Line data Source code
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) 1996-2024, 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 "commands/dbcommands.h"
58 : #include "miscadmin.h"
59 : #include "pgstat.h"
60 : #include "port/pg_lfind.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "utils/acl.h"
64 : #include "utils/builtins.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 */
71 : typedef 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 : */
91 : TransactionId lastOverflowedXid;
92 :
93 : /* oldest xmin of any replication slot */
94 : TransactionId replication_slot_xmin;
95 : /* oldest catalog xmin of any replication slot */
96 : TransactionId replication_slot_catalog_xmin;
97 :
98 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
99 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
100 : } ProcArrayStruct;
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 : */
167 : struct GlobalVisState
168 : {
169 : /* XIDs >= are considered running by some backend */
170 : FullTransactionId definitely_needed;
171 :
172 : /* XIDs < are not considered to be running by any backend */
173 : FullTransactionId maybe_needed;
174 : };
175 :
176 : /*
177 : * Result of ComputeXidHorizons().
178 : */
179 : typedef struct ComputeXidHorizonsResult
180 : {
181 : /*
182 : * The value of TransamVariables->latestCompletedXid when
183 : * ComputeXidHorizons() held ProcArrayLock.
184 : */
185 : FullTransactionId latest_completed;
186 :
187 : /*
188 : * The same for procArray->replication_slot_xmin and.
189 : * procArray->replication_slot_catalog_xmin.
190 : */
191 : TransactionId slot_xmin;
192 : TransactionId slot_catalog_xmin;
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 : */
205 : TransactionId oldest_considered_running;
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 : */
214 : TransactionId shared_oldest_nonremovable;
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 : */
225 : TransactionId shared_oldest_nonremovable_raw;
226 :
227 : /*
228 : * Oldest xid for which deleted tuples need to be retained in non-shared
229 : * catalog tables.
230 : */
231 : TransactionId catalog_oldest_nonremovable;
232 :
233 : /*
234 : * Oldest xid for which deleted tuples need to be retained in normal user
235 : * defined tables.
236 : */
237 : TransactionId data_oldest_nonremovable;
238 :
239 : /*
240 : * Oldest xid for which deleted tuples need to be retained in this
241 : * session's temporary tables.
242 : */
243 : TransactionId temp_oldest_nonremovable;
244 : } ComputeXidHorizonsResult;
245 :
246 : /*
247 : * Return value for GlobalVisHorizonKindForRel().
248 : */
249 : typedef enum GlobalVisHorizonKind
250 : {
251 : VISHORIZON_SHARED,
252 : VISHORIZON_CATALOG,
253 : VISHORIZON_DATA,
254 : VISHORIZON_TEMP,
255 : } GlobalVisHorizonKind;
256 :
257 : /*
258 : * Reason codes for KnownAssignedXidsCompress().
259 : */
260 : typedef enum KAXCompressReason
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 */
266 : } KAXCompressReason;
267 :
268 :
269 : static ProcArrayStruct *procArray;
270 :
271 : static PGPROC *allProcs;
272 :
273 : /*
274 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
275 : */
276 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
277 :
278 : /*
279 : * Bookkeeping for tracking emulated transactions in recovery
280 : */
281 : static TransactionId *KnownAssignedXids;
282 : static bool *KnownAssignedXidsValid;
283 : static TransactionId latestObservedXid = InvalidTransactionId;
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 : */
290 : static TransactionId standbySnapshotPendingXmin;
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 : */
297 : static GlobalVisState GlobalVisSharedRels;
298 : static GlobalVisState GlobalVisCatalogRels;
299 : static GlobalVisState GlobalVisDataRels;
300 : static GlobalVisState GlobalVisTempRels;
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 : */
307 : static TransactionId ComputeXidHorizonsResultLastXmin;
308 :
309 : #ifdef XIDCACHE_DEBUG
310 :
311 : /* counters for XidCache measurement */
312 : static long xc_by_recent_xmin = 0;
313 : static long xc_by_known_xact = 0;
314 : static long xc_by_my_xact = 0;
315 : static long xc_by_latest_xid = 0;
316 : static long xc_by_main_xid = 0;
317 : static long xc_by_child_xid = 0;
318 : static long xc_by_known_assigned = 0;
319 : static long xc_no_overflow = 0;
320 : static 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 :
332 : static 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 */
347 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
348 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
349 : bool exclusive_lock);
350 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
351 : static bool KnownAssignedXidExists(TransactionId xid);
352 : static void KnownAssignedXidsRemove(TransactionId xid);
353 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
354 : TransactionId *subxids);
355 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
356 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
357 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
358 : TransactionId *xmin,
359 : TransactionId xmax);
360 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
361 : static void KnownAssignedXidsDisplay(int trace_level);
362 : static void KnownAssignedXidsReset(void);
363 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
364 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
365 : static void MaintainLatestCompletedXid(TransactionId latestXid);
366 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
367 :
368 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
369 : TransactionId xid);
370 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
371 :
372 : /*
373 : * Report shared-memory space needed by CreateSharedProcArray.
