflow
1253 * - nextXid
1254 *
1255 * If the snapshot overflowed, then we still initialise with what we know,
1256 * but the recovery snapshot isn't fully valid yet because we know there
1257 * are some subxids missing. We don't know the specific subxids that are
1258 * missing, so conservatively assume the last one is latestObservedXid.
1259 * ----------
1260 */
1261 if (running->subxid_status == SUBXIDS_MISSING)
1262 {
1264
1267 }
1268 else
1269 {
1271
1273
1274 /*
1275 * If the 'xids' array didn't include all subtransactions, we have to
1276 * mark any snapshots taken as overflowed.
1277 */
1278 if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1280 else
1281 {
1284 }
1285 }
1286
1287 /*
1288 * If a transaction wrote a commit record in the gap between taking and
1289 * logging the snapshot then latestCompletedXid may already be higher than
1290 * the value from the snapshot, so check before we use the incoming value.
1291 * It also might not yet be set at all.
1292 */
1294
1295 /*
1296 * NB: No need to increment TransamVariables->xactCompletionCount here,
1297 * nobody can see it yet.
1298 */
1299
1300 LWLockRelease(ProcArrayLock);
1301
1304 elog(DEBUG1, "recovery snapshots are now enabled");
1305 else
1306 elog(DEBUG1,
1307 "recovery snapshot waiting for non-overflowed snapshot or "
1308 "until oldest active xid on standby is at least %u (now %u)",
1310 running->oldestRunningXid);
1311}
1312
1313/*
1314 * ProcArrayApplyXidAssignment
1315 * Process an XLOG_XACT_ASSIGNMENT WAL record
1316 */
1317void
1319 int nsubxids, TransactionId *subxids)
1320{
1321 TransactionId max_xid;
1322 int i;
1323
1325
1326 max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1327
1328 /*
1329 * Mark all the subtransactions as observed.
1330 *
1331 * NOTE: This will fail if the subxid contains too many previously
1332 * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1333 * as the code stands, because xid-assignment records should never contain
1334 * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1335 */
1337
1338 /*
1339 * Notice that we update pg_subtrans with the top-level xid, rather than
1340 * the parent xid. This is a difference between normal processing and
1341 * recovery, yet is still correct in all cases. The reason is that
1342 * subtransaction commit is not marked in clog until commit processing, so
1343 * all aborted subtransactions have already been clearly marked in clog.
1344 * As a result we are able to refer directly to the top-level
1345 * transaction's state rather than skipping through all the intermediate
1346 * states in the subtransaction tree. This should be the first time we
1347 * have attempted to SubTransSetParent().
1348 */
1349 for (i = 0; i < nsubxids; i++)
1350 SubTransSetParent(subxids[i], topxid);
1351
1352 /* KnownAssignedXids isn't maintained yet, so we're done for now */
1354 return;
1355
1356 /*
1357 * Uses same locking as transaction commit
1358 */
1359 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1360
1361 /*
1362 * Remove subxids from known-assigned-xacts.
1363 */
1365
1366 /*
1367 * Advance lastOverflowedXid to be at least the last of these subxids.
1368 */
1370 procArray->lastOverflowedXid = max_xid;
1371
1372 LWLockRelease(ProcArrayLock);
1373}
1374
1375/*
1376 * TransactionIdIsInProgress -- is given transaction running in some backend
1377 *
1378 * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1379 * there are four possibilities for finding a running transaction:
1380 *
1381 * 1. The given Xid is a main transaction Id. We will find this out cheaply
1382 * by looking at ProcGlobal->xids.
1383 *
1384 * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1385 * We can find this out cheaply too.
1386 *
1387 * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1388 * if the Xid is running on the primary.
1389 *
1390 * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1391 * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1392 * This is the slowest way, but sadly it has to be done always if the others
1393 * failed, unless we see that the cached subxact sets are complete (none have
1394 * overflowed).
1395 *
1396 * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1397 * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1398 * This buys back some concurrency (and we can't retrieve the main Xids from
1399 * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1400 */
1401bool
1403{
1404 static TransactionId *xids = NULL;
1405 static TransactionId *other_xids;
1406 XidCacheStatus *other_subxidstates;
1407 int nxids = 0;
1408 ProcArrayStruct *arrayP = procArray;
1409 TransactionId topxid;
1410 TransactionId latestCompletedXid;
1411 int mypgxactoff;
1412 int numProcs;
1413 int j;
1414
1415 /*
1416 * Don't bother checking a transaction older than RecentXmin; it could not
1417 * possibly still be running. (Note: in particular, this guarantees that
1418 * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1419 * running.)
1420 */
1422 {
1424 return false;
1425 }
1426
1427 /*
1428 * We may have just checked the status of this transaction, so if it is
1429 * already known to be completed, we can fall out without any access to
1430 * shared memory.
1431 */
1433 {
1435 return false;
1436 }
1437
1438 /*
1439 * Also, we can handle our own transaction (and subtransactions) without
1440 * any access to shared memory.
1441 */
1443 {
1445 return true;
1446 }
1447
1448 /*
1449 * If first time through, get workspace to remember main XIDs in. We
1450 * malloc it permanently to avoid repeated palloc/pfree overhead.
1451 */
1452 if (xids == NULL)
1453 {
1454 /*
1455 * In hot standby mode, reserve enough space to hold all xids in the
1456 * known-assigned list. If we later finish recovery, we no longer need
1457 * the bigger array, but we don't bother to shrink it.
1458 */
1459 int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1460
1461 xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1462 if (xids == NULL)
1463 ereport(ERROR,
1464 (errcode(ERRCODE_OUT_OF_MEMORY),
1465 errmsg("out of memory")));
1466 }
1467
1468 other_xids = ProcGlobal->xids;
1469 other_subxidstates = ProcGlobal->subxidStates;
1470
1471 LWLockAcquire(ProcArrayLock, LW_SHARED);
1472
1473 /*
1474 * Now that we have the lock, we can check latestCompletedXid; if the
1475 * target Xid is after that, it's surely still running.
1476 */
1477 latestCompletedXid =
1479 if (TransactionIdPrecedes(latestCompletedXid, xid))
1480 {
1481 LWLockRelease(ProcArrayLock);
1483 return true;
1484 }
1485
1486 /* No shortcuts, gotta grovel through the array */
1487 mypgxactoff = MyProc->pgxactoff;
1488 numProcs = arrayP->numProcs;
1489 for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1490 {
1491 int pgprocno;
1492 PGPROC *proc;
1493 TransactionId pxid;
1494 int pxids;
1495
1496 /* Ignore ourselves --- dealt with it above */
1497 if (pgxactoff == mypgxactoff)
1498 continue;
1499
1500 /* Fetch xid just once - see GetNewTransactionId */
1501 pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1502
1503 if (!TransactionIdIsValid(pxid))
1504 continue;
1505
1506 /*
1507 * Step 1: check the main Xid
1508 */
1509 if (TransactionIdEquals(pxid, xid))
1510 {
1511 LWLockRelease(ProcArrayLock);
1513 return true;
1514 }
1515
1516 /*
1517 * We can ignore main Xids that are younger than the target Xid, since
1518 * the target could not possibly be their child.
1519 */
1520 if (TransactionIdPrecedes(xid, pxid))
1521 continue;
1522
1523 /*
1524 * Step 2: check the cached child-Xids arrays
1525 */
1526 pxids = other_subxidstates[pgxactoff].count;
1527 pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1528 pgprocno = arrayP->pgprocnos[pgxactoff];
1529 proc = &allProcs[pgprocno];
1530 for (j = pxids - 1; j >= 0; j--)
1531 {
1532 /* Fetch xid just once - see GetNewTransactionId */
1534
1535 if (TransactionIdEquals(cxid, xid))
1536 {
1537 LWLockRelease(ProcArrayLock);
1539 return true;
1540 }
1541 }
1542
1543 /*
1544 * Save the main Xid for step 4. We only need to remember main Xids
1545 * that have uncached children. (Note: there is no race condition
1546 * here because the overflowed flag cannot be cleared, only set, while
1547 * we hold ProcArrayLock. So we can't miss an Xid that we need to
1548 * worry about.)
1549 */
1550 if (other_subxidstates[pgxactoff].overflowed)
1551 xids[nxids++] = pxid;
1552 }
1553
1554 /*
1555 * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1556 * in the list must be treated as running.
1557 */
1558 if (RecoveryInProgress())
1559 {
1560 /* none of the PGPROC entries should have XIDs in hot standby mode */
1561 Assert(nxids == 0);
1562
1563 if (KnownAssignedXidExists(xid))
1564 {
1565 LWLockRelease(ProcArrayLock);
1567 return true;
1568 }
1569
1570 /*
1571 * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1572 * too. Fetch all xids from KnownAssignedXids that are lower than
1573 * xid, since if xid is a subtransaction its parent will always have a
1574 * lower value. Note we will collect both main and subXIDs here, but
1575 * there's no help for it.
1576 */
1578 nxids = KnownAssignedXidsGet(xids, xid);
1579 }
1580
1581 LWLockRelease(ProcArrayLock);
1582
1583 /*
1584 * If none of the relevant caches overflowed, we know the Xid is not
1585 * running without even looking at pg_subtrans.
1586 */
1587 if (nxids == 0)
1588 {
1591 return false;
1592 }
1593
1594 /*
1595 * Step 4: have to check pg_subtrans.
1596 *
1597 * At this point, we know it's either a subtransaction of one of the Xids
1598 * in xids[], or it's not running. If it's an already-failed
1599 * subtransaction, we want to say "not running" even though its parent may
1600 * still be running. So first, check pg_xact to see if it's been aborted.
1601 */
1603
1604 if (TransactionIdDidAbort(xid))
1605 {
1607 return false;
1608 }
1609
1610 /*
1611 * It isn't aborted, so check whether the transaction tree it belongs to
1612 * is still running (or, more precisely, whether it was running when we
1613 * held ProcArrayLock).
1614 */
1615 topxid = SubTransGetTopmostTransaction(xid);
1617 if (!TransactionIdEquals(topxid, xid) &&
1618 pg_lfind32(topxid, xids, nxids))
1619 return true;
1620
1622 return false;
1623}
1624
1625
1626/*
1627 * Determine XID horizons.
1628 *
1629 * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1630 * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1631 * well as "internally" by GlobalVisUpdate() (see comment above struct
1632 * GlobalVisState).
1633 *
1634 * See the definition of ComputeXidHorizonsResult for the various computed
1635 * horizons.
1636 *
1637 * For VACUUM separate horizons (used to decide which deleted tuples must
1638 * be preserved), for shared and non-shared tables are computed. For shared
1639 * relations backends in all databases must be considered, but for non-shared
1640 * relations that's not required, since only backends in my own database could
1641 * ever see the tuples in them. Also, we can ignore concurrently running lazy
1642 * VACUUMs because (a) they must be working on other tables, and (b) they
1643 * don't need to do snapshot-based lookups.
1644 *
1645 * This also computes a horizon used to truncate pg_subtrans. For that
1646 * backends in all databases have to be considered, and concurrently running
1647 * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1648 * accesses.
1649 *
1650 * Note: we include all currently running xids in the set of considered xids.
1651 * This ensures that if a just-started xact has not yet set its snapshot,
1652 * when it does set the snapshot it cannot set xmin less than what we compute.
1653 * See notes in src/backend/access/transam/README.
1654 *
1655 * Note: despite the above, it's possible for the calculated values to move
1656 * backwards on repeated calls. The calculated values are conservative, so
1657 * that anything older is definitely not considered as running by anyone
1658 * anymore, but the exact values calculated depend on a number of things. For
1659 * example, if there are no transactions running in the current database, the
1660 * horizon for normal tables will be latestCompletedXid. If a transaction
1661 * begins after that, its xmin will include in-progress transactions in other
1662 * databases that started earlier, so another call will return a lower value.
1663 * Nonetheless it is safe to vacuum a table in the current database with the
1664 * first result. There are also replication-related effects: a walsender
1665 * process can set its xmin based on transactions that are no longer running
1666 * on the primary but are still being replayed on the standby, thus possibly
1667 * making the values go backwards. In this case there is a possibility that
1668 * we lose data that the standby would like to have, but unless the standby
1669 * uses a replication slot to make its xmin persistent there is little we can
1670 * do about that --- data is only protected if the walsender runs continuously
1671 * while queries are executed on the standby. (The Hot Standby code deals
1672 * with such cases by failing standby queries that needed to access
1673 * already-removed data, so there's no integrity bug.)
1674 *
1675 * Note: the approximate horizons (see definition of GlobalVisState) are
1676 * updated by the computations done here. That's currently required for
1677 * correctness and a small optimization. Without doing so it's possible that
1678 * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1679 * horizon than later when deciding which tuples can be removed - which the
1680 * code doesn't expect (breaking HOT).
1681 */
1682static void
1684{
1685 ProcArrayStruct *arrayP = procArray;
1686 TransactionId kaxmin;
1687 bool in_recovery = RecoveryInProgress();
1688 TransactionId *other_xids = ProcGlobal->xids;
1689
1690 /* inferred after ProcArrayLock is released */
1692
1693 LWLockAcquire(ProcArrayLock, LW_SHARED);
1694
1696
1697 /*
1698 * We initialize the MIN() calculation with latestCompletedXid + 1. This
1699 * is a lower bound for the XIDs that might appear in the ProcArray later,
1700 * and so protects us against overestimating the result due to future
1701 * additions.
1702 */
1703 {
1704 TransactionId initial;
1705
1707 Assert(TransactionIdIsValid(initial));
1708 TransactionIdAdvance(initial);
1709
1710 h->oldest_considered_running = initial;
1711 h->shared_oldest_nonremovable = initial;
1712 h->data_oldest_nonremovable = initial;
1713
1714 /*
1715 * Only modifications made by this backend affect the horizon for
1716 * temporary relations. Instead of a check in each iteration of the
1717 * loop over all PGPROCs it is cheaper to just initialize to the
1718 * current top-level xid any.
1719 *
1720 * Without an assigned xid we could use a horizon as aggressive as
1721 * GetNewTransactionId(), but we can get away with the much cheaper
1722 * latestCompletedXid + 1: If this backend has no xid there, by
1723 * definition, can't be any newer changes in the temp table than
1724 * latestCompletedXid.
1725 */
1728 else
1729 h->temp_oldest_nonremovable = initial;
1730 }
1731
1732 /*
1733 * Fetch slot horizons while ProcArrayLock is held - the
1734 * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1735 * the lock.
1736 */
1739
1740 for (int index = 0; index < arrayP->numProcs; index++)
1741 {
1742 int pgprocno = arrayP->pgprocnos[index];
1743 PGPROC *proc = &allProcs[pgprocno];
1744 int8 statusFlags = ProcGlobal->statusFlags[index];
1745 TransactionId xid;
1746 TransactionId xmin;
1747
1748 /* Fetch xid just once - see GetNewTransactionId */
1749 xid = UINT32_ACCESS_ONCE(other_xids[index]);
1750 xmin = UINT32_ACCESS_ONCE(proc->xmin);
1751
1752 /*
1753 * Consider both the transaction's Xmin, and its Xid.
1754 *
1755 * We must check both because a transaction might have an Xmin but not
1756 * (yet) an Xid; conversely, if it has an Xid, that could determine
1757 * some not-yet-set Xmin.
1758 */
1759 xmin = TransactionIdOlder(xmin, xid);
1760
1761 /* if neither is set, this proc doesn't influence the horizon */
1762 if (!TransactionIdIsValid(xmin))
1763 continue;
1764
1765 /*
1766 * Don't ignore any procs when determining which transactions might be
1767 * considered running. While slots should ensure logical decoding
1768 * backends are protected even without this check, it can't hurt to
1769 * include them here as well..
1770 */
1773
1774 /*
1775 * Skip over backends either vacuuming (which is ok with rows being
1776 * removed, as long as pg_subtrans is not truncated) or doing logical
1777 * decoding (which manages xmin separately, check below).
1778 */
1779 if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1780 continue;
1781
1782 /* shared tables need to take backends in all databases into account */
1785
1786 /*
1787 * Normally sessions in other databases are ignored for anything but
1788 * the shared horizon.
1789 *
1790 * However, include them when MyDatabaseId is not (yet) set. A
1791 * backend in the process of starting up must not compute a "too
1792 * aggressive" horizon, otherwise we could end up using it to prune
1793 * still-needed data away. If the current backend never connects to a
1794 * database this is harmless, because data_oldest_nonremovable will
1795 * never be utilized.
1796 *
1797 * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1798 * be included. (This flag is used for hot standby feedback, which
1799 * can't be tied to a specific database.)
1800 *
1801 * Also, while in recovery we cannot compute an accurate per-database
1802 * horizon, as all xids are managed via the KnownAssignedXids
1803 * machinery.
1804 */
1805 if (proc->databaseId == MyDatabaseId ||
1807 (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1808 in_recovery)
1809 {
1812 }
1813 }
1814
1815 /*
1816 * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1817 * after lock is released.
1818 */
1819 if (in_recovery)
1821
1822 /*
1823 * No other information from shared state is needed, release the lock
1824 * immediately. The rest of the computations can be done without a lock.
1825 */
1826 LWLockRelease(ProcArrayLock);
1827
1828 if (in_recovery)
1829 {
1836 /* temp relations cannot be accessed in recovery */
1837 }
1838
1843
1844 /*
1845 * Check whether there are replication slots requiring an older xmin.
1846 */
1851
1852 /*
1853 * The only difference between catalog / data horizons is that the slot's
1854 * catalog xmin is applied to the catalog one (so catalogs can be accessed
1855 * for logical decoding). Initialize with data horizon, and then back up
1856 * further if necessary. Have to back up the shared horizon as well, since
1857 * that also can contain catalogs.