374 : */
375 : Size
376 3298 : ProcArrayShmemSize(void)
377 : {
378 : Size size;
379 :
380 : /* Size of the ProcArray structure itself */
381 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
382 :
383 3298 : size = offsetof(ProcArrayStruct, pgprocnos);
384 3298 : 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 :
402 3298 : if (EnableHotStandby)
403 : {
404 3278 : size = add_size(size,
405 : mul_size(sizeof(TransactionId),
406 3278 : TOTAL_MAX_CACHED_SUBXIDS));
407 3278 : size = add_size(size,
408 3278 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
409 : }
410 :
411 3298 : return size;
412 : }
413 :
414 : /*
415 : * Initialize the shared PGPROC array during postmaster startup.
416 : */
417 : void
418 1768 : CreateSharedProcArray(void)
419 : {
420 : bool found;
421 :
422 : /* Create or attach to the ProcArray shared structure */
423 1768 : procArray = (ProcArrayStruct *)
424 1768 : ShmemInitStruct("Proc Array",
425 : add_size(offsetof(ProcArrayStruct, pgprocnos),
426 : mul_size(sizeof(int),
427 1768 : PROCARRAY_MAXPROCS)),
428 : &found);
429 :
430 1768 : if (!found)
431 : {
432 : /*
433 : * We're the first - initialize.
434 : */
435 1768 : procArray->numProcs = 0;
436 1768 : procArray->maxProcs = PROCARRAY_MAXPROCS;
437 1768 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
438 1768 : procArray->numKnownAssignedXids = 0;
439 1768 : procArray->tailKnownAssignedXids = 0;
440 1768 : procArray->headKnownAssignedXids = 0;
441 1768 : procArray->lastOverflowedXid = InvalidTransactionId;
442 1768 : procArray->replication_slot_xmin = InvalidTransactionId;
443 1768 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
444 1768 : TransamVariables->xactCompletionCount = 1;
445 : }
446 :
447 1768 : allProcs = ProcGlobal->allProcs;
448 :
449 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
450 1768 : if (EnableHotStandby)
451 : {
452 1758 : KnownAssignedXids = (TransactionId *)
453 1758 : ShmemInitStruct("KnownAssignedXids",
454 : mul_size(sizeof(TransactionId),
455 1758 : TOTAL_MAX_CACHED_SUBXIDS),
456 : &found);
457 1758 : KnownAssignedXidsValid = (bool *)
458 1758 : ShmemInitStruct("KnownAssignedXidsValid",
459 1758 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
460 : &found);
461 : }
462 1768 : }
463 :
464 : /*
465 : * Add the specified PGPROC to the shared array.