1858 */
1867
1868 /*
1869 * It's possible that slots backed up the horizons further than
1870 * oldest_considered_running. Fix.
1871 */
1881
1882 /*
1883 * shared horizons have to be at least as old as the oldest visible in
1884 * current db
1885 */
1890
1891 /*
1892 * Horizons need to ensure that pg_subtrans access is still possible for
1893 * the relevant backends.
1894 */
1905 h->slot_xmin));
1908 h->slot_catalog_xmin));
1909
1910 /* update approximate horizons with the computed horizons */
1912}
1913
1914/*
1915 * Determine what kind of visibility horizon needs to be used for a
1916 * relation. If rel is NULL, the most conservative horizon is used.
1917 */
1918static inline GlobalVisHorizonKind
1920{
1921 /*
1922 * Other relkinds currently don't contain xids, nor always the necessary
1923 * logical decoding markers.
1924 */
1925 Assert(!rel ||
1926 rel->rd_rel->relkind == RELKIND_RELATION ||
1927 rel->rd_rel->relkind == RELKIND_MATVIEW ||
1928 rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1929
1930 if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1931 return VISHORIZON_SHARED;
1932 else if (IsCatalogRelation(rel) ||
1934 return VISHORIZON_CATALOG;
1935 else if (!RELATION_IS_LOCAL(rel))
1936 return VISHORIZON_DATA;
1937 else
1938 return VISHORIZON_TEMP;
1939}
1940
1941/*
1942 * Return the oldest XID for which deleted tuples must be preserved in the
1943 * passed table.
1944 *
1945 * If rel is not NULL the horizon may be considerably more recent than
1946 * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1947 * that is correct (but not optimal) for all relations will be returned.
1948 *
1949 * This is used by VACUUM to decide which deleted tuples must be preserved in
1950 * the passed in table.
1951 */
1954{
1955 ComputeXidHorizonsResult horizons;
1956
1957 ComputeXidHorizons(&horizons);
1958
1959 switch (GlobalVisHorizonKindForRel(rel))
1960 {
1961 case VISHORIZON_SHARED:
1962 return horizons.shared_oldest_nonremovable;
1963 case VISHORIZON_CATALOG:
1964 return horizons.catalog_oldest_nonremovable;
1965 case VISHORIZON_DATA:
1966 return horizons.data_oldest_nonremovable;
1967 case VISHORIZON_TEMP:
1968 return horizons.temp_oldest_nonremovable;
1969 }
1970
1971 /* just to prevent compiler warnings */
1972 return InvalidTransactionId;
1973}
1974
1975/*
1976 * Return the oldest transaction id any currently running backend might still
1977 * consider running. This should not be used for visibility / pruning
1978 * determinations (see GetOldestNonRemovableTransactionId()), but for
1979 * decisions like up to where pg_subtrans can be truncated.
1980 */
1983{
1984 ComputeXidHorizonsResult horizons;
1985
1986 ComputeXidHorizons(&horizons);
1987
1988 return horizons.oldest_considered_running;
1989}
1990
1991/*
1992 * Return the visibility horizons for a hot standby feedback message.
1993 */
1994void
1996{
1997 ComputeXidHorizonsResult horizons;
1998
1999 ComputeXidHorizons(&horizons);
2000
2001 /*
2002 * Don't want to use shared_oldest_nonremovable here, as that contains the
2003 * effect of replication slot's catalog_xmin. We want to send a separate
2004 * feedback for the catalog horizon, so the primary can remove data table
2005 * contents more aggressively.
2006 */
2007 *xmin = horizons.shared_oldest_nonremovable_raw;
2008 *catalog_xmin = horizons.slot_catalog_xmin;
2009}
2010
2011/*
2012 * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2013 *
2014 * We have to export this for use by snapmgr.c.
2015 */
2016int
2018{
2019 return procArray->maxProcs;
2020}
2021
2022/*
2023 * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2024 *
2025 * We have to export this for use by snapmgr.c.
2026 */
2027int
2029{
2031}
2032
2033/*
2034 * Helper function for GetSnapshotData() that checks if the bulk of the
2035 * visibility information in the snapshot is still valid. If so, it updates
2036 * the fields that need to change and returns true. Otherwise it returns
2037 * false.
2038 *
2039 * This very likely can be evolved to not need ProcArrayLock held (at very
2040 * least in the case we already hold a snapshot), but that's for another day.
2041 */
2042static bool
2044{
2045 uint64 curXactCompletionCount;
2046
2047 Assert(LWLockHeldByMe(ProcArrayLock));
2048
2049 if (unlikely(snapshot->snapXactCompletionCount == 0))
2050 return false;
2051
2052 curXactCompletionCount = TransamVariables->xactCompletionCount;
2053 if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2054 return false;
2055
2056 /*
2057 * If the current xactCompletionCount is still the same as it was at the
2058 * time the snapshot was built, we can be sure that rebuilding the
2059 * contents of the snapshot the hard way would result in the same snapshot
2060 * contents:
2061 *
2062 * As explained in transam/README, the set of xids considered running by
2063 * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2064 * contents only depend on transactions with xids and xactCompletionCount
2065 * is incremented whenever a transaction with an xid finishes (while
2066 * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2067 * ensures we would detect if the snapshot would have changed.
2068 *
2069 * As the snapshot contents are the same as it was before, it is safe to
2070 * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2071 * visible under the snapshot could already have been removed (that'd
2072 * require the set of running transactions to change) and it fulfills the
2073 * requirement that concurrent GetSnapshotData() calls yield the same
2074 * xmin.
2075 */
2077 MyProc->xmin = TransactionXmin = snapshot->xmin;
2078
2079 RecentXmin = snapshot->xmin;
2081
2082 snapshot->curcid = GetCurrentCommandId(false);
2083 snapshot->active_count = 0;
2084 snapshot->regd_count = 0;
2085 snapshot->copied = false;
2086
2087 return true;
2088}
2089
2090/*
2091 * GetSnapshotData -- returns information about running transactions.
2092 *
2093 * The returned snapshot includes xmin (lowest still-running xact ID),
2094 * xmax (highest completed xact ID + 1), and a list of running xact IDs
2095 * in the range xmin <= xid < xmax. It is used as follows:
2096 * All xact IDs < xmin are considered finished.
2097 * All xact IDs >= xmax are considered still running.
2098 * For an xact ID xmin <= xid < xmax, consult list to see whether
2099 * it is considered running or not.
2100 * This ensures that the set of transactions seen as "running" by the
2101 * current xact will not change after it takes the snapshot.
2102 *
2103 * All running top-level XIDs are included in the snapshot, except for lazy
2104 * VACUUM processes. We also try to include running subtransaction XIDs,
2105 * but since PGPROC has only a limited cache area for subxact XIDs, full
2106 * information may not be available. If we find any overflowed subxid arrays,
2107 * we have to mark the snapshot's subxid data as overflowed, and extra work
2108 * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2109 *
2110 * We also update the following backend-global variables:
2111 * TransactionXmin: the oldest xmin of any snapshot in use in the
2112 * current transaction (this is the same as MyProc->xmin).
2113 * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2114 * older than this are known not running any more.
2115 *
2116 * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2117 * for the benefit of the GlobalVisTest* family of functions.
2118 *
2119 * Note: this function should probably not be called with an argument that's
2120 * not statically allocated (see xip allocation below).
2121 */
2124{
2125 ProcArrayStruct *arrayP = procArray;
2126 TransactionId *other_xids = ProcGlobal->xids;
2127 TransactionId xmin;
2128 TransactionId xmax;
2129 int count = 0;
2130 int subcount = 0;
2131 bool suboverflowed = false;
2132 FullTransactionId latest_completed;
2133 TransactionId oldestxid;
2134 int mypgxactoff;
2135 TransactionId myxid;
2136 uint64 curXactCompletionCount;
2137
2138 TransactionId replication_slot_xmin = InvalidTransactionId;
2139 TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2140
2141 Assert(snapshot != NULL);
2142
2143 /*
2144 * Allocating space for maxProcs xids is usually overkill; numProcs would
2145 * be sufficient. But it seems better to do the malloc while not holding
2146 * the lock, so we can't look at numProcs. Likewise, we allocate much
2147 * more subxip storage than is probably needed.
2148 *
2149 * This does open a possibility for avoiding repeated malloc/free: since
2150 * maxProcs does not change at runtime, we can simply reuse the previous
2151 * xip arrays if any. (This relies on the fact that all callers pass
2152 * static SnapshotData structs.)
2153 */
2154 if (snapshot->xip == NULL)
2155 {
2156 /*
2157 * First call for this snapshot. Snapshot is same size whether or not
2158 * we are in recovery, see later comments.
2159 */
2160 snapshot->xip = (TransactionId *)
2162 if (snapshot->xip == NULL)
2163 ereport(ERROR,
2164 (errcode(ERRCODE_OUT_OF_MEMORY),
2165 errmsg("out of memory")));
2166 Assert(snapshot->subxip == NULL);
2167 snapshot->subxip = (TransactionId *)
2169 if (snapshot->subxip == NULL)
2170 ereport(ERROR,
2171 (errcode(ERRCODE_OUT_OF_MEMORY),
2172 errmsg("out of memory")));
2173 }
2174
2175 /*
2176 * It is sufficient to get shared lock on ProcArrayLock, even if we are
2177 * going to set MyProc->xmin.
2178 */
2179 LWLockAcquire(ProcArrayLock, LW_SHARED);
2180
2181 if (GetSnapshotDataReuse(snapshot))
2182 {
2183 LWLockRelease(ProcArrayLock);
2184 return snapshot;
2185 }
2186
2187 latest_completed = TransamVariables->latestCompletedXid;
2188 mypgxactoff = MyProc->pgxactoff;
2189 myxid = other_xids[mypgxactoff];
2190 Assert(myxid == MyProc->xid);
2191
2192 oldestxid = TransamVariables->oldestXid;
2193 curXactCompletionCount = TransamVariables->xactCompletionCount;
2194
2195 /* xmax is always latestCompletedXid + 1 */
2196 xmax = XidFromFullTransactionId(latest_completed);
2199
2200 /* initialize xmin calculation with xmax */
2201 xmin = xmax;
2202
2203 /* take own xid into account, saves a check inside the loop */
2204 if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2205 xmin = myxid;
2206
2208
2209 if (!snapshot->takenDuringRecovery)
2210 {
2211 int numProcs = arrayP->numProcs;
2212 TransactionId *xip = snapshot->xip;
2213 int *pgprocnos = arrayP->pgprocnos;
2214 XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2215 uint8 *allStatusFlags = ProcGlobal->statusFlags;
2216
2217 /*
2218 * First collect set of pgxactoff/xids that need to be included in the
2219 * snapshot.
2220 */
2221 for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2222 {
2223 /* Fetch xid just once - see GetNewTransactionId */
2224 TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2225 uint8 statusFlags;
2226
2227 Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2228
2229 /*
2230 * If the transaction has no XID assigned, we can skip it; it
2231 * won't have sub-XIDs either.
2232 */
2233 if (likely(xid == InvalidTransactionId))
2234 continue;
2235
2236 /*
2237 * We don't include our own XIDs (if any) in the snapshot. It
2238 * needs to be included in the xmin computation, but we did so
2239 * outside the loop.
2240 */
2241 if (pgxactoff == mypgxactoff)
2242 continue;
2243
2244 /*
2245 * The only way we are able to get here with a non-normal xid is
2246 * during bootstrap - with this backend using
2247 * BootstrapTransactionId. But the above test should filter that
2248 * out.
2249 */
2251
2252 /*
2253 * If the XID is >= xmax, we can skip it; such transactions will
2254 * be treated as running anyway (and any sub-XIDs will also be >=
2255 * xmax).
2256 */
2257 if (!NormalTransactionIdPrecedes(xid, xmax))
2258 continue;
2259
2260 /*
2261 * Skip over backends doing logical decoding which manages xmin
2262 * separately (check below) and ones running LAZY VACUUM.
2263 */
2264 statusFlags = allStatusFlags[pgxactoff];
2265 if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2266 continue;
2267
2268 if (NormalTransactionIdPrecedes(xid, xmin))
2269 xmin = xid;
2270
2271 /* Add XID to snapshot. */
2272 xip[count++] = xid;
2273
2274 /*
2275 * Save subtransaction XIDs if possible (if we've already
2276 * overflowed, there's no point). Note that the subxact XIDs must
2277 * be later than their parent, so no need to check them against
2278 * xmin. We could filter against xmax, but it seems better not to
2279 * do that much work while holding the ProcArrayLock.
2280 *
2281 * The other backend can add more subxids concurrently, but cannot
2282 * remove any. Hence it's important to fetch nxids just once.
2283 * Should be safe to use memcpy, though. (We needn't worry about
2284 * missing any xids added concurrently, because they must postdate
2285 * xmax.)
2286 *
2287 * Again, our own XIDs are not included in the snapshot.
2288 */
2289 if (!suboverflowed)
2290 {
2291
2292 if (subxidStates[pgxactoff].overflowed)
2293 suboverflowed = true;
2294 else
2295 {
2296 int nsubxids = subxidStates[pgxactoff].count;
2297
2298 if (nsubxids > 0)
2299 {
2300 int pgprocno = pgprocnos[pgxactoff];
2301 PGPROC *proc = &allProcs[pgprocno];
2302
2303 pg_read_barrier(); /* pairs with GetNewTransactionId */
2304
2305 memcpy(snapshot->subxip + subcount,
2306 proc->subxids.xids,
2307 nsubxids * sizeof(TransactionId));
2308 subcount += nsubxids;
2309 }
2310 }
2311 }
2312 }
2313 }
2314 else
2315 {
2316 /*
2317 * We're in hot standby, so get XIDs from KnownAssignedXids.
2318 *
2319 * We store all xids directly into subxip[]. Here's why:
2320 *
2321 * In recovery we don't know which xids are top-level and which are
2322 * subxacts, a design choice that greatly simplifies xid processing.
2323 *
2324 * It seems like we would want to try to put xids into xip[] only, but
2325 * that is fairly small. We would either need to make that bigger or
2326 * to increase the rate at which we WAL-log xid assignment; neither is
2327 * an appealing choice.
2328 *
2329 * We could try to store xids into xip[] first and then into subxip[]
2330 * if there are too many xids. That only works if the snapshot doesn't
2331 * overflow because we do not search subxip[] in that case. A simpler
2332 * way is to just store all xids in the subxip array because this is
2333 * by far the bigger array. We just leave the xip array empty.
2334 *
2335 * Either way we need to change the way XidInMVCCSnapshot() works
2336 * depending upon when the snapshot was taken, or change normal
2337 * snapshot processing so it matches.
2338 *
2339 * Note: It is possible for recovery to end before we finish taking
2340 * the snapshot, and for newly assigned transaction ids to be added to
2341 * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2342 * those newly added transaction ids would be filtered away, so we
2343 * need not be concerned about them.
2344 */
2345 subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2346 xmax);
2347
2349 suboverflowed = true;
2350 }
2351
2352
2353 /*
2354 * Fetch into local variable while ProcArrayLock is held - the
2355 * LWLockRelease below is a barrier, ensuring this happens inside the
2356 * lock.
2357 */
2358 replication_slot_xmin = procArray->replication_slot_xmin;
2359 replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2360
2362 MyProc->xmin = TransactionXmin = xmin;
2363
2364 LWLockRelease(ProcArrayLock);
2365
2366 /* maintain state for GlobalVis* */
2367 {
2368 TransactionId def_vis_xid;
2369 TransactionId def_vis_xid_data;
2370 FullTransactionId def_vis_fxid;
2371 FullTransactionId def_vis_fxid_data;
2372 FullTransactionId oldestfxid;
2373
2374 /*
2375 * Converting oldestXid is only safe when xid horizon cannot advance,
2376 * i.e. holding locks. While we don't hold the lock anymore, all the
2377 * necessary data has been gathered with lock held.
2378 */
2379 oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2380
2381 /* Check whether there's a replication slot requiring an older xmin. */
2382 def_vis_xid_data =
2383 TransactionIdOlder(xmin, replication_slot_xmin);
2384
2385 /*
2386 * Rows in non-shared, non-catalog tables possibly could be vacuumed
2387 * if older than this xid.
2388 */
2389 def_vis_xid = def_vis_xid_data;
2390
2391 /*
2392 * Check whether there's a replication slot requiring an older catalog
2393 * xmin.
2394 */
2395 def_vis_xid =
2396 TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2397
2398 def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2399 def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2400
2401 /*
2402 * Check if we can increase upper bound. As a previous
2403 * GlobalVisUpdate() might have computed more aggressive values, don't
2404 * overwrite them if so.
2405 */
2407 FullTransactionIdNewer(def_vis_fxid,
2410 FullTransactionIdNewer(def_vis_fxid,
2413 FullTransactionIdNewer(def_vis_fxid_data,
2415 /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2416 if (TransactionIdIsNormal(myxid))
2418 FullXidRelativeTo(latest_completed, myxid);
2419 else
2420 {
2421 GlobalVisTempRels.definitely_needed = latest_completed;
2423 }
2424
2425 /*
2426 * Check if we know that we can initialize or increase the lower
2427 * bound. Currently the only cheap way to do so is to use
2428 * TransamVariables->oldestXid as input.
2429 *
2430 * We should definitely be able to do better. We could e.g. put a
2431 * global lower bound value into TransamVariables.
2432 */
2435 oldestfxid);
2438 oldestfxid);
2441 oldestfxid);
2442 /* accurate value known */
2444 }
2445
2446 RecentXmin = xmin;
2448
2449 snapshot->xmin = xmin;
2450 snapshot->xmax = xmax;
2451 snapshot->xcnt = count;
2452 snapshot->subxcnt = subcount;
2453 snapshot->suboverflowed = suboverflowed;
2454 snapshot->snapXactCompletionCount = curXactCompletionCount;
2455
2456 snapshot->curcid = GetCurrentCommandId(false);
2457
2458 /*
2459 * This is a new snapshot, so set both refcounts are zero, and mark it as
2460 * not copied in persistent memory.
2461 */
2462 snapshot->active_count = 0;
2463 snapshot->regd_count = 0;
2464 snapshot->copied = false;
2465
2466 return snapshot;
2467}
2468
2469/*
2470 * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2471 *
2472 * This is called when installing a snapshot imported from another
2473 * transaction. To ensure that OldestXmin doesn't go backwards, we must
2474 * check that the source transaction is still running, and we'd better do
2475 * that atomically with installing the new xmin.
2476 *
2477 * Returns true if successful, false if source xact is no longer running.
2478 */
2479bool
2481 VirtualTransactionId *sourcevxid)
2482{
2483 bool result = false;
2484 ProcArrayStruct *arrayP = procArray;
2485 int index;
2486
2488 if (!sourcevxid)
2489 return false;
2490
2491 /* Get lock so source xact can't end while we're doing this */
2492 LWLockAcquire(ProcArrayLock, LW_SHARED);
2493
2494 /*
2495 * Find the PGPROC entry of the source transaction. (This could use
2496 * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2497 * performance critical.)
2498 */
2499 for (index = 0; index < arrayP->numProcs; index++)
2500 {
2501 int pgprocno = arrayP->pgprocnos[index];
2502 PGPROC *proc = &allProcs[pgprocno];
2503 int statusFlags = ProcGlobal->statusFlags[index];
2504 TransactionId xid;
2505
2506 /* Ignore procs running LAZY VACUUM */
2507 if (statusFlags & PROC_IN_VACUUM)
2508 continue;
2509
2510 /* We are only interested in the specific virtual transaction. */
2511 if (proc->vxid.procNumber != sourcevxid->procNumber)
2512 continue;
2513 if (proc->vxid.lxid != sourcevxid->localTransactionId)
2514 continue;
2515
2516 /*
2517 * We check the transaction's database ID for paranoia's sake: if it's
2518 * in another DB then its xmin does not cover us. Caller should have
2519 * detected this already, so we just treat any funny cases as
2520 * "transaction not found".
2521 */
2522 if (proc->databaseId != MyDatabaseId)
2523 continue;
2524
2525 /*
2526 * Likewise, let's just make real sure its xmin does cover us.
2527 */
2528 xid = UINT32_ACCESS_ONCE(proc->xmin);
2529 if (!TransactionIdIsNormal(xid) ||
2531 continue;
2532
2533 /*
2534 * We're good. Install the new xmin. As in GetSnapshotData, set
2535 * TransactionXmin too. (Note that because snapmgr.c called
2536 * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2537 * we don't check that.)
2538 */
2539 MyProc->xmin = TransactionXmin = xmin;
2540
2541 result = true;
2542 break;
2543 }
2544
2545 LWLockRelease(ProcArrayLock);
2546
2547 return result;
2548}
2549
2550/*
2551 * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2552 *
2553 * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2554 * PGPROC of the transaction from which we imported the snapshot, rather than
2555 * an XID.
2556 *
2557 * Note that this function also copies statusFlags from the source `proc` in
2558 * order to avoid the case where MyProc's xmin needs to be skipped for
2559 * computing xid horizon.
2560 *
2561 * Returns true if successful, false if source xact is no longer running.
2562 */
2563bool
2565{
2566 bool result = false;
2567 TransactionId xid;
2568
2570 Assert(proc != NULL);
2571
2572 /*
2573 * Get an exclusive lock so that we can copy statusFlags from source proc.
2574 */
2575 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2576
2577 /*
2578 * Be certain that the referenced PGPROC has an advertised xmin which is
2579 * no later than the one we're installing, so that the system-wide xmin
2580 * can't go backwards. Also, make sure it's running in the same database,
2581 * so that the per-database xmin cannot go backwards.
2582 */
2583 xid = UINT32_ACCESS_ONCE(proc->xmin);
2584 if (proc->databaseId == MyDatabaseId &&
2585 TransactionIdIsNormal(xid) &&
2587 {
2588 /*
2589 * Install xmin and propagate the statusFlags that affect how the
2590 * value is interpreted by vacuum.
2591 */
2592 MyProc->xmin = TransactionXmin = xmin;
2593 MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2594 (proc->statusFlags & PROC_XMIN_FLAGS);
2596
2597 result = true;
2598 }
2599
2600 LWLockRelease(ProcArrayLock);
2601
2602 return result;
2603}
2604
2605/*
2606 * GetRunningTransactionData -- returns information about running transactions.
2607 *
2608 * Similar to GetSnapshotData but returns more information. We include
2609 * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2610 * prepared transactions.
2611 *
2612 * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2613 * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2614 * array until the caller has WAL-logged this snapshot, and releases the
2615 * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2616 * lock is released.
2617 *
2618 * The returned data structure is statically allocated; caller should not
2619 * modify it, and must not assume it is valid past the next call.
2620 *
2621 * This is never executed during recovery so there is no need to look at
2622 * KnownAssignedXids.
2623 *
2624 * Dummy PGPROCs from prepared transaction are included, meaning that this
2625 * may return entries with duplicated TransactionId values coming from
2626 * transaction finishing to prepare. Nothing is done about duplicated
2627 * entries here to not hold on ProcArrayLock more than necessary.
2628 *
2629 * We don't worry about updating other counters, we want to keep this as
2630 * simple as possible and leave GetSnapshotData() as the primary code for
2631 * that bookkeeping.
2632 *
2633 * Note that if any transaction has overflowed its cached subtransactions
2634 * then there is no real need include any subtransactions.
2635 */
2638{
2639 /* result workspace */
2640 static RunningTransactionsData CurrentRunningXactsData;
2641
2642 ProcArrayStruct *arrayP = procArray;
2643 TransactionId *other_xids = ProcGlobal->xids;
2644 RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2645 TransactionId latestCompletedXid;
2646 TransactionId oldestRunningXid;
2647 TransactionId oldestDatabaseRunningXid;
2648 TransactionId *xids;
2649 int index;
2650 int count;
2651 int subcount;
2652 bool suboverflowed;
2653
2655
2656 /*
2657 * Allocating space for maxProcs xids is usually overkill; numProcs would
2658 * be sufficient. But it seems better to do the malloc while not holding
2659 * the lock, so we can't look at numProcs. Likewise, we allocate much
2660 * more subxip storage than is probably needed.
2661 *
2662 * Should only be allocated in bgwriter, since only ever executed during
2663 * checkpoints.
2664 */
2665 if (CurrentRunningXacts->xids == NULL)
2666 {
2667 /*
2668 * First call
2669 */
2670 CurrentRunningXacts->xids = (TransactionId *)
2672 if (CurrentRunningXacts->xids == NULL)
2673 ereport(ERROR,
2674 (errcode(ERRCODE_OUT_OF_MEMORY),
2675 errmsg("out of memory")));
2676 }
2677
2678 xids = CurrentRunningXacts->xids;
2679
2680 count = subcount = 0;
2681 suboverflowed = false;
2682
2683 /*
2684 * Ensure that no xids enter or leave the procarray while we obtain
2685 * snapshot.
2686 */
2687 LWLockAcquire(ProcArrayLock, LW_SHARED);
2688 LWLockAcquire(XidGenLock, LW_SHARED);
2689
2690 latestCompletedXid =
2692 oldestDatabaseRunningXid = oldestRunningXid =
2694
2695 /*
2696 * Spin over procArray collecting all xids
2697 */
2698 for (index = 0; index < arrayP->numProcs; index++)
2699 {
2700 TransactionId xid;
2701
2702 /* Fetch xid just once - see GetNewTransactionId */
2703 xid = UINT32_ACCESS_ONCE(other_xids[index]);
2704
2705 /*
2706 * We don't need to store transactions that don't have a TransactionId
2707 * yet because they will not show as running on a standby server.
2708 */
2709 if (!TransactionIdIsValid(xid))
2710 continue;
2711
2712 /*
2713 * Be careful not to exclude any xids before calculating the values of
2714 * oldestRunningXid and suboverflowed, since these are used to clean
2715 * up transaction information held on standbys.
2716 */
2717 if (TransactionIdPrecedes(xid, oldestRunningXid))
2718 oldestRunningXid = xid;
2719
2720 /*
2721 * Also, update the oldest running xid within the current database. As
2722 * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2723 * TransactionId comparison first.
2724 */
2725 if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2726 {
2727 int pgprocno = arrayP->pgprocnos[index];
2728 PGPROC *proc = &allProcs[pgprocno];
2729
2730 if (proc->databaseId == MyDatabaseId)
2731 oldestDatabaseRunningXid = xid;
2732 }
2733
2735 suboverflowed = true;
2736
2737 /*
2738 * If we wished to exclude xids this would be the right place for it.
2739 * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2740 * but they do during truncation at the end when they get the lock and
2741 * truncate, so it is not much of a problem to include them if they
2742 * are seen and it is cleaner to include them.
2743 */
2744
2745 xids[count++] = xid;
2746 }
2747
2748 /*
2749 * Spin over procArray collecting all subxids, but only if there hasn't
2750 * been a suboverflow.
2751 */
2752 if (!suboverflowed)
2753 {
2754 XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2755
2756 for (index = 0; index < arrayP->numProcs; index++)
2757 {
2758 int pgprocno = arrayP->pgprocnos[index];
2759 PGPROC *proc = &allProcs[pgprocno];
2760 int nsubxids;
2761
2762 /*
2763 * Save subtransaction XIDs. Other backends can't add or remove
2764 * entries while we're holding XidGenLock.
2765 */
2766 nsubxids = other_subxidstates[index].count;
2767 if (nsubxids > 0)
2768 {
2769 /* barrier not really required, as XidGenLock is held, but ... */
2770 pg_read_barrier(); /* pairs with GetNewTransactionId */
2771
2772 memcpy(&xids[count], proc->subxids.xids,
2773 nsubxids * sizeof(TransactionId));
2774 count += nsubxids;
2775 subcount += nsubxids;
2776
2777 /*
2778 * Top-level XID of a transaction is always less than any of
2779 * its subxids, so we don't need to check if any of the
2780 * subxids are smaller than oldestRunningXid
2781 */
2782 }
2783 }
2784 }
2785
2786 /*
2787 * It's important *not* to include the limits set by slots here because
2788 * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2789 * were to be included here the initial value could never increase because
2790 * of a circular dependency where slots only increase their limits when
2791 * running xacts increases oldestRunningXid and running xacts only
2792 * increases if slots do.
2793 */
2794
2795 CurrentRunningXacts->xcnt = count - subcount;
2796 CurrentRunningXacts->subxcnt = subcount;
2797 CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2799 CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2800 CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2801 CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2802
2803 Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2804 Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2805 Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2806
2807 /* We don't release the locks here, the caller is responsible for that */
2808
2809 return CurrentRunningXacts;
2810}
2811
2812/*
2813 * GetOldestActiveTransactionId()
2814 *
2815 * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2816 * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2817 *
2818 * If allDbs is true, we look at all databases, though there is no need to
2819 * include WALSender since this has no effect on hot standby conflicts. If
2820 * allDbs is false, skip processes attached to other databases.
2821 *
2822 * This is never executed during recovery so there is no need to look at
2823 * KnownAssignedXids.
2824 *
2825 * We don't worry about updating other counters, we want to keep this as
2826 * simple as possible and leave GetSnapshotData() as the primary code for
2827 * that bookkeeping.
2828 *
2829 * inCommitOnly indicates getting the oldestActiveXid among the transactions
2830 * in the commit critical section.
2831 */
2833GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
2834{
2835 ProcArrayStruct *arrayP = procArray;
2836 TransactionId *other_xids = ProcGlobal->xids;
2837 TransactionId oldestRunningXid;
2838 int index;
2839
2841
2842 /*
2843 * Read nextXid, as the upper bound of what's still active.
2844 *
2845 * Reading a TransactionId is atomic, but we must grab the lock to make
2846 * sure that all XIDs < nextXid are already present in the proc array (or
2847 * have already completed), when we spin over it.
2848 */
2849 LWLockAcquire(XidGenLock, LW_SHARED);
2851 LWLockRelease(XidGenLock);
2852
2853 /*
2854 * Spin over procArray collecting all xids and subxids.
2855 */
2856 LWLockAcquire(ProcArrayLock, LW_SHARED);
2857 for (index = 0; index < arrayP->numProcs; index++)
2858 {
2859 TransactionId xid;
2860 int pgprocno = arrayP->pgprocnos[index];
2861 PGPROC *proc = &allProcs[pgprocno];
2862
2863 /* Fetch xid just once - see GetNewTransactionId */
2864 xid = UINT32_ACCESS_ONCE(other_xids[index]);
2865
2866 if (!TransactionIdIsNormal(xid))
2867 continue;
2868
2869 if (inCommitOnly &&
2871 continue;
2872
2873 if (!allDbs && proc->databaseId != MyDatabaseId)
2874 continue;
2875
2876 if (TransactionIdPrecedes(xid, oldestRunningXid))
2877 oldestRunningXid = xid;
2878
2879 /*
2880 * Top-level XID of a transaction is always less than any of its
2881 * subxids, so we don't need to check if any of the subxids are
2882 * smaller than oldestRunningXid
2883 */
2884 }
2885 LWLockRelease(ProcArrayLock);
2886
2887 return oldestRunningXid;
2888}
2889
2890/*
2891 * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2892 *
2893 * Returns the oldest xid that we can guarantee not to have been affected by
2894 * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2895 * transaction aborted. Note that the value can (and most of the time will) be
2896 * much more conservative than what really has been affected by vacuum, but we
2897 * currently don't have better data available.
2898 *
2899 * This is useful to initialize the cutoff xid after which a new changeset
2900 * extraction replication slot can start decoding changes.
2901 *
2902 * Must be called with ProcArrayLock held either shared or exclusively,
2903 * although most callers will want to use exclusive mode since it is expected
2904 * that the caller will immediately use the xid to peg the xmin horizon.
2905 */
2908{
2909 ProcArrayStruct *arrayP = procArray;
2910 TransactionId oldestSafeXid;
2911 int index;
2912 bool recovery_in_progress = RecoveryInProgress();
2913
2914 Assert(LWLockHeldByMe(ProcArrayLock));
2915
2916 /*
2917 * Acquire XidGenLock, so no transactions can acquire an xid while we're
2918 * running. If no transaction with xid were running concurrently a new xid
2919 * could influence the RecentXmin et al.
2920 *
2921 * We initialize the computation to nextXid since that's guaranteed to be
2922 * a safe, albeit pessimal, value.
2923 */
2924 LWLockAcquire(XidGenLock, LW_SHARED);
2926
2927 /*
2928 * If there's already a slot pegging the xmin horizon, we can start with
2929 * that value, it's guaranteed to be safe since it's computed by this
2930 * routine initially and has been enforced since. We can always use the
2931 * slot's general xmin horizon, but the catalog horizon is only usable
2932 * when only catalog data is going to be looked at.
2933 */
2936 oldestSafeXid))
2937 oldestSafeXid = procArray->replication_slot_xmin;
2938
2939 if (catalogOnly &&
2942 oldestSafeXid))
2944
2945 /*
2946 * If we're not in recovery, we walk over the procarray and collect the
2947 * lowest xid. Since we're called with ProcArrayLock held and have
2948 * acquired XidGenLock, no entries can vanish concurrently, since
2949 * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2950 * with ProcArrayLock held.
2951 *
2952 * In recovery we can't lower the safe value besides what we've computed
2953 * above, so we'll have to wait a bit longer there. We unfortunately can
2954 * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2955 * machinery can miss values and return an older value than is safe.
2956 */
2957 if (!recovery_in_progress)
2958 {
2959 TransactionId *other_xids = ProcGlobal->xids;
2960
2961 /*
2962 * Spin over procArray collecting min(ProcGlobal->xids[i])
2963 */
2964 for (index = 0; index < arrayP->numProcs; index++)
2965 {
2966 TransactionId xid;
2967
2968 /* Fetch xid just once - see GetNewTransactionId */
2969 xid = UINT32_ACCESS_ONCE(other_xids[index]);
2970
2971 if (!TransactionIdIsNormal(xid))
2972 continue;
2973
2974 if (TransactionIdPrecedes(xid, oldestSafeXid))
2975 oldestSafeXid = xid;
2976 }
2977 }
2978
2979 LWLockRelease(XidGenLock);
2980
2981 return oldestSafeXid;
2982}
2983
2984/*
2985 * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2986 * delaying checkpoint because they have critical actions in progress.
2987 *
2988 * Constructs an array of VXIDs of transactions that are currently in commit
2989 * critical sections, as shown by having specified delayChkptFlags bits set
2990 * in their PGPROC.
2991 *
2992 * Returns a palloc'd array that should be freed by the caller.
2993 * *nvxids is the number of valid entries.
2994 *
2995 * Note that because backends set or clear delayChkptFlags without holding any
2996 * lock, the result is somewhat indeterminate, but we don't really care. Even
2997 * in a multiprocessor with delayed writes to shared memory, it should be
2998 * certain that setting of delayChkptFlags will propagate to shared memory
2999 * when the backend takes a lock, so we cannot fail to see a virtual xact as
3000 * delayChkptFlags if it's already inserted its commit record. Whether it
3001 * takes a little while for clearing of delayChkptFlags to propagate is
3002 * unimportant for correctness.