466 : */
467 : void
468 26480 : ProcArrayAdd(PGPROC *proc)
469 : {
470 26480 : int pgprocno = GetNumberFromPGProc(proc);
471 26480 : ProcArrayStruct *arrayP = procArray;
472 : int index;
473 : int movecount;
474 :
475 : /* See ProcGlobal comment explaining why both locks are held */
476 26480 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
477 26480 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
478 :
479 26480 : 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 : */
486 0 : ereport(FATAL,
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 58864 : for (index = 0; index < arrayP->numProcs; index++)
501 : {
502 51772 : 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 51772 : if (this_procno > pgprocno)
509 19388 : break;
510 : }
511 :
512 26480 : movecount = arrayP->numProcs - index;
513 26480 : memmove(&arrayP->pgprocnos[index + 1],
514 26480 : &arrayP->pgprocnos[index],
515 : movecount * sizeof(*arrayP->pgprocnos));
516 26480 : memmove(&ProcGlobal->xids[index + 1],
517 26480 : &ProcGlobal->xids[index],
518 : movecount * sizeof(*ProcGlobal->xids));
519 26480 : memmove(&ProcGlobal->subxidStates[index + 1],
520 26480 : &ProcGlobal->subxidStates[index],
521 : movecount * sizeof(*ProcGlobal->subxidStates));
522 26480 : memmove(&ProcGlobal->statusFlags[index + 1],
523 26480 : &ProcGlobal->statusFlags[index],
524 : movecount * sizeof(*ProcGlobal->statusFlags));
525 :
526 26480 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
527 26480 : proc->pgxactoff = index;
528 26480 : ProcGlobal->xids[index] = proc->xid;
529 26480 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
530 26480 : ProcGlobal->statusFlags[index] = proc->statusFlags;
531 :
532 26480 : arrayP->numProcs++;
533 :
534 : /* adjust pgxactoff for all following PGPROCs */
535 26480 : index++;
536 73532 : for (; index < arrayP->numProcs; index++)
537 : {
538 47052 : 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 47052 : 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 26480 : LWLockRelease(XidGenLock);
551 26480 : LWLockRelease(ProcArrayLock);
552 26480 : }
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 : */
564 : void
565 26434 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
566 : {
567 26434 : 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 26434 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
579 26434 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
580 :
581 26434 : myoff = proc->pgxactoff;
582 :
583 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
584 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
585 :
586 26434 : if (TransactionIdIsValid(latestXid))
587 : {
588 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
589 :
590 : /* Advance global latestCompletedXid while holding the lock */
591 762 : MaintainLatestCompletedXid(latestXid);
592 :
593 : /* Same with xactCompletionCount */
594 762 : TransamVariables->xactCompletionCount++;
595 :
596 762 : ProcGlobal->xids[myoff] = InvalidTransactionId;
597 762 : ProcGlobal->subxidStates[myoff].overflowed = false;
598 762 : ProcGlobal->subxidStates[myoff].count = 0;
599 : }
600 : else
601 : {
602 : /* Shouldn't be trying to remove a live transaction here */
603 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
604 : }
605 :
606 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
607 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
608 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
609 :
610 26434 : ProcGlobal->statusFlags[myoff] = 0;
611 :
612 : /* Keep the PGPROC array sorted. See notes above */
613 26434 : movecount = arrayP->numProcs - myoff - 1;
614 26434 : memmove(&arrayP->pgprocnos[myoff],
615 26434 : &arrayP->pgprocnos[myoff + 1],
616 : movecount * sizeof(*arrayP->pgprocnos));
617 26434 : memmove(&ProcGlobal->xids[myoff],
618 26434 : &ProcGlobal->xids[myoff + 1],
619 : movecount * sizeof(*ProcGlobal->xids));
620 26434 : memmove(&ProcGlobal->subxidStates[myoff],
621 26434 : &ProcGlobal->subxidStates[myoff + 1],
622 : movecount * sizeof(*ProcGlobal->subxidStates));
623 26434 : memmove(&ProcGlobal->statusFlags[myoff],
624 26434 : &ProcGlobal->statusFlags[myoff + 1],
625 : movecount * sizeof(*ProcGlobal->statusFlags));
626 :
627 26434 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
628 26434 : arrayP->numProcs--;
629 :
630 : /*
631 : * Adjust pgxactoff of following procs for removed PGPROC (note that
632 : * numProcs already has been decremented).
633 : */
634 77002 : for (int index = myoff; index < arrayP->numProcs; index++)
635 : {
636 50568 : 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 50568 : 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 26434 : LWLockRelease(XidGenLock);
649 26434 : LWLockRelease(ProcArrayLock);
650 26434 : }
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 : */
666 : void
667 565472 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
668 : {
669 565472 : 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 : */
677 : Assert(TransactionIdIsValid(proc->xid));
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 231246 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
685 : {
686 230830 : ProcArrayEndTransactionInternal(proc, latestXid);
687 230830 : LWLockRelease(ProcArrayLock);
688 : }
689 : else
690 416 : 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 : */
699 : Assert(!TransactionIdIsValid(proc->xid));
700 : Assert(proc->subxidStatus.count == 0);
701 : Assert(!proc->subxidStatus.overflowed);
702 :
703 334226 : proc->vxid.lxid = InvalidLocalTransactionId;
704 334226 : proc->xmin = InvalidTransactionId;
705 :
706 : /* be sure this is cleared in abort */
707 334226 : proc->delayChkptFlags = 0;
708 :
709 334226 : proc->recoveryConflictPending = false;
710 :
711 : /* must be cleared with xid/xmin: */
712 : /* avoid unnecessarily dirtying shared cachelines */
713 334226 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
714 : {
715 : Assert(!LWLockHeldByMe(ProcArrayLock));
716 19902 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
717 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
718 19902 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
719 19902 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
720 19902 : LWLockRelease(ProcArrayLock);
721 : }
722 : }
723 565472 : }
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 : */
730 : static inline void
731 231246 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
732 : {
733 231246 : 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 : */
739 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
740 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
741 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
742 :
743 231246 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
744 231246 : proc->xid = InvalidTransactionId;
745 231246 : proc->vxid.lxid = InvalidLocalTransactionId;
746 231246 : proc->xmin = InvalidTransactionId;
747 :
748 : /* be sure this is cleared in abort */
749 231246 : proc->delayChkptFlags = 0;
750 :
751 231246 : proc->recoveryConflictPending = false;
752 :
753 : /* must be cleared with xid/xmin: */
754 : /* avoid unnecessarily dirtying shared cachelines */
755 231246 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
756 : {
757 1194 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
758 1194 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
759 : }
760 :
761 : /* Clear the subtransaction-XID cache too while holding the lock */
762 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
763 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
764 231246 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
765 : {
766 914 : ProcGlobal->subxidStates[pgxactoff].count = 0;
767 914 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
768 914 : proc->subxidStatus.count = 0;
769 914 : proc->subxidStatus.overflowed = false;
770 : }
771 :
772 : /* Also advance global latestCompletedXid while holding the lock */
773 231246 : MaintainLatestCompletedXid(latestXid);
774 :
775 : /* Same with xactCompletionCount */
776 231246 : TransamVariables->xactCompletionCount++;
777 231246 : }
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 : */
791 : static void
792 416 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
793 : {
794 416 : int pgprocno = GetNumberFromPGProc(proc);
795 416 : PROC_HDR *procglobal = ProcGlobal;
796 : uint32 nextidx;
797 : uint32 wakeidx;
798 :
799 : /* We should definitely have an XID to clear. */
800 : Assert(TransactionIdIsValid(proc->xid));
801 :
802 : /* Add ourselves to the list of processes needing a group XID clear. */
803 416 : proc->procArrayGroupMember = true;
804 416 : proc->procArrayGroupMemberXid = latestXid;
805 416 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
806 : while (true)
807 : {
808 418 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
809 :
810 418 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
811 : &nextidx,
812 : (uint32) pgprocno))
813 416 : 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 : */
822 416 : if (nextidx != INVALID_PROC_NUMBER)
823 : {
824 12 : int extraWaits = 0;
825 :
826 : /* Sleep until the leader clears our XID. */
827 12 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
828 : for (;;)
829 : {
830 : /* acts as a read barrier */
831 12 : PGSemaphoreLock(proc->sem);
832 12 : if (!proc->procArrayGroupMember)
833 12 : break;
834 0 : extraWaits++;
835 : }
836 12 : pgstat_report_wait_end();
837 :
838 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER);
839 :
840 : /* Fix semaphore count for any absorbed wakeups */
841 12 : while (extraWaits-- > 0)
842 0 : PGSemaphoreUnlock(proc->sem);
843 12 : return;
844 : }
845 :
846 : /* We are the leader. Acquire the lock on behalf of everyone. */
847 404 : 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 : */