3003 */
3006{
3007 VirtualTransactionId *vxids;
3008 ProcArrayStruct *arrayP = procArray;
3009 int count = 0;
3010 int index;
3011
3012 Assert(type != 0);
3013
3014 /* allocate what's certainly enough result space */
3015 vxids = (VirtualTransactionId *)
3016 palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3017
3018 LWLockAcquire(ProcArrayLock, LW_SHARED);
3019
3020 for (index = 0; index < arrayP->numProcs; index++)
3021 {
3022 int pgprocno = arrayP->pgprocnos[index];
3023 PGPROC *proc = &allProcs[pgprocno];
3024
3025 if ((proc->delayChkptFlags & type) != 0)
3026 {
3028
3029 GET_VXID_FROM_PGPROC(vxid, *proc);
3031 vxids[count++] = vxid;
3032 }
3033 }
3034
3035 LWLockRelease(ProcArrayLock);
3036
3037 *nvxids = count;
3038 return vxids;
3039}
3040
3041/*
3042 * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3043 *
3044 * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3045 * of the specified VXIDs are still in critical sections of code.
3046 *
3047 * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3048 * those numbers should be small enough for it not to be a problem.
3049 */
3050bool
3052{
3053 bool result = false;
3054 ProcArrayStruct *arrayP = procArray;
3055 int index;
3056
3057 Assert(type != 0);
3058
3059 LWLockAcquire(ProcArrayLock, LW_SHARED);
3060
3061 for (index = 0; index < arrayP->numProcs; index++)
3062 {
3063 int pgprocno = arrayP->pgprocnos[index];
3064 PGPROC *proc = &allProcs[pgprocno];
3066
3067 GET_VXID_FROM_PGPROC(vxid, *proc);
3068
3069 if ((proc->delayChkptFlags & type) != 0 &&
3071 {
3072 int i;
3073
3074 for (i = 0; i < nvxids; i++)
3075 {
3076 if (VirtualTransactionIdEquals(vxid, vxids[i]))
3077 {
3078 result = true;
3079 break;
3080 }
3081 }
3082 if (result)
3083 break;
3084 }
3085 }
3086
3087 LWLockRelease(ProcArrayLock);
3088
3089 return result;
3090}
3091
3092/*
3093 * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3094 *
3095 * The result may be out of date arbitrarily quickly, so the caller
3096 * must be careful about how this information is used. NULL is
3097 * returned if the backend is not active.
3098 */
3099PGPROC *
3101{
3102 PGPROC *result;
3103
3104 if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3105 return NULL;
3106 result = GetPGProcByNumber(procNumber);
3107
3108 if (result->pid == 0)
3109 return NULL;
3110
3111 return result;
3112}
3113
3114/*
3115 * ProcNumberGetTransactionIds -- get a backend's transaction status
3116 *
3117 * Get the xid, xmin, nsubxid and overflow status of the backend. The
3118 * result may be out of date arbitrarily quickly, so the caller must be
3119 * careful about how this information is used.
3120 */
3121void
3123 TransactionId *xmin, int *nsubxid, bool *overflowed)
3124{
3125 PGPROC *proc;
3126
3127 *xid = InvalidTransactionId;
3128 *xmin = InvalidTransactionId;
3129 *nsubxid = 0;
3130 *overflowed = false;
3131
3132 if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3133 return;
3134 proc = GetPGProcByNumber(procNumber);
3135
3136 /* Need to lock out additions/removals of backends */
3137 LWLockAcquire(ProcArrayLock, LW_SHARED);
3138
3139 if (proc->pid != 0)
3140 {
3141 *xid = proc->xid;
3142 *xmin = proc->xmin;
3143 *nsubxid = proc->subxidStatus.count;
3144 *overflowed = proc->subxidStatus.overflowed;
3145 }
3146
3147 LWLockRelease(ProcArrayLock);
3148}
3149
3150/*
3151 * BackendPidGetProc -- get a backend's PGPROC given its PID
3152 *
3153 * Returns NULL if not found. Note that it is up to the caller to be
3154 * sure that the question remains meaningful for long enough for the
3155 * answer to be used ...
3156 */
3157PGPROC *
3159{
3160 PGPROC *result;
3161
3162 if (pid == 0) /* never match dummy PGPROCs */
3163 return NULL;
3164
3165 LWLockAcquire(ProcArrayLock, LW_SHARED);
3166
3167 result = BackendPidGetProcWithLock(pid);
3168
3169 LWLockRelease(ProcArrayLock);
3170
3171 return result;
3172}
3173
3174/*
3175 * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3176 *
3177 * Same as above, except caller must be holding ProcArrayLock. The found
3178 * entry, if any, can be assumed to be valid as long as the lock remains held.
3179 */
3180PGPROC *
3182{
3183 PGPROC *result = NULL;
3184 ProcArrayStruct *arrayP = procArray;
3185 int index;
3186
3187 if (pid == 0) /* never match dummy PGPROCs */
3188 return NULL;
3189
3190 for (index = 0; index < arrayP->numProcs; index++)
3191 {
3192 PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3193
3194 if (proc->pid == pid)
3195 {
3196 result = proc;
3197 break;
3198 }
3199 }
3200
3201 return result;
3202}
3203
3204/*
3205 * BackendXidGetPid -- get a backend's pid given its XID
3206 *
3207 * Returns 0 if not found or it's a prepared transaction. Note that
3208 * it is up to the caller to be sure that the question remains
3209 * meaningful for long enough for the answer to be used ...
3210 *
3211 * Only main transaction Ids are considered. This function is mainly
3212 * useful for determining what backend owns a lock.
3213 *
3214 * Beware that not every xact has an XID assigned. However, as long as you
3215 * only call this using an XID found on disk, you're safe.
3216 */
3217int
3219{
3220 int result = 0;
3221 ProcArrayStruct *arrayP = procArray;
3222 TransactionId *other_xids = ProcGlobal->xids;
3223 int index;
3224
3225 if (xid == InvalidTransactionId) /* never match invalid xid */
3226 return 0;
3227
3228 LWLockAcquire(ProcArrayLock, LW_SHARED);
3229
3230 for (index = 0; index < arrayP->numProcs; index++)
3231 {
3232 if (other_xids[index] == xid)
3233 {
3234 int pgprocno = arrayP->pgprocnos[index];
3235 PGPROC *proc = &allProcs[pgprocno];
3236
3237 result = proc->pid;
3238 break;
3239 }
3240 }
3241
3242 LWLockRelease(ProcArrayLock);
3243
3244 return result;
3245}
3246
3247/*
3248 * IsBackendPid -- is a given pid a running backend
3249 *
3250 * This is not called by the backend, but is called by external modules.
3251 */
3252bool
3254{
3255 return (BackendPidGetProc(pid) != NULL);
3256}
3257
3258
3259/*
3260 * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3261 *
3262 * The array is palloc'd. The number of valid entries is returned into *nvxids.
3263 *
3264 * The arguments allow filtering the set of VXIDs returned. Our own process
3265 * is always skipped. In addition:
3266 * If limitXmin is not InvalidTransactionId, skip processes with
3267 * xmin > limitXmin.
3268 * If excludeXmin0 is true, skip processes with xmin = 0.
3269 * If allDbs is false, skip processes attached to other databases.
3270 * If excludeVacuum isn't zero, skip processes for which
3271 * (statusFlags & excludeVacuum) is not zero.
3272 *
3273 * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3274 * allow skipping backends whose oldest live snapshot is no older than
3275 * some snapshot we have. Since we examine the procarray with only shared
3276 * lock, there are race conditions: a backend could set its xmin just after
3277 * we look. Indeed, on multiprocessors with weak memory ordering, the
3278 * other backend could have set its xmin *before* we look. We know however
3279 * that such a backend must have held shared ProcArrayLock overlapping our
3280 * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3281 * any snapshot the other backend is taking concurrently with our scan cannot
3282 * consider any transactions as still running that we think are committed
3283 * (since backends must hold ProcArrayLock exclusive to commit).
3284 */
3286GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3287 bool allDbs, int excludeVacuum,
3288 int *nvxids)
3289{
3290 VirtualTransactionId *vxids;
3291 ProcArrayStruct *arrayP = procArray;
3292 int count = 0;
3293 int index;
3294
3295 /* allocate what's certainly enough result space */
3296 vxids = (VirtualTransactionId *)
3297 palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3298
3299 LWLockAcquire(ProcArrayLock, LW_SHARED);
3300
3301 for (index = 0; index < arrayP->numProcs; index++)
3302 {
3303 int pgprocno = arrayP->pgprocnos[index];
3304 PGPROC *proc = &allProcs[pgprocno];
3305 uint8 statusFlags = ProcGlobal->statusFlags[index];
3306
3307 if (proc == MyProc)
3308 continue;
3309
3310 if (excludeVacuum & statusFlags)
3311 continue;
3312
3313 if (allDbs || proc->databaseId == MyDatabaseId)
3314 {
3315 /* Fetch xmin just once - might change on us */
3316 TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3317
3318 if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3319 continue;
3320
3321 /*
3322 * InvalidTransactionId precedes all other XIDs, so a proc that
3323 * hasn't set xmin yet will not be rejected by this test.
3324 */
3325 if (!TransactionIdIsValid(limitXmin) ||
3326 TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3327 {
3329
3330 GET_VXID_FROM_PGPROC(vxid, *proc);
3332 vxids[count++] = vxid;
3333 }
3334 }
3335 }
3336
3337 LWLockRelease(ProcArrayLock);
3338
3339 *nvxids = count;
3340 return vxids;
3341}
3342
3343/*
3344 * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3345 *
3346 * Usage is limited to conflict resolution during recovery on standby servers.
3347 * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3348 * semantics, or InvalidTransactionId in cases where caller cannot accurately
3349 * determine a safe snapshotConflictHorizon value.
3350 *
3351 * If limitXmin is InvalidTransactionId then we want to kill everybody,
3352 * so we're not worried if they have a snapshot or not, nor does it really
3353 * matter what type of lock we hold. Caller must avoid calling here with
3354 * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3355 * during original execution, since that actually indicates that there is
3356 * definitely no need for a recovery conflict (the snapshotConflictHorizon
3357 * convention for InvalidTransactionId values is the opposite of our own!).
3358 *
3359 * All callers that are checking xmins always now supply a valid and useful
3360 * value for limitXmin. The limitXmin is always lower than the lowest
3361 * numbered KnownAssignedXid that is not already a FATAL error. This is
3362 * because we only care about cleanup records that are cleaning up tuple
3363 * versions from committed transactions. In that case they will only occur
3364 * at the point where the record is less than the lowest running xid. That
3365 * allows us to say that if any backend takes a snapshot concurrently with
3366 * us then the conflict assessment made here would never include the snapshot
3367 * that is being derived. So we take LW_SHARED on the ProcArray and allow
3368 * concurrent snapshots when limitXmin is valid. We might think about adding
3369 * Assert(limitXmin < lowest(KnownAssignedXids))
3370 * but that would not be true in the case of FATAL errors lagging in array,
3371 * but we already know those are bogus anyway, so we skip that test.
3372 *
3373 * If dbOid is valid we skip backends attached to other databases.
3374 *
3375 * Be careful to *not* pfree the result from this function. We reuse
3376 * this array sufficiently often that we use malloc for the result.
3377 */
3380{
3381 static VirtualTransactionId *vxids;
3382 ProcArrayStruct *arrayP = procArray;
3383 int count = 0;
3384 int index;
3385
3386 /*
3387 * If first time through, get workspace to remember main XIDs in. We
3388 * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3389 * result space, remembering room for a terminator.
3390 */
3391 if (vxids == NULL)
3392 {
3393 vxids = (VirtualTransactionId *)
3394 malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3395 if (vxids == NULL)
3396 ereport(ERROR,
3397 (errcode(ERRCODE_OUT_OF_MEMORY),
3398 errmsg("out of memory")));
3399 }
3400
3401 LWLockAcquire(ProcArrayLock, LW_SHARED);
3402
3403 for (index = 0; index < arrayP->numProcs; index++)
3404 {
3405 int pgprocno = arrayP->pgprocnos[index];
3406 PGPROC *proc = &allProcs[pgprocno];
3407
3408 /* Exclude prepared transactions */
3409 if (proc->pid == 0)
3410 continue;
3411
3412 if (!OidIsValid(dbOid) ||
3413 proc->databaseId == dbOid)
3414 {
3415 /* Fetch xmin just once - can't change on us, but good coding */
3416 TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3417
3418 /*
3419 * We ignore an invalid pxmin because this means that backend has
3420 * no snapshot currently. We hold a Share lock to avoid contention
3421 * with users taking snapshots. That is not a problem because the
3422 * current xmin is always at least one higher than the latest
3423 * removed xid, so any new snapshot would never conflict with the
3424 * test here.
3425 */
3426 if (!TransactionIdIsValid(limitXmin) ||
3427 (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3428 {
3430
3431 GET_VXID_FROM_PGPROC(vxid, *proc);
3433 vxids[count++] = vxid;
3434 }
3435 }
3436 }
3437
3438 LWLockRelease(ProcArrayLock);
3439
3440 /* add the terminator */
3441 vxids[count].procNumber = INVALID_PROC_NUMBER;
3443
3444 return vxids;
3445}
3446
3447/*
3448 * CancelVirtualTransaction - used in recovery conflict processing
3449 *
3450 * Returns pid of the process signaled, or 0 if not found.
3451 */
3452pid_t
3454{
3455 return SignalVirtualTransaction(vxid, sigmode, true);
3456}
3457
3458pid_t
3460 bool conflictPending)
3461{
3462 ProcArrayStruct *arrayP = procArray;
3463 int index;
3464 pid_t pid = 0;
3465
3466 LWLockAcquire(ProcArrayLock, LW_SHARED);
3467
3468 for (index = 0; index < arrayP->numProcs; index++)
3469 {
3470 int pgprocno = arrayP->pgprocnos[index];
3471 PGPROC *proc = &allProcs[pgprocno];
3472 VirtualTransactionId procvxid;
3473
3474 GET_VXID_FROM_PGPROC(procvxid, *proc);
3475
3476 if (procvxid.procNumber == vxid.procNumber &&
3477 procvxid.localTransactionId == vxid.localTransactionId)
3478 {
3479 proc->recoveryConflictPending = conflictPending;
3480 pid = proc->pid;
3481 if (pid != 0)
3482 {
3483 /*
3484 * Kill the pid if it's still here. If not, that's what we
3485 * wanted so ignore any errors.
3486 */
3487 (void) SendProcSignal(pid, sigmode, vxid.procNumber);
3488 }
3489 break;
3490 }
3491 }
3492
3493 LWLockRelease(ProcArrayLock);
3494
3495 return pid;
3496}
3497
3498/*
3499 * MinimumActiveBackends --- count backends (other than myself) that are
3500 * in active transactions. Return true if the count exceeds the
3501 * minimum threshold passed. This is used as a heuristic to decide if
3502 * a pre-XLOG-flush delay is worthwhile during commit.
3503 *
3504 * Do not count backends that are blocked waiting for locks, since they are
3505 * not going to get to run until someone else commits.
3506 */
3507bool
3509{
3510 ProcArrayStruct *arrayP = procArray;
3511 int count = 0;
3512 int index;
3513
3514 /* Quick short-circuit if no minimum is specified */
3515 if (min == 0)
3516 return true;
3517
3518 /*
3519 * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3520 * bogus, but since we are only testing fields for zero or nonzero, it
3521 * should be OK. The result is only used for heuristic purposes anyway...
3522 */
3523 for (index = 0; index < arrayP->numProcs; index++)
3524 {
3525 int pgprocno = arrayP->pgprocnos[index];
3526 PGPROC *proc = &allProcs[pgprocno];
3527
3528 /*
3529 * Since we're not holding a lock, need to be prepared to deal with
3530 * garbage, as someone could have incremented numProcs but not yet
3531 * filled the structure.
3532 *
3533 * If someone just decremented numProcs, 'proc' could also point to a
3534 * PGPROC entry that's no longer in the array. It still points to a
3535 * PGPROC struct, though, because freed PGPROC entries just go to the
3536 * free list and are recycled. Its contents are nonsense in that case,
3537 * but that's acceptable for this function.
3538 */
3539 if (pgprocno == -1)
3540 continue; /* do not count deleted entries */
3541 if (proc == MyProc)
3542 continue; /* do not count myself */
3543 if (proc->xid == InvalidTransactionId)
3544 continue; /* do not count if no XID assigned */
3545 if (proc->pid == 0)
3546 continue; /* do not count prepared xacts */
3547 if (proc->waitLock != NULL)
3548 continue; /* do not count if blocked on a lock */
3549 count++;
3550 if (count >= min)
3551 break;
3552 }
3553
3554 return count >= min;
3555}
3556
3557/*
3558 * CountDBBackends --- count backends that are using specified database
3559 */
3560int
3562{
3563 ProcArrayStruct *arrayP = procArray;
3564 int count = 0;
3565 int index;
3566
3567 LWLockAcquire(ProcArrayLock, LW_SHARED);
3568
3569 for (index = 0; index < arrayP->numProcs; index++)
3570 {
3571 int pgprocno = arrayP->pgprocnos[index];
3572 PGPROC *proc = &allProcs[pgprocno];
3573
3574 if (proc->pid == 0)
3575 continue; /* do not count prepared xacts */
3576 if (!OidIsValid(databaseid) ||
3577 proc->databaseId == databaseid)
3578 count++;
3579 }
3580
3581 LWLockRelease(ProcArrayLock);
3582
3583 return count;
3584}
3585
3586/*
3587 * CountDBConnections --- counts database backends (only regular backends)
3588 */
3589int
3591{
3592 ProcArrayStruct *arrayP = procArray;
3593 int count = 0;
3594 int index;
3595
3596 LWLockAcquire(ProcArrayLock, LW_SHARED);
3597
3598 for (index = 0; index < arrayP->numProcs; index++)
3599 {
3600 int pgprocno = arrayP->pgprocnos[index];
3601 PGPROC *proc = &allProcs[pgprocno];
3602
3603 if (proc->pid == 0)
3604 continue; /* do not count prepared xacts */
3605 if (!proc->isRegularBackend)
3606 continue; /* count only regular backend processes */
3607 if (!OidIsValid(databaseid) ||
3608 proc->databaseId == databaseid)
3609 count++;
3610 }
3611
3612 LWLockRelease(ProcArrayLock);
3613
3614 return count;
3615}
3616
3617/*
3618 * CancelDBBackends --- cancel backends that are using specified database
3619 */
3620void
3621CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3622{
3623 ProcArrayStruct *arrayP = procArray;
3624 int index;
3625
3626 /* tell all backends to die */
3627 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3628
3629 for (index = 0; index < arrayP->numProcs; index++)
3630 {
3631 int pgprocno = arrayP->pgprocnos[index];
3632 PGPROC *proc = &allProcs[pgprocno];
3633
3634 if (databaseid == InvalidOid || proc->databaseId == databaseid)
3635 {
3636 VirtualTransactionId procvxid;
3637 pid_t pid;
3638
3639 GET_VXID_FROM_PGPROC(procvxid, *proc);
3640
3641 proc->recoveryConflictPending = conflictPending;
3642 pid = proc->pid;
3643 if (pid != 0)
3644 {
3645 /*
3646 * Kill the pid if it's still here. If not, that's what we
3647 * wanted so ignore any errors.
3648 */
3649 (void) SendProcSignal(pid, sigmode, procvxid.procNumber);
3650 }
3651 }
3652 }
3653
3654 LWLockRelease(ProcArrayLock);
3655}
3656
3657/*
3658 * CountUserBackends --- count backends that are used by specified user
3659 * (only regular backends, not any type of background worker)
3660 */
3661int
3663{
3664 ProcArrayStruct *arrayP = procArray;
3665 int count = 0;
3666 int index;
3667
3668 LWLockAcquire(ProcArrayLock, LW_SHARED);
3669
3670 for (index = 0; index < arrayP->numProcs; index++)
3671 {
3672 int pgprocno = arrayP->pgprocnos[index];
3673 PGPROC *proc = &allProcs[pgprocno];
3674
3675 if (proc->pid == 0)
3676 continue; /* do not count prepared xacts */
3677 if (!proc->isRegularBackend)
3678 continue; /* count only regular backend processes */
3679 if (proc->roleId == roleid)
3680 count++;
3681 }
3682
3683 LWLockRelease(ProcArrayLock);
3684
3685 return count;
3686}
3687
3688/*
3689 * CountOtherDBBackends -- check for other backends running in the given DB
3690 *
3691 * If there are other backends in the DB, we will wait a maximum of 5 seconds
3692 * for them to exit. Autovacuum backends are encouraged to exit early by
3693 * sending them SIGTERM, but normal user backends are just waited for.
3694 *
3695 * The current backend is always ignored; it is caller's responsibility to
3696 * check whether the current backend uses the given DB, if it's important.
3697 *
3698 * Returns true if there are (still) other backends in the DB, false if not.
3699 * Also, *nbackends and *nprepared are set to the number of other backends
3700 * and prepared transactions in the DB, respectively.
3701 *
3702 * This function is used to interlock DROP DATABASE and related commands
3703 * against there being any active backends in the target DB --- dropping the
3704 * DB while active backends remain would be a Bad Thing. Note that we cannot
3705 * detect here the possibility of a newly-started backend that is trying to
3706 * connect to the doomed database, so additional interlocking is needed during
3707 * backend startup. The caller should normally hold an exclusive lock on the
3708 * target DB before calling this, which is one reason we mustn't wait
3709 * indefinitely.
3710 */
3711bool
3712CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3713{
3714 ProcArrayStruct *arrayP = procArray;
3715
3716#define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3717 int autovac_pids[MAXAUTOVACPIDS];
3718 int tries;
3719
3720 /* 50 tries with 100ms sleep between tries makes 5 sec total wait */
3721 for (tries = 0; tries < 50; tries++)
3722 {
3723 int nautovacs = 0;
3724 bool found = false;
3725 int index;
3726
3728
3729 *nbackends = *nprepared = 0;
3730
3731 LWLockAcquire(ProcArrayLock, LW_SHARED);
3732
3733 for (index = 0; index < arrayP->numProcs; index++)
3734 {
3735 int pgprocno = arrayP->pgprocnos[index];
3736 PGPROC *proc = &allProcs[pgprocno];
3737 uint8 statusFlags = ProcGlobal->statusFlags[index];
3738
3739 if (proc->databaseId != databaseId)
3740 continue;
3741 if (proc == MyProc)
3742 continue;
3743
3744 found = true;
3745
3746 if (proc->pid == 0)
3747 (*nprepared)++;
3748 else
3749 {
3750 (*nbackends)++;
3751 if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3752 nautovacs < MAXAUTOVACPIDS)
3753 autovac_pids[nautovacs++] = proc->pid;
3754 }
3755 }
3756
3757 LWLockRelease(ProcArrayLock);
3758
3759 if (!found)
3760 return false; /* no conflicting backends, so done */
3761
3762 /*
3763 * Send SIGTERM to any conflicting autovacuums before sleeping. We
3764 * postpone this step until after the loop because we don't want to
3765 * hold ProcArrayLock while issuing kill(). We have no idea what might
3766 * block kill() inside the kernel...
3767 */
3768 for (index = 0; index < nautovacs; index++)
3769 (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3770
3771 /* sleep, then try again */
3772 pg_usleep(100 * 1000L); /* 100ms */
3773 }
3774
3775 return true; /* timed out, still conflicts */
3776}
3777
3778/*
3779 * Terminate existing connections to the specified database. This routine
3780 * is used by the DROP DATABASE command when user has asked to forcefully
3781 * drop the database.
3782 *
3783 * The current backend is always ignored; it is caller's responsibility to
3784 * check whether the current backend uses the given DB, if it's important.
3785 *
3786 * If the target database has a prepared transaction or permissions checks
3787 * fail for a connection, this fails without terminating anything.
3788 */
3789void
3791{
3792 ProcArrayStruct *arrayP = procArray;
3793 List *pids = NIL;
3794 int nprepared = 0;
3795 int i;
3796
3797 LWLockAcquire(ProcArrayLock, LW_SHARED);
3798
3799 for (i = 0; i < procArray->numProcs; i++)
3800 {
3801 int pgprocno = arrayP->pgprocnos[i];
3802 PGPROC *proc = &allProcs[pgprocno];
3803
3804 if (proc->databaseId != databaseId)
3805 continue;
3806 if (proc == MyProc)
3807 continue;
3808
3809 if (proc->pid != 0)
3810 pids = lappend_int(pids, proc->pid);
3811 else
3812 nprepared++;
3813 }
3814
3815 LWLockRelease(ProcArrayLock);
3816
3817 if (nprepared > 0)
3818 ereport(ERROR,
3819 (errcode(ERRCODE_OBJECT_IN_USE),
3820 errmsg("database \"%s\" is being used by prepared transactions",
3821 get_database_name(databaseId)),
3822 errdetail_plural("There is %d prepared transaction using the database.",
3823 "There are %d prepared transactions using the database.",
3824 nprepared,
3825 nprepared)));
3826
3827 if (pids)
3828 {
3829 ListCell *lc;
3830
3831 /*
3832 * Permissions checks relax the pg_terminate_backend checks in two
3833 * ways, both by omitting the !OidIsValid(proc->roleId) check:
3834 *
3835 * - Accept terminating autovacuum workers, since DROP DATABASE
3836 * without FORCE terminates them.
3837 *
3838 * - Accept terminating bgworkers. For bgworker authors, it's
3839 * convenient to be able to recommend FORCE if a worker is blocking
3840 * DROP DATABASE unexpectedly.
3841 *
3842 * Unlike pg_terminate_backend, we don't raise some warnings - like
3843 * "PID %d is not a PostgreSQL server process", because for us already
3844 * finished session is not a problem.
3845 */
3846 foreach(lc, pids)
3847 {
3848 int pid = lfirst_int(lc);
3849 PGPROC *proc = BackendPidGetProc(pid);
3850
3851 if (proc != NULL)
3852 {
3853 if (superuser_arg(proc->roleId) && !superuser())
3854 ereport(ERROR,
3855 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3856 errmsg("permission denied to terminate process"),
3857 errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3858 "SUPERUSER", "SUPERUSER")));
3859
3860 if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3861 !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3862 ereport(ERROR,
3863 (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3864 errmsg("permission denied to terminate process"),
3865 errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3866 "pg_signal_backend")));
3867 }
3868 }
3869
3870 /*
3871 * There's a race condition here: once we release the ProcArrayLock,
3872 * it's possible for the session to exit before we issue kill. That
3873 * race condition possibility seems too unlikely to worry about. See
3874 * pg_signal_backend.
3875 */
3876 foreach(lc, pids)
3877 {
3878 int pid = lfirst_int(lc);
3879 PGPROC *proc = BackendPidGetProc(pid);
3880
3881 if (proc != NULL)
3882 {
3883 /*
3884 * If we have setsid(), signal the backend's whole process
3885 * group
3886 */
3887#ifdef HAVE_SETSID
3888 (void) kill(-pid, SIGTERM);
3889#else
3890 (void) kill(pid, SIGTERM);
3891#endif
3892 }
3893 }
3894 }
3895}
3896
3897/*
3898 * ProcArraySetReplicationSlotXmin
3899 *
3900 * Install limits to future computations of the xmin horizon to prevent vacuum
3901 * and HOT pruning from removing affected rows still needed by clients with
3902 * replication slots.
3903 */
3904void
3906 bool already_locked)
3907{
3908 Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3909
3910 if (!already_locked)
3911 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3912
3915
3916 if (!already_locked)
3917 LWLockRelease(ProcArrayLock);
3918
3919 elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3920 xmin, catalog_xmin);
3921}
3922
3923/*
3924 * ProcArrayGetReplicationSlotXmin
3925 *
3926 * Return the current slot xmin limits. That's useful to be able to remove
3927 * data that's older than those limits.
3928 */
3929void
3931 TransactionId *catalog_xmin)
3932{
3933 LWLockAcquire(ProcArrayLock, LW_SHARED);
3934
3935 if (xmin != NULL)
3937
3938 if (catalog_xmin != NULL)
3940
3941 LWLockRelease(ProcArrayLock);
3942}
3943
3944/*
3945 * XidCacheRemoveRunningXids
3946 *
3947 * Remove a bunch of TransactionIds from the list of known-running
3948 * subtransactions for my backend. Both the specified xid and those in
3949 * the xids[] array (of length nxids) are removed from the subxids cache.
3950 * latestXid must be the latest XID among the group.
3951 */
3952void
3954 int nxids, const TransactionId *xids,
3955 TransactionId latestXid)
3956{
3957 int i,
3958 j;
3959 XidCacheStatus *mysubxidstat;
3960
3962
3963 /*
3964 * We must hold ProcArrayLock exclusively in order to remove transactions
3965 * from the PGPROC array. (See src/backend/access/transam/README.) It's
3966 * possible this could be relaxed since we know this routine is only used
3967 * to abort subtransactions, but pending closer analysis we'd best be
3968 * conservative.
3969 *
3970 * Note that we do not have to be careful about memory ordering of our own
3971 * reads wrt. GetNewTransactionId() here - only this process can modify
3972 * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
3973 * careful about our own writes being well ordered.
3974 */
3975 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3976
3977 mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
3978
3979 /*
3980 * Under normal circumstances xid and xids[] will be in increasing order,
3981 * as will be the entries in subxids. Scan backwards to avoid O(N^2)
3982 * behavior when removing a lot of xids.
3983 */
3984 for (i = nxids - 1; i >= 0; i--)
3985 {
3986 TransactionId anxid = xids[i];
3987
3988 for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
3989 {
3990 if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
3991 {
3994 mysubxidstat->count--;
3996 break;
3997 }
3998 }
3999
4000 /*
4001 * Ordinarily we should have found it, unless the cache has
4002 * overflowed. However it's also possible for this routine to be
4003 * invoked multiple times for the same subtransaction, in case of an
4004 * error during AbortSubTransaction. So instead of Assert, emit a
4005 * debug warning.
4006 */
4007 if (j < 0 && !MyProc->subxidStatus.overflowed)
4008 elog(WARNING, "did not find subXID %u in MyProc", anxid);
4009 }
4010
4011 for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4012 {
4014 {
4017 mysubxidstat->count--;
4019 break;
4020 }
4021 }
4022 /* Ordinarily we should have found it, unless the cache has overflowed */
4023 if (j < 0 && !MyProc->subxidStatus.overflowed)
4024 elog(WARNING, "did not find subXID %u in MyProc", xid);
4025
4026 /* Also advance global latestCompletedXid while holding the lock */
4027 MaintainLatestCompletedXid(latestXid);
4028
4029 /* ... and xactCompletionCount */
4031
4032 LWLockRelease(ProcArrayLock);
4033}
4034
4035#ifdef XIDCACHE_DEBUG
4036
4037/*
4038 * Print stats about effectiveness of XID cache
4039 */
4040static void
4041DisplayXidCache(void)
4042{
4043 fprintf(stderr,
4044 "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4045 xc_by_recent_xmin,
4046 xc_by_known_xact,
4047 xc_by_my_xact,
4048 xc_by_latest_xid,
4049 xc_by_main_xid,
4050 xc_by_child_xid,
4051 xc_by_known_assigned,
4052 xc_no_overflow,
4053 xc_slow_answer);
4054}
4055#endif /* XIDCACHE_DEBUG */
4056
4057/*
4058 * If rel != NULL, return test state appropriate for relation, otherwise
4059 * return state usable for all relations. The latter may consider XIDs as
4060 * not-yet-visible-to-everyone that a state for a specific relation would
4061 * already consider visible-to-everyone.
4062 *
4063 * This needs to be called while a snapshot is active or registered, otherwise
4064 * there are wraparound and other dangers.
4065 *
4066 * See comment for GlobalVisState for details.
4067 */
4070{
4071 GlobalVisState *state = NULL;
4072
4073 /* XXX: we should assert that a snapshot is pushed or registered */
4075
4076 switch (GlobalVisHorizonKindForRel(rel))
4077 {
4078 case VISHORIZON_SHARED:
4080 break;
4081 case VISHORIZON_CATALOG:
4083 break;
4084 case VISHORIZON_DATA:
4086 break;
4087 case VISHORIZON_TEMP:
4089 break;
4090 }
4091
4092 Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4093 FullTransactionIdIsValid(state->maybe_needed));
4094
4095 return state;
4096}
4097
4098/*
4099 * Return true if it's worth updating the accurate maybe_needed boundary.
4100 *
4101 * As it is somewhat expensive to determine xmin horizons, we don't want to
4102 * repeatedly do so when there is a low likelihood of it being beneficial.
4103 *
4104 * The current heuristic is that we update only if RecentXmin has changed
4105 * since the last update. If the oldest currently running transaction has not
4106 * finished, it is unlikely that recomputing the horizon would be useful.
4107 */
4108static bool
4110{
4111 /* hasn't been updated yet */
4113 return true;
4114
4115 /*
4116 * If the maybe_needed/definitely_needed boundaries are the same, it's
4117 * unlikely to be beneficial to refresh boundaries.
4118 */
4119 if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4120 state->definitely_needed))
4121 return false;
4122
4123 /* does the last snapshot built have a different xmin? */
4125}
4126
4127static void
4129{
4132 horizons->shared_oldest_nonremovable);
4135 horizons->catalog_oldest_nonremovable);
4138 horizons->data_oldest_nonremovable);
4141 horizons->temp_oldest_nonremovable);
4142
4143 /*
4144 * In longer running transactions it's possible that transactions we
4145 * previously needed to treat as running aren't around anymore. So update
4146 * definitely_needed to not be earlier than maybe_needed.
4147 */
4158
4160}
4161
4162/*
4163 * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4164 * using ComputeXidHorizons().
4165 */
4166static void
4168{
4169 ComputeXidHorizonsResult horizons;
4170
4171 /* updates the horizons as a side-effect */
4172 ComputeXidHorizons(&horizons);
4173}
4174
4175/*
4176 * Return true if no snapshot still considers fxid to be running.
4177 *
4178 * The state passed needs to have been initialized for the relation fxid is
4179 * from (NULL is also OK), otherwise the result may not be correct.
4180 *
4181 * See comment for GlobalVisState for details.
4182 */
4183bool
4185 FullTransactionId fxid)
4186{
4187 /*
4188 * If fxid is older than maybe_needed bound, it definitely is visible to
4189 * everyone.
4190 */
4191 if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4192 return true;
4193
4194 /*
4195 * If fxid is >= definitely_needed bound, it is very likely to still be
4196 * considered running.
4197 */
4198 if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4199 return false;
4200
4201 /*
4202 * fxid is between maybe_needed and definitely_needed, i.e. there might or
4203 * might not exist a snapshot considering fxid running. If it makes sense,
4204 * update boundaries and recheck.