854 404 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
855 : INVALID_PROC_NUMBER);
856 :
857 : /* Remember head of list so we can perform wakeups after dropping lock. */
858 404 : wakeidx = nextidx;
859 :
860 : /* Walk the list and clear all XIDs. */
861 820 : while (nextidx != INVALID_PROC_NUMBER)
862 : {
863 416 : PGPROC *nextproc = &allProcs[nextidx];
864 :
865 416 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
866 :
867 : /* Move to next proc in list. */
868 416 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
869 : }
870 :
871 : /* We're done with the lock now. */
872 404 : 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 : */
881 820 : while (wakeidx != INVALID_PROC_NUMBER)
882 : {
883 416 : PGPROC *nextproc = &allProcs[wakeidx];
884 :
885 416 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
886 416 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
887 :
888 : /* ensure all previous writes are visible before follower continues. */
889 416 : pg_write_barrier();
890 :
891 416 : nextproc->procArrayGroupMember = false;
892 :
893 416 : if (nextproc != MyProc)
894 12 : 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 : */
906 : void
907 746 : ProcArrayClearTransaction(PGPROC *proc)
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 : */
925 746 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
926 :
927 746 : pgxactoff = proc->pgxactoff;
928 :
929 746 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
930 746 : proc->xid = InvalidTransactionId;
931 :
932 746 : proc->vxid.lxid = InvalidLocalTransactionId;
933 746 : proc->xmin = InvalidTransactionId;
934 746 : proc->recoveryConflictPending = false;
935 :
936 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
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 : */
946 746 : TransamVariables->xactCompletionCount++;
947 :
948 : /* Clear the subtransaction-XID cache too */
949 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
950 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
951 746 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
952 : {
953 272 : ProcGlobal->subxidStates[pgxactoff].count = 0;
954 272 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
955 272 : proc->subxidStatus.count = 0;
956 272 : proc->subxidStatus.overflowed = false;
957 : }
958 :
959 746 : LWLockRelease(ProcArrayLock);
960 746 : }
961 :
962 : /*
963 : * Update TransamVariables->latestCompletedXid to point to latestXid if
964 : * currently older.
965 : */
966 : static void
967 233300 : MaintainLatestCompletedXid(TransactionId latestXid)
968 : {
969 233300 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
970 :
971 : Assert(FullTransactionIdIsValid(cur_latest));
972 : Assert(!RecoveryInProgress());
973 : Assert(LWLockHeldByMe(ProcArrayLock));
974 :
975 233300 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
976 : {
977 213990 : TransamVariables->latestCompletedXid =
978 213990 : FullXidRelativeTo(cur_latest, latestXid);
979 : }
980 :
981 : Assert(IsBootstrapProcessingMode() ||
982 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
983 233300 : }
984 :
985 : /*
986 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
987 : */
988 : static void
989 40076 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
990 : {
991 40076 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
992 : FullTransactionId rel;
993 :
994 : Assert(AmStartupProcess() || !IsUnderPostmaster);
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 : */
1002 40076 : rel = TransamVariables->nextXid;
1003 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1004 :
1005 79966 : if (!FullTransactionIdIsValid(cur_latest) ||
1006 39890 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1007 : {
1008 33652 : TransamVariables->latestCompletedXid =
1009 33652 : FullXidRelativeTo(rel, latestXid);
1010 : }
1011 :
1012 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1013 40076 : }
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 : */
1022 : void
1023 186 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1024 : {
1025 : Assert(standbyState == STANDBY_INITIALIZED);
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 186 : latestObservedXid = initializedUptoXID;
1035 186 : TransactionIdRetreat(latestObservedXid);
1036 186 : }
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 : */
1053 : void
1054 506 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1055 : {
1056 : TransactionId *xids;
1057 : TransactionId advanceNextXid;
1058 : int nxids;
1059 : int i;
1060 :
1061 : Assert(standbyState >= STANDBY_INITIALIZED);
1062 : Assert(TransactionIdIsValid(running->nextXid));
1063 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1064 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1065 :
1066 : /*
1067 : * Remove stale transactions, if any.
1068 : */
1069 506 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
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 506 : advanceNextXid = running->nextXid;
1077 506 : TransactionIdRetreat(advanceNextXid);
1078 506 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1079 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1080 :
1081 : /*
1082 : * Remove stale locks, if any.
1083 : */
1084 506 : StandbyReleaseOldLocks(running->oldestRunningXid);
1085 :
1086 : /*
1087 : * If our snapshot is already valid, nothing else to do...