4205 */
4207 {
4209
4210 Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4211
4212 return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4213 }
4214 else
4215 return false;
4216}
4217
4218/*
4219 * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4220 *
4221 * It is crucial that this only gets called for xids from a source that
4222 * protects against xid wraparounds (e.g. from a table and thus protected by
4223 * relfrozenxid).
4224 */
4225bool
4227{
4228 FullTransactionId fxid;
4229
4230 /*
4231 * Convert 32 bit argument to FullTransactionId. We can do so safely
4232 * because we know the xid has to, at the very least, be between
4233 * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4234 * lock to determine either, we can just compare with
4235 * state->definitely_needed, which was based on those value at the time
4236 * the current snapshot was built.
4237 */
4238 fxid = FullXidRelativeTo(state->definitely_needed, xid);
4239
4241}
4242
4243/*
4244 * Convenience wrapper around GlobalVisTestFor() and
4245 * GlobalVisTestIsRemovableFullXid(), see their comments.
4246 */
4247bool
4249{
4251
4252 state = GlobalVisTestFor(rel);
4253
4255}
4256
4257/*
4258 * Convenience wrapper around GlobalVisTestFor() and
4259 * GlobalVisTestIsRemovableXid(), see their comments.
4260 */
4261bool
4263{
4265
4266 state = GlobalVisTestFor(rel);
4267
4269}
4270
4271/*
4272 * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4273 * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4274 *
4275 * Be very careful about when to use this function. It can only safely be used
4276 * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4277 * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4278 * held by the backend and xid is from a table (where vacuum/freezing ensures
4279 * the xid has to be within that range), or if xid is from the procarray and
4280 * prevents xid wraparound that way.
4281 */
4282static inline FullTransactionId
4284{
4286
4288 Assert(TransactionIdIsValid(rel_xid));
4289
4290 /* not guaranteed to find issues, but likely to catch mistakes */
4292
4294 + (int32) (xid - rel_xid));
4295}
4296
4297
4298/* ----------------------------------------------
4299 * KnownAssignedTransactionIds sub-module
4300 * ----------------------------------------------
4301 */
4302
4303/*
4304 * In Hot Standby mode, we maintain a list of transactions that are (or were)
4305 * running on the primary at the current point in WAL. These XIDs must be
4306 * treated as running by standby transactions, even though they are not in
4307 * the standby server's PGPROC array.
4308 *
4309 * We record all XIDs that we know have been assigned. That includes all the
4310 * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4311 * been assigned. We can deduce the existence of unobserved XIDs because we
4312 * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4313 * list expands as new XIDs are observed or inferred, and contracts when
4314 * transaction completion records arrive.
4315 *
4316 * During hot standby we do not fret too much about the distinction between
4317 * top-level XIDs and subtransaction XIDs. We store both together in the
4318 * KnownAssignedXids list. In backends, this is copied into snapshots in
4319 * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4320 * doesn't care about the distinction either. Subtransaction XIDs are
4321 * effectively treated as top-level XIDs and in the typical case pg_subtrans
4322 * links are *not* maintained (which does not affect visibility).
4323 *
4324 * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4325 * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4326 * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4327 * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4328 * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4329 * and then remove them from KnownAssignedXids. This prevents overflow of
4330 * KnownAssignedXids and snapshots, at the cost that status checks for these
4331 * subXIDs will take a slower path through TransactionIdIsInProgress().
4332 * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4333 * though it should be complete for top-level XIDs; this is the same situation
4334 * that holds with respect to the PGPROC entries in normal running.
4335 *
4336 * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4337 * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4338 * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4339 * As long as that is within the range of interesting XIDs, we have to assume
4340 * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4341 * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4342 * subXID arrives - that is not an error.)
4343 *
4344 * Should a backend on primary somehow disappear before it can write an abort
4345 * record, then we just leave those XIDs in KnownAssignedXids. They actually
4346 * aborted but we think they were running; the distinction is irrelevant
4347 * because either way any changes done by the transaction are not visible to
4348 * backends in the standby. We prune KnownAssignedXids when
4349 * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4350 * array due to such dead XIDs.
4351 */
4352
4353/*
4354 * RecordKnownAssignedTransactionIds
4355 * Record the given XID in KnownAssignedXids, as well as any preceding
4356 * unobserved XIDs.
4357 *
4358 * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4359 * associated with a transaction. Must be called for each record after we
4360 * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4361 *
4362 * Called during recovery in analogy with and in place of GetNewTransactionId()
4363 */
4364void
4366{
4370
4371 elog(DEBUG4, "record known xact %u latestObservedXid %u",
4372 xid, latestObservedXid);
4373
4374 /*
4375 * When a newly observed xid arrives, it is frequently the case that it is
4376 * *not* the next xid in sequence. When this occurs, we must treat the
4377 * intervening xids as running also.
4378 */
4380 {
4381 TransactionId next_expected_xid;
4382
4383 /*
4384 * Extend subtrans like we do in GetNewTransactionId() during normal
4385 * operation using individual extend steps. Note that we do not need
4386 * to extend clog since its extensions are WAL logged.
4387 *
4388 * This part has to be done regardless of standbyState since we
4389 * immediately start assigning subtransactions to their toplevel
4390 * transactions.
4391 */
4392 next_expected_xid = latestObservedXid;
4393 while (TransactionIdPrecedes(next_expected_xid, xid))
4394 {
4395 TransactionIdAdvance(next_expected_xid);
4396 ExtendSUBTRANS(next_expected_xid);
4397 }
4398 Assert(next_expected_xid == xid);
4399
4400 /*
4401 * If the KnownAssignedXids machinery isn't up yet, there's nothing
4402 * more to do since we don't track assigned xids yet.
4403 */
4405 {
4406 latestObservedXid = xid;
4407 return;
4408 }
4409
4410 /*
4411 * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4412 */
4413 next_expected_xid = latestObservedXid;
4414 TransactionIdAdvance(next_expected_xid);
4415 KnownAssignedXidsAdd(next_expected_xid, xid, false);
4416
4417 /*
4418 * Now we can advance latestObservedXid
4419 */
4420 latestObservedXid = xid;
4421
4422 /* TransamVariables->nextXid must be beyond any observed xid */
4424 }
4425}
4426
4427/*
4428 * ExpireTreeKnownAssignedTransactionIds
4429 * Remove the given XIDs from KnownAssignedXids.
4430 *
4431 * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4432 */
4433void
4435 TransactionId *subxids, TransactionId max_xid)
4436{
4438
4439 /*
4440 * Uses same locking as transaction commit
4441 */
4442 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4443
4444 KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4445
4446 /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4448
4449 /* ... and xactCompletionCount */
4451
4452 LWLockRelease(ProcArrayLock);
4453}
4454
4455/*
4456 * ExpireAllKnownAssignedTransactionIds
4457 * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4458 */
4459void
4461{
4462 FullTransactionId latestXid;
4463
4464 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4466
4467 /* Reset latestCompletedXid to nextXid - 1 */
4469 latestXid = TransamVariables->nextXid;
4470 FullTransactionIdRetreat(&latestXid);
4472
4473 /*
4474 * Any transactions that were in-progress were effectively aborted, so
4475 * advance xactCompletionCount.
4476 */
4478
4479 /*
4480 * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4481 * the call of this function. But do this for unification with what
4482 * ExpireOldKnownAssignedTransactionIds() do.
4483 */
4485 LWLockRelease(ProcArrayLock);
4486}
4487
4488/*
4489 * ExpireOldKnownAssignedTransactionIds
4490 * Remove KnownAssignedXids entries preceding the given XID and
4491 * potentially reset lastOverflowedXid.
4492 */
4493void
4495{
4496 TransactionId latestXid;
4497
4498 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4499
4500 /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4501 latestXid = xid;
4502 TransactionIdRetreat(latestXid);
4504
4505 /* ... and xactCompletionCount */
4507
4508 /*
4509 * Reset lastOverflowedXid if we know all transactions that have been
4510 * possibly running are being gone. Not doing so could cause an incorrect
4511 * lastOverflowedXid value, which makes extra snapshots be marked as
4512 * suboverflowed.
4513 */
4517 LWLockRelease(ProcArrayLock);
4518}
4519
4520/*
4521 * KnownAssignedTransactionIdsIdleMaintenance
4522 * Opportunistically do maintenance work when the startup process
4523 * is about to go idle.
4524 */
4525void
4527{
4529}
4530
4531
4532/*
4533 * Private module functions to manipulate KnownAssignedXids
4534 *
4535 * There are 5 main uses of the KnownAssignedXids data structure:
4536 *
4537 * * backends taking snapshots - all valid XIDs need to be copied out
4538 * * backends seeking to determine presence of a specific XID
4539 * * startup process adding new known-assigned XIDs
4540 * * startup process removing specific XIDs as transactions end
4541 * * startup process pruning array when special WAL records arrive
4542 *
4543 * This data structure is known to be a hot spot during Hot Standby, so we
4544 * go to some lengths to make these operations as efficient and as concurrent
4545 * as possible.
4546 *
4547 * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4548 * order, to be exact --- to allow binary search for specific XIDs. Note:
4549 * in general TransactionIdPrecedes would not provide a total order, but
4550 * we know that the entries present at any instant should not extend across
4551 * a large enough fraction of XID space to wrap around (the primary would
4552 * shut down for fear of XID wrap long before that happens). So it's OK to
4553 * use TransactionIdPrecedes as a binary-search comparator.
4554 *
4555 * It's cheap to maintain the sortedness during insertions, since new known
4556 * XIDs are always reported in XID order; we just append them at the right.
4557 *
4558 * To keep individual deletions cheap, we need to allow gaps in the array.
4559 * This is implemented by marking array elements as valid or invalid using
4560 * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4561 * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4562 * XID entry itself. This preserves the property that the XID entries are
4563 * sorted, so we can do binary searches easily. Periodically we compress
4564 * out the unused entries; that's much cheaper than having to compress the
4565 * array immediately on every deletion.
4566 *
4567 * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4568 * are those with indexes tail <= i < head; items outside this subscript range
4569 * have unspecified contents. When head reaches the end of the array, we
4570 * force compression of unused entries rather than wrapping around, since
4571 * allowing wraparound would greatly complicate the search logic. We maintain
4572 * an explicit tail pointer so that pruning of old XIDs can be done without
4573 * immediately moving the array contents. In most cases only a small fraction
4574 * of the array contains valid entries at any instant.
4575 *
4576 * Although only the startup process can ever change the KnownAssignedXids
4577 * data structure, we still need interlocking so that standby backends will
4578 * not observe invalid intermediate states. The convention is that backends
4579 * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4580 * the array, the startup process must hold ProcArrayLock exclusively, for
4581 * the usual transactional reasons (compare commit/abort of a transaction
4582 * during normal running). Compressing unused entries out of the array
4583 * likewise requires exclusive lock. To add XIDs to the array, we just insert
4584 * them into slots to the right of the head pointer and then advance the head
4585 * pointer. This doesn't require any lock at all, but on machines with weak
4586 * memory ordering, we need to be careful that other processors see the array
4587 * element changes before they see the head pointer change. We handle this by
4588 * using memory barriers when reading or writing the head/tail pointers (unless
4589 * the caller holds ProcArrayLock exclusively).
4590 *
4591 * Algorithmic analysis:
4592 *
4593 * If we have a maximum of M slots, with N XIDs currently spread across
4594 * S elements then we have N <= S <= M always.
4595 *
4596 * * Adding a new XID is O(1) and needs no lock (unless compression must
4597 * happen)
4598 * * Compressing the array is O(S) and requires exclusive lock
4599 * * Removing an XID is O(logS) and requires exclusive lock
4600 * * Taking a snapshot is O(S) and requires shared lock
4601 * * Checking for an XID is O(logS) and requires shared lock
4602 *
4603 * In comparison, using a hash table for KnownAssignedXids would mean that
4604 * taking snapshots would be O(M). If we can maintain S << M then the
4605 * sorted array technique will deliver significantly faster snapshots.
4606 * If we try to keep S too small then we will spend too much time compressing,
4607 * so there is an optimal point for any workload mix. We use a heuristic to
4608 * decide when to compress the array, though trimming also helps reduce
4609 * frequency of compressing. The heuristic requires us to track the number of
4610 * currently valid XIDs in the array (N). Except in special cases, we'll
4611 * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4612 * taking a snapshot to be O(N), which it would have to be anyway.
4613 */
4614
4615
4616/*
4617 * Compress KnownAssignedXids by shifting valid data down to the start of the
4618 * array, removing any gaps.
4619 *
4620 * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4621 * we do it only if a heuristic indicates it's a good time to do it.
4622 *
4623 * Compression requires holding ProcArrayLock in exclusive mode.
4624 * Caller must pass haveLock = true if it already holds the lock.
4625 */
4626static void
4628{
4629 ProcArrayStruct *pArray = procArray;
4630 int head,
4631 tail,
4632 nelements;
4633 int compress_index;
4634 int i;
4635
4636 /* Counters for compression heuristics */
4637 static unsigned int transactionEndsCounter;
4638 static TimestampTz lastCompressTs;
4639
4640 /* Tuning constants */
4641#define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4642#define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4643
4644 /*
4645 * Since only the startup process modifies the head/tail pointers, we
4646 * don't need a lock to read them here.
4647 */
4648 head = pArray->headKnownAssignedXids;
4649 tail = pArray->tailKnownAssignedXids;
4650 nelements = head - tail;
4651
4652 /*
4653 * If we can choose whether to compress, use a heuristic to avoid
4654 * compressing too often or not often enough. "Compress" here simply
4655 * means moving the values to the beginning of the array, so it is not as
4656 * complex or costly as typical data compression algorithms.
4657 */
4658 if (nelements == pArray->numKnownAssignedXids)
4659 {
4660 /*
4661 * When there are no gaps between head and tail, don't bother to
4662 * compress, except in the KAX_NO_SPACE case where we must compress to
4663 * create some space after the head.
4664 */
4665 if (reason != KAX_NO_SPACE)
4666 return;
4667 }
4668 else if (reason == KAX_TRANSACTION_END)
4669 {
4670 /*
4671 * Consider compressing only once every so many commits. Frequency
4672 * determined by benchmarks.
4673 */
4674 if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4675 return;
4676
4677 /*
4678 * Furthermore, compress only if the used part of the array is less
4679 * than 50% full (see comments above).
4680 */
4681 if (nelements < 2 * pArray->numKnownAssignedXids)
4682 return;
4683 }
4684 else if (reason == KAX_STARTUP_PROCESS_IDLE)
4685 {
4686 /*
4687 * We're about to go idle for lack of new WAL, so we might as well
4688 * compress. But not too often, to avoid ProcArray lock contention
4689 * with readers.
4690 */
4691 if (lastCompressTs != 0)
4692 {
4693 TimestampTz compress_after;
4694
4695 compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4697 if (GetCurrentTimestamp() < compress_after)
4698 return;
4699 }
4700 }
4701
4702 /* Need to compress, so get the lock if we don't have it. */
4703 if (!haveLock)
4704 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4705
4706 /*
4707 * We compress the array by reading the valid values from tail to head,
4708 * re-aligning data to 0th element.
4709 */
4710 compress_index = 0;
4711 for (i = tail; i < head; i++)
4712 {
4714 {
4715 KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4716 KnownAssignedXidsValid[compress_index] = true;
4717 compress_index++;
4718 }
4719 }
4720 Assert(compress_index == pArray->numKnownAssignedXids);
4721
4722 pArray->tailKnownAssignedXids = 0;
4723 pArray->headKnownAssignedXids = compress_index;
4724
4725 if (!haveLock)
4726 LWLockRelease(ProcArrayLock);
4727
4728 /* Update timestamp for maintenance. No need to hold lock for this. */
4729 lastCompressTs = GetCurrentTimestamp();
4730}
4731
4732/*
4733 * Add xids into KnownAssignedXids at the head of the array.
4734 *
4735 * xids from from_xid to to_xid, inclusive, are added to the array.
4736 *
4737 * If exclusive_lock is true then caller already holds ProcArrayLock in
4738 * exclusive mode, so we need no extra locking here. Else caller holds no
4739 * lock, so we need to be sure we maintain sufficient interlocks against
4740 * concurrent readers. (Only the startup process ever calls this, so no need
4741 * to worry about concurrent writers.)
4742 */
4743static void
4745 bool exclusive_lock)
4746{
4747 ProcArrayStruct *pArray = procArray;
4748 TransactionId next_xid;
4749 int head,
4750 tail;
4751 int nxids;
4752 int i;
4753
4754 Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4755
4756 /*
4757 * Calculate how many array slots we'll need. Normally this is cheap; in
4758 * the unusual case where the XIDs cross the wrap point, we do it the hard
4759 * way.
4760 */
4761 if (to_xid >= from_xid)
4762 nxids = to_xid - from_xid + 1;
4763 else
4764 {
4765 nxids = 1;
4766 next_xid = from_xid;
4767 while (TransactionIdPrecedes(next_xid, to_xid))
4768 {
4769 nxids++;
4770 TransactionIdAdvance(next_xid);
4771 }
4772 }
4773
4774 /*
4775 * Since only the startup process modifies the head/tail pointers, we
4776 * don't need a lock to read them here.
4777 */
4778 head = pArray->headKnownAssignedXids;
4779 tail = pArray->tailKnownAssignedXids;
4780
4781 Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4782 Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4783
4784 /*
4785 * Verify that insertions occur in TransactionId sequence. Note that even
4786 * if the last existing element is marked invalid, it must still have a
4787 * correctly sequenced XID value.