1088 : */
1089 506 : if (standbyState == STANDBY_SNAPSHOT_READY)
1090 320 : 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 : */
1103 186 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
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 0 : if (!running->subxid_overflow || 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 : */
1115 0 : KnownAssignedXidsReset();
1116 0 : standbyState = STANDBY_INITIALIZED;
1117 : }
1118 : else
1119 : {
1120 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1121 : running->oldestRunningXid))
1122 : {
1123 0 : standbyState = STANDBY_SNAPSHOT_READY;
1124 0 : elog(DEBUG1,
1125 : "recovery snapshots are now enabled");
1126 : }
1127 : else
1128 0 : elog(DEBUG1,
1129 : "recovery snapshot waiting for non-overflowed snapshot or "
1130 : "until oldest active xid on standby is at least %u (now %u)",
1131 : standbySnapshotPendingXmin,
1132 : running->oldestRunningXid);
1133 0 : return;
1134 : }
1135 : }
1136 :
1137 : Assert(standbyState == STANDBY_INITIALIZED);
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 186 : 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 186 : 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 186 : nxids = 0;
1171 192 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1172 : {
1173 6 : 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 : */
1181 6 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1182 0 : continue;
1183 :
1184 6 : xids[nxids++] = xid;
1185 : }
1186 :
1187 186 : if (nxids > 0)
1188 : {
1189 6 : if (procArray->numKnownAssignedXids != 0)
1190 : {
1191 0 : LWLockRelease(ProcArrayLock);
1192 0 : 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 6 : 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 12 : for (i = 0; i < nxids; i++)
1212 : {
1213 6 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1214 : {
1215 0 : elog(DEBUG1,
1216 : "found duplicated transaction %u for KnownAssignedXids insertion",
1217 : xids[i]);
1218 0 : continue;
1219 : }
1220 6 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1221 : }
1222 :
1223 6 : KnownAssignedXidsDisplay(DEBUG3);
1224 : }
1225 :
1226 186 : 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 : */
1238 : Assert(TransactionIdIsNormal(latestObservedXid));
1239 186 : TransactionIdAdvance(latestObservedXid);
1240 186 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1241 : {
1242 0 : ExtendSUBTRANS(latestObservedXid);
1243 0 : TransactionIdAdvance(latestObservedXid);
1244 : }
1245 186 : 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 186 : if (running->subxid_overflow)
1262 : {
1263 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1264 :
1265 0 : standbySnapshotPendingXmin = latestObservedXid;
1266 0 : procArray->lastOverflowedXid = latestObservedXid;
1267 : }
1268 : else
1269 : {
1270 186 : standbyState = STANDBY_SNAPSHOT_READY;
1271 :
1272 186 : standbySnapshotPendingXmin = InvalidTransactionId;
1273 : }
1274 :
1275 : /*
1276 : * If a transaction wrote a commit record in the gap between taking and
1277 : * logging the snapshot then latestCompletedXid may already be higher than
1278 : * the value from the snapshot, so check before we use the incoming value.
1279 : * It also might not yet be set at all.
1280 : */
1281 186 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1282 :
1283 : /*
1284 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1285 : * nobody can see it yet.
1286 : */
1287 :
1288 186 : LWLockRelease(ProcArrayLock);
1289 :
1290 186 : KnownAssignedXidsDisplay(DEBUG3);
1291 186 : if (standbyState == STANDBY_SNAPSHOT_READY)
1292 186 : elog(DEBUG1, "recovery snapshots are now enabled");
1293 : else
1294 0 : elog(DEBUG1,
1295 : "recovery snapshot waiting for non-overflowed snapshot or "
1296 : "until oldest active xid on standby is at least %u (now %u)",
1297 : standbySnapshotPendingXmin,
1298 : running->oldestRunningXid);
1299 : }
1300 :
1301 : /*
1302 : * ProcArrayApplyXidAssignment
1303 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1304 : */
1305 : void
1306 42 : ProcArrayApplyXidAssignment(TransactionId topxid,
1307 : int nsubxids, TransactionId *subxids)
1308 : {
1309 : TransactionId max_xid;
1310 : int i;
1311 :
1312 : Assert(standbyState >= STANDBY_INITIALIZED);
1313 :
1314 42 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1315 :
1316 : /*
1317 : * Mark all the subtransactions as observed.