4788 */
4789 if (head > tail &&
4791 {
4793 elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4794 }
4795
4796 /*
4797 * If our xids won't fit in the remaining space, compress out free space
4798 */
4799 if (head + nxids > pArray->maxKnownAssignedXids)
4800 {
4801 KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4802
4803 head = pArray->headKnownAssignedXids;
4804 /* note: we no longer care about the tail pointer */
4805
4806 /*
4807 * If it still won't fit then we're out of memory
4808 */
4809 if (head + nxids > pArray->maxKnownAssignedXids)
4810 elog(ERROR, "too many KnownAssignedXids");
4811 }
4812
4813 /* Now we can insert the xids into the space starting at head */
4814 next_xid = from_xid;
4815 for (i = 0; i < nxids; i++)
4816 {
4817 KnownAssignedXids[head] = next_xid;
4818 KnownAssignedXidsValid[head] = true;
4819 TransactionIdAdvance(next_xid);
4820 head++;
4821 }
4822
4823 /* Adjust count of number of valid entries */
4824 pArray->numKnownAssignedXids += nxids;
4825
4826 /*
4827 * Now update the head pointer. We use a write barrier to ensure that
4828 * other processors see the above array updates before they see the head
4829 * pointer change. The barrier isn't required if we're holding
4830 * ProcArrayLock exclusively.
4831 */
4832 if (!exclusive_lock)
4834
4835 pArray->headKnownAssignedXids = head;
4836}
4837
4838/*
4839 * KnownAssignedXidsSearch
4840 *
4841 * Searches KnownAssignedXids for a specific xid and optionally removes it.
4842 * Returns true if it was found, false if not.
4843 *
4844 * Caller must hold ProcArrayLock in shared or exclusive mode.
4845 * Exclusive lock must be held for remove = true.
4846 */
4847static bool
4849{
4850 ProcArrayStruct *pArray = procArray;
4851 int first,
4852 last;
4853 int head;
4854 int tail;
4855 int result_index = -1;
4856
4857 tail = pArray->tailKnownAssignedXids;
4858 head = pArray->headKnownAssignedXids;
4859
4860 /*
4861 * Only the startup process removes entries, so we don't need the read
4862 * barrier in that case.
4863 */
4864 if (!remove)
4865 pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4866
4867 /*
4868 * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4869 * array here, since even invalid entries will contain sorted XIDs.
4870 */
4871 first = tail;
4872 last = head - 1;
4873 while (first <= last)
4874 {
4875 int mid_index;
4876 TransactionId mid_xid;
4877
4878 mid_index = (first + last) / 2;
4879 mid_xid = KnownAssignedXids[mid_index];
4880
4881 if (xid == mid_xid)
4882 {
4883 result_index = mid_index;
4884 break;
4885 }
4886 else if (TransactionIdPrecedes(xid, mid_xid))
4887 last = mid_index - 1;
4888 else
4889 first = mid_index + 1;
4890 }
4891
4892 if (result_index < 0)
4893 return false; /* not in array */
4894
4895 if (!KnownAssignedXidsValid[result_index])
4896 return false; /* in array, but invalid */
4897
4898 if (remove)
4899 {
4900 KnownAssignedXidsValid[result_index] = false;
4901
4902 pArray->numKnownAssignedXids--;
4903 Assert(pArray->numKnownAssignedXids >= 0);
4904
4905 /*
4906 * If we're removing the tail element then advance tail pointer over
4907 * any invalid elements. This will speed future searches.
4908 */
4909 if (result_index == tail)
4910 {
4911 tail++;
4912 while (tail < head && !KnownAssignedXidsValid[tail])
4913 tail++;
4914 if (tail >= head)
4915 {
4916 /* Array is empty, so we can reset both pointers */
4917 pArray->headKnownAssignedXids = 0;
4918 pArray->tailKnownAssignedXids = 0;
4919 }
4920 else
4921 {
4922 pArray->tailKnownAssignedXids = tail;
4923 }
4924 }
4925 }
4926
4927 return true;
4928}
4929
4930/*
4931 * Is the specified XID present in KnownAssignedXids[]?
4932 *
4933 * Caller must hold ProcArrayLock in shared or exclusive mode.
4934 */
4935static bool
4937{
4939
4940 return KnownAssignedXidsSearch(xid, false);
4941}
4942
4943/*
4944 * Remove the specified XID from KnownAssignedXids[].
4945 *
4946 * Caller must hold ProcArrayLock in exclusive mode.
4947 */
4948static void
4950{
4952
4953 elog(DEBUG4, "remove KnownAssignedXid %u", xid);
4954
4955 /*
4956 * Note: we cannot consider it an error to remove an XID that's not
4957 * present. We intentionally remove subxact IDs while processing
4958 * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
4959 * removed again when the top-level xact commits or aborts.
4960 *
4961 * It might be possible to track such XIDs to distinguish this case from
4962 * actual errors, but it would be complicated and probably not worth it.
4963 * So, just ignore the search result.
4964 */
4965 (void) KnownAssignedXidsSearch(xid, true);
4966}
4967
4968/*
4969 * KnownAssignedXidsRemoveTree
4970 * Remove xid (if it's not InvalidTransactionId) and all the subxids.
4971 *
4972 * Caller must hold ProcArrayLock in exclusive mode.
4973 */
4974static void
4976 TransactionId *subxids)
4977{
4978 int i;
4979
4980 if (TransactionIdIsValid(xid))
4982
4983 for (i = 0; i < nsubxids; i++)
4984 KnownAssignedXidsRemove(subxids[i]);
4985
4986 /* Opportunistically compress the array */
4988}
4989
4990/*
4991 * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
4992 * then clear the whole table.
4993 *
4994 * Caller must hold ProcArrayLock in exclusive mode.
4995 */
4996static void
4998{
4999 ProcArrayStruct *pArray = procArray;
5000 int count = 0;
5001 int head,
5002 tail,
5003 i;
5004
5005 if (!TransactionIdIsValid(removeXid))
5006 {
5007 elog(DEBUG4, "removing all KnownAssignedXids");
5008 pArray->numKnownAssignedXids = 0;
5009 pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5010 return;
5011 }
5012
5013 elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5014
5015 /*
5016 * Mark entries invalid starting at the tail. Since array is sorted, we
5017 * can stop as soon as we reach an entry >= removeXid.
5018 */
5019 tail = pArray->tailKnownAssignedXids;
5020 head = pArray->headKnownAssignedXids;
5021
5022 for (i = tail; i < head; i++)
5023 {
5025 {
5026 TransactionId knownXid = KnownAssignedXids[i];
5027
5028 if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5029 break;
5030
5031 if (!StandbyTransactionIdIsPrepared(knownXid))
5032 {
5033 KnownAssignedXidsValid[i] = false;
5034 count++;
5035 }
5036 }
5037 }
5038
5039 pArray->numKnownAssignedXids -= count;
5040 Assert(pArray->numKnownAssignedXids >= 0);
5041
5042 /*
5043 * Advance the tail pointer if we've marked the tail item invalid.
5044 */
5045 for (i = tail; i < head; i++)
5046 {
5048 break;
5049 }
5050 if (i >= head)
5051 {
5052 /* Array is empty, so we can reset both pointers */
5053 pArray->headKnownAssignedXids = 0;
5054 pArray->tailKnownAssignedXids = 0;
5055 }
5056 else
5057 {
5058 pArray->tailKnownAssignedXids = i;
5059 }
5060
5061 /* Opportunistically compress the array */
5063}
5064
5065/*
5066 * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5067 * We filter out anything >= xmax.
5068 *
5069 * Returns the number of XIDs stored into xarray[]. Caller is responsible
5070 * that array is large enough.
5071 *
5072 * Caller must hold ProcArrayLock in (at least) shared mode.
5073 */
5074static int
5076{
5078
5079 return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5080}
5081
5082/*
5083 * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5084 * we reduce *xmin to the lowest xid value seen if not already lower.
5085 *
5086 * Caller must hold ProcArrayLock in (at least) shared mode.
5087 */
5088static int
5090 TransactionId xmax)
5091{
5092 int count = 0;
5093 int head,
5094 tail;
5095 int i;
5096
5097 /*
5098 * Fetch head just once, since it may change while we loop. We can stop
5099 * once we reach the initially seen head, since we are certain that an xid
5100 * cannot enter and then leave the array while we hold ProcArrayLock. We
5101 * might miss newly-added xids, but they should be >= xmax so irrelevant
5102 * anyway.
5103 */
5106
5107 pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5108
5109 for (i = tail; i < head; i++)
5110 {
5111 /* Skip any gaps in the array */
5113 {
5114 TransactionId knownXid = KnownAssignedXids[i];
5115
5116 /*
5117 * Update xmin if required. Only the first XID need be checked,
5118 * since the array is sorted.
5119 */
5120 if (count == 0 &&
5121 TransactionIdPrecedes(knownXid, *xmin))
5122 *xmin = knownXid;
5123
5124 /*
5125 * Filter out anything >= xmax, again relying on sorted property
5126 * of array.
5127 */
5128 if (TransactionIdIsValid(xmax) &&
5129 TransactionIdFollowsOrEquals(knownXid, xmax))
5130 break;
5131
5132 /* Add knownXid into output array */
5133 xarray[count++] = knownXid;
5134 }
5135 }
5136
5137 return count;
5138}
5139
5140/*
5141 * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5142 * if nothing there.
5143 */
5144static TransactionId
5146{
5147 int head,
5148 tail;
5149 int i;
5150
5151 /*
5152 * Fetch head just once, since it may change while we loop.
5153 */
5156
5157 pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5158
5159 for (i = tail; i < head; i++)
5160 {
5161 /* Skip any gaps in the array */
5163 return KnownAssignedXids[i];
5164 }
5165
5166 return InvalidTransactionId;
5167}
5168
5169/*
5170 * Display KnownAssignedXids to provide debug trail
5171 *
5172 * Currently this is only called within startup process, so we need no
5173 * special locking.
5174 *
5175 * Note this is pretty expensive, and much of the expense will be incurred
5176 * even if the elog message will get discarded. It's not currently called
5177 * in any performance-critical places, however, so no need to be tenser.
5178 */
5179static void
5181{
5182 ProcArrayStruct *pArray = procArray;
5184 int head,
5185 tail,
5186 i;
5187 int nxids = 0;
5188
5189 tail = pArray->tailKnownAssignedXids;
5190 head = pArray->headKnownAssignedXids;
5191
5193
5194 for (i = tail; i < head; i++)
5195 {
5197 {
5198 nxids++;
5199 appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5200 }
5201 }
5202
5203 elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5204 nxids,
5205 pArray->numKnownAssignedXids,
5206 pArray->tailKnownAssignedXids,
5207 pArray->headKnownAssignedXids,
5208 buf.data);
5209
5210 pfree(buf.data);
5211}
5212
5213/*
5214 * KnownAssignedXidsReset
5215 * Resets KnownAssignedXids to be empty
5216 */
5217static void
5219{
5220 ProcArrayStruct *pArray = procArray;
5221
5222 LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5223
5224 pArray->numKnownAssignedXids = 0;
5225 pArray->tailKnownAssignedXids = 0;
5226 pArray->headKnownAssignedXids = 0;
5227
5228 LWLockRelease(ProcArrayLock);
5229}
bool has_privs_of_role(Oid member, Oid role)
Definition: acl.c:5284
static bool pg_atomic_compare_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 *expected, uint32 newval)
Definition: atomics.h:347
#define pg_read_barrier()
Definition: atomics.h:154
#define pg_write_barrier()
Definition: atomics.h:155
static void pg_atomic_write_u32(volatile pg_atomic_uint32 *ptr, uint32 val)
Definition: atomics.h:274
static uint32 pg_atomic_read_u32(volatile pg_atomic_uint32 *ptr)
Definition: atomics.h:237
static uint32 pg_atomic_exchange_u32(volatile pg_atomic_uint32 *ptr, uint32 newval)
Definition: atomics.h:328
TimestampTz GetCurrentTimestamp(void)
Definition: timestamp.c:1645
#define likely(x)
Definition: c.h:402
uint8_t uint8
Definition: c.h:537
#define FLEXIBLE_ARRAY_MEMBER
Definition: c.h:471
int8_t int8
Definition: c.h:533
int32_t int32
Definition: c.h:535
uint64_t uint64
Definition: c.h:540
#define unlikely(x)
Definition: c.h:403
uint32_t uint32
Definition: c.h:539
uint32 TransactionId
Definition: c.h:658
#define OidIsValid(objectId)
Definition: c.h:775
size_t Size
Definition: c.h:611
bool IsCatalogRelation(Relation relation)
Definition: catalog.c:104
#define fprintf(file, fmt, msg)
Definition: cubescan.l:21
int64 TimestampTz
Definition: timestamp.h:39
int errdetail(const char *fmt,...)
Definition: elog.c:1207
int errdetail_plural(const char *fmt_singular, const char *fmt_plural, unsigned long n,...)
Definition: elog.c:1299
int errcode(int sqlerrcode)
Definition: elog.c:854
int errmsg(const char *fmt,...)
Definition: elog.c:1071
#define LOG
Definition: elog.h:31
#define DEBUG3
Definition: elog.h:28
#define FATAL
Definition: elog.h:41
#define WARNING
Definition: elog.h:36
#define DEBUG1
Definition: elog.h:30
#define ERROR
Definition: elog.h:39
#define elog(elevel,...)
Definition: elog.h:226
#define ereport(elevel,...)