1318 : *
1319 : * NOTE: This will fail if the subxid contains too many previously
1320 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1321 : * as the code stands, because xid-assignment records should never contain
1322 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1323 : */
1324 42 : RecordKnownAssignedTransactionIds(max_xid);
1325 :
1326 : /*
1327 : * Notice that we update pg_subtrans with the top-level xid, rather than
1328 : * the parent xid. This is a difference between normal processing and
1329 : * recovery, yet is still correct in all cases. The reason is that
1330 : * subtransaction commit is not marked in clog until commit processing, so
1331 : * all aborted subtransactions have already been clearly marked in clog.
1332 : * As a result we are able to refer directly to the top-level
1333 : * transaction's state rather than skipping through all the intermediate
1334 : * states in the subtransaction tree. This should be the first time we
1335 : * have attempted to SubTransSetParent().
1336 : */
1337 2730 : for (i = 0; i < nsubxids; i++)
1338 2688 : SubTransSetParent(subxids[i], topxid);
1339 :
1340 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1341 42 : if (standbyState == STANDBY_INITIALIZED)
1342 0 : return;
1343 :
1344 : /*
1345 : * Uses same locking as transaction commit
1346 : */
1347 42 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1348 :
1349 : /*
1350 : * Remove subxids from known-assigned-xacts.
1351 : */
1352 42 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1353 :
1354 : /*
1355 : * Advance lastOverflowedXid to be at least the last of these subxids.
1356 : */
1357 42 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1358 42 : procArray->lastOverflowedXid = max_xid;
1359 :
1360 42 : LWLockRelease(ProcArrayLock);
1361 : }
1362 :
1363 : /*
1364 : * TransactionIdIsInProgress -- is given transaction running in some backend
1365 : *
1366 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1367 : * there are four possibilities for finding a running transaction:
1368 : *
1369 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1370 : * by looking at ProcGlobal->xids.
1371 : *
1372 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1373 : * We can find this out cheaply too.
1374 : *
1375 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1376 : * if the Xid is running on the primary.
1377 : *
1378 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1379 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1380 : * This is the slowest way, but sadly it has to be done always if the others
1381 : * failed, unless we see that the cached subxact sets are complete (none have
1382 : * overflowed).
1383 : *
1384 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1385 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1386 : * This buys back some concurrency (and we can't retrieve the main Xids from
1387 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1388 : */
1389 : bool
1390 10509812 : TransactionIdIsInProgress(TransactionId xid)
1391 : {
1392 : static TransactionId *xids = NULL;
1393 : static TransactionId *other_xids;
1394 : XidCacheStatus *other_subxidstates;
1395 10509812 : int nxids = 0;
1396 10509812 : ProcArrayStruct *arrayP = procArray;
1397 : TransactionId topxid;
1398 : TransactionId latestCompletedXid;
1399 : int mypgxactoff;
1400 : int numProcs;
1401 : int j;
1402 :
1403 : /*
1404 : * Don't bother checking a transaction older than RecentXmin; it could not
1405 : * possibly still be running. (Note: in particular, this guarantees that
1406 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1407 : * running.)
1408 : */
1409 10509812 : if (TransactionIdPrecedes(xid, RecentXmin))
1410 : {
1411 : xc_by_recent_xmin_inc();
1412 8296626 : return false;
1413 : }
1414 :
1415 : /*
1416 : * We may have just checked the status of this transaction, so if it is
1417 : * already known to be completed, we can fall out without any access to
1418 : * shared memory.
1419 : */
1420 2213186 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1421 : {
1422 : xc_by_known_xact_inc();
1423 1797798 : return false;
1424 : }
1425 :
1426 : /*
1427 : * Also, we can handle our own transaction (and subtransactions) without
1428 : * any access to shared memory.
1429 : */
1430 415388 : if (TransactionIdIsCurrentTransactionId(xid))
1431 : {
1432 : xc_by_my_xact_inc();
1433 386418 : return true;
1434 : }
1435 :
1436 : /*
1437 : * If first time through, get workspace to remember main XIDs in. We
1438 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1439 : */
1440 28970 : if (xids == NULL)
1441 : {
1442 : /*
1443 : * In hot standby mode, reserve enough space to hold all xids in the
1444 : * known-assigned list. If we later finish recovery, we no longer need
1445 : * the bigger array, but we don't bother to shrink it.