Definition: elog.h:150
#define DEBUG4
Definition: elog.h:27
bool IsUnderPostmaster
Definition: globals.c:120
Oid MyDatabaseId
Definition: globals.c:94
Assert(PointerIsAligned(start, uint64))
#define malloc(a)
Definition: header.h:50
int j
Definition: isn.c:78
int i
Definition: isn.c:77
List * lappend_int(List *list, int datum)
Definition: list.c:357
#define VirtualTransactionIdIsValid(vxid)
Definition: lock.h:69
#define GET_VXID_FROM_PGPROC(vxid_dst, proc)
Definition: lock.h:79
#define InvalidLocalTransactionId
Definition: lock.h:67
#define VirtualTransactionIdEquals(vxid1, vxid2)
Definition: lock.h:73
char * get_database_name(Oid dbid)
Definition: lsyscache.c:1259
bool LWLockHeldByMe(LWLock *lock)
Definition: lwlock.c:1977
bool LWLockAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1174
bool LWLockHeldByMeInMode(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:2021
void LWLockRelease(LWLock *lock)
Definition: lwlock.c:1894
bool LWLockConditionalAcquire(LWLock *lock, LWLockMode mode)
Definition: lwlock.c:1345
@ LW_SHARED
Definition: lwlock.h:113
@ LW_EXCLUSIVE
Definition: lwlock.h:112
void pfree(void *pointer)
Definition: mcxt.c:1594
void * palloc(Size size)
Definition: mcxt.c:1365
#define AmStartupProcess()
Definition: miscadmin.h:389
#define IsBootstrapProcessingMode()
Definition: miscadmin.h:476
#define CHECK_FOR_INTERRUPTS()
Definition: miscadmin.h:122
Oid GetUserId(void)
Definition: miscinit.c:469
static bool pg_lfind32(uint32 key, const uint32 *base, uint32 nelem)
Definition: pg_lfind.h:153
#define NIL
Definition: pg_list.h:68
#define lfirst_int(lc)
Definition: pg_list.h:173
static char * buf
Definition: pg_test_fsync.c:72
#define qsort(a, b, c, d)
Definition: port.h:479
void PGSemaphoreUnlock(PGSemaphore sema)
Definition: posix_sema.c:339
void PGSemaphoreLock(PGSemaphore sema)
Definition: posix_sema.c:319
#define InvalidOid
Definition: postgres_ext.h:37
unsigned int Oid
Definition: postgres_ext.h:32
#define PROC_IN_LOGICAL_DECODING
Definition: proc.h:61
#define NUM_AUXILIARY_PROCS
Definition: proc.h:463
#define DELAY_CHKPT_IN_COMMIT
Definition: proc.h:137
#define PROC_XMIN_FLAGS
Definition: proc.h:72
#define PROC_AFFECTS_ALL_HORIZONS
Definition: proc.h:62
#define PROC_IN_VACUUM
Definition: proc.h:58
#define GetPGProcByNumber(n)
Definition: proc.h:440
#define GetNumberFromPGProc(proc)
Definition: proc.h:441
#define PROC_VACUUM_STATE_MASK
Definition: proc.h:65
#define PROC_IS_AUTOVACUUM
Definition: proc.h:57
KAXCompressReason
Definition: procarray.c:261
@ KAX_PRUNE
Definition: procarray.c:263
@ KAX_NO_SPACE
Definition: procarray.c:262
@ KAX_TRANSACTION_END
Definition: procarray.c:264
@ KAX_STARTUP_PROCESS_IDLE
Definition: procarray.c:265
static GlobalVisState GlobalVisDataRels
Definition: procarray.c:299
bool GlobalVisTestIsRemovableFullXid(GlobalVisState *state, FullTransactionId fxid)
Definition: procarray.c:4184
TransactionId GetOldestNonRemovableTransactionId(Relation rel)
Definition: procarray.c:1953
#define TOTAL_MAX_CACHED_SUBXIDS
static GlobalVisState GlobalVisSharedRels
Definition: procarray.c:297
void ProcArrayGetReplicationSlotXmin(TransactionId *xmin, TransactionId *catalog_xmin)
Definition: procarray.c:3930
static GlobalVisState GlobalVisCatalogRels
Definition: procarray.c:298
VirtualTransactionId * GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0, bool allDbs, int excludeVacuum, int *nvxids)
Definition: procarray.c:3286
bool GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
Definition: procarray.c:4226
bool GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
Definition: procarray.c:4248
static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
Definition: procarray.c:4627
pid_t SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode, bool conflictPending)
Definition: procarray.c:3459
Size ProcArrayShmemSize(void)
Definition: procarray.c:376
TransactionId GetOldestSafeDecodingTransactionId(bool catalogOnly)
Definition: procarray.c:2907
void XidCacheRemoveRunningXids(TransactionId xid, int nxids, const TransactionId *xids, TransactionId latestXid)
Definition: procarray.c:3953
static FullTransactionId FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
Definition: procarray.c:4283
bool MinimumActiveBackends(int min)
Definition: procarray.c:3508
void TerminateOtherDBBackends(Oid databaseId)
Definition: procarray.c:3790
#define xc_no_overflow_inc()
Definition: procarray.c:342
static TransactionId standbySnapshotPendingXmin
Definition: procarray.c:290
void ExpireAllKnownAssignedTransactionIds(void)
Definition: procarray.c:4460
#define UINT32_ACCESS_ONCE(var)
Definition: procarray.c:68
RunningTransactions GetRunningTransactionData(void)
Definition: procarray.c:2637
static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids, TransactionId *subxids)
Definition: procarray.c:4975
static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin, TransactionId xmax)
Definition: procarray.c:5089
#define xc_by_recent_xmin_inc()
Definition: procarray.c:335
void ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
Definition: procarray.c:667
void ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid, TransactionId *xmin, int *nsubxid, bool *overflowed)
Definition: procarray.c:3122
static PGPROC * allProcs
Definition: procarray.c:271
void RecordKnownAssignedTransactionIds(TransactionId xid)
Definition: procarray.c:4365
static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
Definition: procarray.c:5075
TransactionId GetOldestTransactionIdConsideredRunning(void)
Definition: procarray.c:1982
static TransactionId latestObservedXid
Definition: procarray.c:283
static ProcArrayStruct * procArray
Definition: procarray.c:269
int GetMaxSnapshotSubxidCount(void)
Definition: procarray.c:2028
int CountDBConnections(Oid databaseid)
Definition: procarray.c:3590
static GlobalVisState GlobalVisTempRels
Definition: procarray.c:300
#define xc_by_my_xact_inc()
Definition: procarray.c:337
#define xc_by_known_assigned_inc()
Definition: procarray.c:341
struct ProcArrayStruct ProcArrayStruct
void CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
Definition: procarray.c:3621
#define PROCARRAY_MAXPROCS
void GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
Definition: procarray.c:1995
static bool GlobalVisTestShouldUpdate(GlobalVisState *state)
Definition: procarray.c:4109
static void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
Definition: procarray.c:731
static void KnownAssignedXidsRemovePreceding(TransactionId removeXid)
Definition: procarray.c:4997
void ProcArrayAdd(PGPROC *proc)
Definition: procarray.c:468
struct ComputeXidHorizonsResult ComputeXidHorizonsResult
static TransactionId * KnownAssignedXids
Definition: procarray.c:281
#define xc_by_child_xid_inc()
Definition: procarray.c:340
pid_t CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
Definition: procarray.c:3453
Snapshot GetSnapshotData(Snapshot snapshot)
Definition: procarray.c:2123
static bool * KnownAssignedXidsValid
Definition: procarray.c:282
bool HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
Definition: procarray.c:3051
static void KnownAssignedXidsRemove(TransactionId xid)
Definition: procarray.c:4949
void KnownAssignedTransactionIdsIdleMaintenance(void)
Definition: procarray.c:4526
static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
Definition: procarray.c:4128
int GetMaxSnapshotXidCount(void)
Definition: procarray.c:2017
int CountDBBackends(Oid databaseid)
Definition: procarray.c:3561
PGPROC * BackendPidGetProcWithLock(int pid)
Definition: procarray.c:3181
bool GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
Definition: procarray.c:4262
#define MAXAUTOVACPIDS
PGPROC * BackendPidGetProc(int pid)
Definition: procarray.c:3158
bool ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
Definition: procarray.c:2564
PGPROC * ProcNumberGetProc(ProcNumber procNumber)
Definition: procarray.c:3100
#define KAX_COMPRESS_FREQUENCY
GlobalVisState * GlobalVisTestFor(Relation rel)
Definition: procarray.c:4069
static TransactionId KnownAssignedXidsGetOldestXmin(void)
Definition: procarray.c:5145
void ProcArrayApplyRecoveryInfo(RunningTransactions running)
Definition: procarray.c:1054
void ProcArrayClearTransaction(PGPROC *proc)
Definition: procarray.c:907
int CountUserBackends(Oid roleid)
Definition: procarray.c:3662
static TransactionId ComputeXidHorizonsResultLastXmin
Definition: procarray.c:307
static void GlobalVisUpdate(void)
Definition: procarray.c:4167
#define xc_slow_answer_inc()
Definition: procarray.c:343
static void KnownAssignedXidsDisplay(int trace_level)
Definition: procarray.c:5180
#define xc_by_main_xid_inc()
Definition: procarray.c:339
static void MaintainLatestCompletedXidRecovery(TransactionId latestXid)
Definition: procarray.c:989
static void ComputeXidHorizons(ComputeXidHorizonsResult *h)
Definition: procarray.c:1683
void ProcArrayApplyXidAssignment(TransactionId topxid, int nsubxids, TransactionId *subxids)
Definition: procarray.c:1318
static bool KnownAssignedXidExists(TransactionId xid)
Definition: procarray.c:4936
TransactionId GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
Definition: procarray.c:2833
void ProcArrayShmemInit(void)
Definition: procarray.c:418
bool CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
Definition: procarray.c:3712
GlobalVisHorizonKind
Definition: procarray.c:250
@ VISHORIZON_SHARED
Definition: procarray.c:251
@ VISHORIZON_DATA
Definition: procarray.c:253
@ VISHORIZON_CATALOG
Definition: procarray.c:252
@ VISHORIZON_TEMP
Definition: procarray.c:254
int BackendXidGetPid(TransactionId xid)
Definition: procarray.c:3218
#define xc_by_latest_xid_inc()
Definition: procarray.c:338
bool IsBackendPid(int pid)
Definition: procarray.c:3253
#define xc_by_known_xact_inc()
Definition: procarray.c:336
static bool KnownAssignedXidsSearch(TransactionId xid, bool remove)
Definition: procarray.c:4848
static void KnownAssignedXidsReset(void)
Definition: procarray.c:5218
static GlobalVisHorizonKind GlobalVisHorizonKindForRel(Relation rel)
Definition: procarray.c:1919
void ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin, bool already_locked)
Definition: procarray.c:3905
void ProcArrayInitRecovery(TransactionId initializedUptoXID)
Definition: procarray.c:1023
void ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
Definition: procarray.c:565
#define KAX_COMPRESS_IDLE_INTERVAL
VirtualTransactionId * GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
Definition: procarray.c:3379
static void MaintainLatestCompletedXid(TransactionId latestXid)
Definition: procarray.c:967
static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
Definition: procarray.c:792
void ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids, TransactionId *subxids, TransactionId max_xid)
Definition: procarray.c:4434
VirtualTransactionId * GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
Definition: procarray.c:3005
static TransactionId cachedXidIsNotInProgress
Definition: procarray.c:276
bool ProcArrayInstallImportedXmin(TransactionId xmin, VirtualTransactionId *sourcevxid)
Definition: procarray.c:2480
static bool GetSnapshotDataReuse(Snapshot snapshot)
Definition: procarray.c:2043
static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid, bool exclusive_lock)
Definition: procarray.c:4744
bool TransactionIdIsInProgress(TransactionId xid)
Definition: procarray.c:1402
void ExpireOldKnownAssignedTransactionIds(TransactionId xid)
Definition: procarray.c:4494
#define INVALID_PROC_NUMBER
Definition: procnumber.h:26
int ProcNumber
Definition: procnumber.h:24
int SendProcSignal(pid_t pid, ProcSignalReason reason, ProcNumber procNumber)
Definition: procsignal.c:284
ProcSignalReason
Definition: procsignal.h:31
#define RELATION_IS_LOCAL(relation)
Definition: rel.h:657
#define RelationIsAccessibleInLogicalDecoding(relation)
Definition: rel.h:693
Size add_size(Size s1, Size s2)
Definition: shmem.c:493
Size mul_size(Size s1, Size s2)
Definition: shmem.c:510
void * ShmemInitStruct(const char *name, Size size, bool *foundPtr)
Definition: shmem.c:387
void pg_usleep(long microsec)
Definition: signal.c:53
TransactionId RecentXmin
Definition: snapmgr.c:159
TransactionId TransactionXmin
Definition: snapmgr.c:158
PGPROC * MyProc
Definition: proc.c:66
PROC_HDR * ProcGlobal
Definition: proc.c:78
void StandbyReleaseOldLocks(TransactionId oldxid)
Definition: standby.c:1130
@ SUBXIDS_IN_SUBTRANS
Definition: standby.h:82
@ SUBXIDS_MISSING
Definition: standby.h:81
@ SUBXIDS_IN_ARRAY
Definition: standby.h:80
void appendStringInfo(StringInfo str, const char *fmt,...)
Definition: stringinfo.c:145
void initStringInfo(StringInfo str)
Definition: stringinfo.c:97
TransactionId slot_catalog_xmin
Definition: procarray.c:192
TransactionId data_oldest_nonremovable
Definition: procarray.c:237
TransactionId temp_oldest_nonremovable
Definition: procarray.c:243
TransactionId shared_oldest_nonremovable
Definition: procarray.c:214
TransactionId oldest_considered_running
Definition: procarray.c:205
TransactionId slot_xmin
Definition: procarray.c:191
FullTransactionId latest_completed
Definition: procarray.c:185
TransactionId catalog_oldest_nonremovable
Definition: procarray.c:231
TransactionId shared_oldest_nonremovable_raw
Definition: procarray.c:225
FullTransactionId definitely_needed
Definition: procarray.c:170
FullTransactionId maybe_needed
Definition: procarray.c:173
Definition: pg_list.h:54
Definition: proc.h:179
bool isRegularBackend
Definition: proc.h:230
TransactionId xmin
Definition: proc.h:194
bool procArrayGroupMember
Definition: proc.h:286
LocalTransactionId lxid
Definition: proc.h:217
pg_atomic_uint32 procArrayGroupNext
Definition: proc.h:288
uint8 statusFlags
Definition: proc.h:259
bool recoveryConflictPending
Definition: proc.h:237
Oid databaseId
Definition: proc.h:224
struct PGPROC::@128 vxid
ProcNumber procNumber
Definition: proc.h:212
int pid
Definition: proc.h:199
int pgxactoff
Definition: proc.h:201
XidCacheStatus subxidStatus
Definition: proc.h:280
LOCK * waitLock
Definition: proc.h:249
TransactionId xid
Definition: proc.h:189
struct XidCache subxids
Definition: proc.h:282
int delayChkptFlags
Definition: proc.h:257
TransactionId procArrayGroupMemberXid
Definition: proc.h:294
PGSemaphore sem
Definition: proc.h:183
Oid roleId
Definition: proc.h:225
Definition: proc.h:386
uint8 * statusFlags
Definition: proc.h:403
XidCacheStatus * subxidStates
Definition: proc.h:397
PGPROC * allProcs
Definition: proc.h:388
TransactionId * xids
Definition: proc.h:391
pg_atomic_uint32 procArrayGroupFirst
Definition: proc.h:416
uint32 allProcCount
Definition: proc.h:406
TransactionId replication_slot_xmin
Definition: procarray.c:94
int maxKnownAssignedXids
Definition: procarray.c:79
TransactionId replication_slot_catalog_xmin
Definition: procarray.c:96
int numKnownAssignedXids
Definition: procarray.c:80
int pgprocnos[FLEXIBLE_ARRAY_MEMBER]
Definition: procarray.c:99
TransactionId lastOverflowedXid
Definition: procarray.c:91
int tailKnownAssignedXids
Definition: procarray.c:81
int headKnownAssignedXids
Definition: procarray.c:82
Form_pg_class rd_rel
Definition: rel.h:111
TransactionId oldestRunningXid
Definition: standby.h:92
TransactionId nextXid
Definition: standby.h:91
TransactionId oldestDatabaseRunningXid
Definition: standby.h:93
TransactionId latestCompletedXid
Definition: standby.h:95
subxids_array_status subxid_status
Definition: standby.h:90
TransactionId * xids
Definition: standby.h:97
TransactionId xmin
Definition: snapshot.h:153
int32 subxcnt
Definition: snapshot.h:177
bool copied
Definition: snapshot.h:181
uint32 regd_count
Definition: snapshot.h:201
uint32 active_count
Definition: snapshot.h:200
CommandId curcid
Definition: snapshot.h:183
uint32 xcnt
Definition: snapshot.h:165
TransactionId * subxip
Definition: snapshot.h:176
uint64 snapXactCompletionCount
Definition: snapshot.h:209
TransactionId xmax
Definition: snapshot.h:154
TransactionId * xip
Definition: snapshot.h:164
bool suboverflowed
Definition: snapshot.h:178
bool takenDuringRecovery
Definition: snapshot.h:180
FullTransactionId latestCompletedXid
Definition: transam.h:238
FullTransactionId nextXid
Definition: transam.h:220
uint64 xactCompletionCount
Definition: transam.h:248
TransactionId oldestXid
Definition: transam.h:222
LocalTransactionId localTransactionId
Definition: lock.h:64
ProcNumber procNumber
Definition: lock.h:63
bool overflowed
Definition: proc.h:46
uint8 count
Definition: proc.h:44
TransactionId xids[PGPROC_MAX_CACHED_SUBXIDS]
Definition: proc.h:51
Definition: type.h:96
Definition: regguts.h:323
void SubTransSetParent(TransactionId xid, TransactionId parent)
Definition: subtrans.c:84
TransactionId SubTransGetTopmostTransaction(TransactionId xid)
Definition: subtrans.c:162
void ExtendSUBTRANS(TransactionId newestXact)
Definition: subtrans.c:353
bool superuser_arg(Oid roleid)
Definition: superuser.c:56
bool superuser(void)
Definition: superuser.c:46
TransactionId TransactionIdLatest(TransactionId mainxid, int nxids, const TransactionId *xids)
Definition: transam.c:345
bool TransactionIdDidCommit(TransactionId transactionId)
Definition: transam.c:126
bool TransactionIdPrecedes(TransactionId id1, TransactionId id2)
Definition: transam.c:280
bool TransactionIdPrecedesOrEquals(TransactionId id1, TransactionId id2)
Definition: transam.c:299
bool TransactionIdDidAbort(TransactionId transactionId)
Definition: transam.c:188
bool TransactionIdFollows(TransactionId id1, TransactionId id2)
Definition: transam.c:314
bool TransactionIdFollowsOrEquals(TransactionId id1, TransactionId id2)
Definition: transam.c:329
#define FullTransactionIdIsNormal(x)
Definition: transam.h:58
static FullTransactionId FullTransactionIdNewer(FullTransactionId a, FullTransactionId b)
Definition: transam.h:360
#define TransactionIdRetreat(dest)
Definition: transam.h:141
#define InvalidTransactionId
Definition: transam.h:31
static void FullTransactionIdRetreat(FullTransactionId *dest)
Definition: transam.h:103
#define U64FromFullTransactionId(x)
Definition: transam.h:49
static FullTransactionId FullTransactionIdFromU64(uint64 value)
Definition: transam.h:81
#define FullTransactionIdFollowsOrEquals(a, b)
Definition: transam.h:54
#define AssertTransactionIdInAllowableRange(xid)
Definition: transam.h:301
#define TransactionIdEquals(id1, id2)
Definition: transam.h:43
#define NormalTransactionIdPrecedes(id1, id2)
Definition: transam.h:147
#define XidFromFullTransactionId(x)
Definition: transam.h:48
static void FullTransactionIdAdvance(FullTransactionId *dest)
Definition: transam.h:128
#define TransactionIdIsValid(xid)
Definition: transam.h:41
#define TransactionIdIsNormal(xid)
Definition: transam.h:42
#define TransactionIdAdvance(dest)
Definition: transam.h:91
#define FullTransactionIdPrecedes(a, b)
Definition: transam.h:51
#define FullTransactionIdIsValid(x)
Definition: transam.h:55
static TransactionId TransactionIdOlder(TransactionId a, TransactionId b)
Definition: transam.h:334
bool StandbyTransactionIdIsPrepared(TransactionId xid)
Definition: twophase.c:1467
#define TimestampTzPlusMilliseconds(tz, ms)
Definition: timestamp.h:85
void AdvanceNextFullTransactionIdPastXid(TransactionId xid)
Definition: varsup.c:304
TransamVariablesData * TransamVariables
Definition: varsup.c:34
static void pgstat_report_wait_start(uint32 wait_event_info)
Definition: wait_event.h:69
static void pgstat_report_wait_end(void)
Definition: wait_event.h:85
const char * type
#define kill(pid, sig)
Definition: win32_port.h:493
bool TransactionIdIsCurrentTransactionId(TransactionId xid)
Definition: xact.c:941
CommandId GetCurrentCommandId(bool used)
Definition: xact.c:829
int xidLogicalComparator(const void *arg1, const void *arg2)
Definition: xid.c:169
bool RecoveryInProgress(void)
Definition: xlog.c:6383
bool EnableHotStandby
Definition: xlog.c:122
HotStandbyState standbyState
Definition: xlogutils.c:53
@ STANDBY_SNAPSHOT_READY
Definition: xlogutils.h:55
@ STANDBY_SNAPSHOT_PENDING
Definition: xlogutils.h:54
@ STANDBY_INITIALIZED
Definition: xlogutils.h:53