1446 : */
1447 1180 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1448 :
1449 1180 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1450 1180 : if (xids == NULL)
1451 0 : ereport(ERROR,
1452 : (errcode(ERRCODE_OUT_OF_MEMORY),
1453 : errmsg("out of memory")));
1454 : }
1455 :
1456 28970 : other_xids = ProcGlobal->xids;
1457 28970 : other_subxidstates = ProcGlobal->subxidStates;
1458 :
1459 28970 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1460 :
1461 : /*
1462 : * Now that we have the lock, we can check latestCompletedXid; if the
1463 : * target Xid is after that, it's surely still running.
1464 : */
1465 28970 : latestCompletedXid =
1466 28970 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1467 28970 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1468 : {
1469 9304 : LWLockRelease(ProcArrayLock);
1470 : xc_by_latest_xid_inc();
1471 9304 : return true;
1472 : }
1473 :
1474 : /* No shortcuts, gotta grovel through the array */
1475 19666 : mypgxactoff = MyProc->pgxactoff;
1476 19666 : numProcs = arrayP->numProcs;
1477 178010 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1478 : {
1479 : int pgprocno;
1480 : PGPROC *proc;
1481 : TransactionId pxid;
1482 : int pxids;
1483 :
1484 : /* Ignore ourselves --- dealt with it above */
1485 161516 : if (pgxactoff == mypgxactoff)
1486 18058 : continue;
1487 :
1488 : /* Fetch xid just once - see GetNewTransactionId */
1489 143458 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1490 :
1491 143458 : if (!TransactionIdIsValid(pxid))
1492 100948 : continue;
1493 :
1494 : /*
1495 : * Step 1: check the main Xid
1496 : */
1497 42510 : if (TransactionIdEquals(pxid, xid))
1498 : {
1499 2916 : LWLockRelease(ProcArrayLock);
1500 : xc_by_main_xid_inc();
1501 2916 : return true;
1502 : }
1503 :
1504 : /*
1505 : * We can ignore main Xids that are younger than the target Xid, since
1506 : * the target could not possibly be their child.
1507 : */
1508 39594 : if (TransactionIdPrecedes(xid, pxid))
1509 15560 : continue;
1510 :
1511 : /*
1512 : * Step 2: check the cached child-Xids arrays
1513 : */
1514 24034 : pxids = other_subxidstates[pgxactoff].count;
1515 24034 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1516 24034 : pgprocno = arrayP->pgprocnos[pgxactoff];
1517 24034 : proc = &allProcs[pgprocno];
1518 52454 : for (j = pxids - 1; j >= 0; j--)
1519 : {
1520 : /* Fetch xid just once - see GetNewTransactionId */
1521 28676 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1522 :
1523 28676 : if (TransactionIdEquals(cxid, xid))
1524 : {
1525 256 : LWLockRelease(ProcArrayLock);
1526 : xc_by_child_xid_inc();
1527 256 : return true;
1528 : }
1529 : }
1530 :
1531 : /*
1532 : * Save the main Xid for step 4. We only need to remember main Xids
1533 : * that have uncached children. (Note: there is no race condition
1534 : * here because the overflowed flag cannot be cleared, only set, while
1535 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1536 : * worry about.)
1537 : */
1538 23778 : if (other_subxidstates[pgxactoff].overflowed)
1539 314 : xids[nxids++] = pxid;
1540 : }
1541 :
1542 : /*
1543 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1544 : * in the list must be treated as running.
1545 : */
1546 16494 : if (RecoveryInProgress())
1547 : {
1548 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1549 : Assert(nxids == 0);
1550 :
1551 0 : if (KnownAssignedXidExists(xid))
1552 : {
1553 0 : LWLockRelease(ProcArrayLock);
1554 : xc_by_known_assigned_inc();
1555 0 : return true;
1556 : }
1557 :
1558 : /*
1559 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1560 : * too. Fetch all xids from KnownAssignedXids that are lower than
1561 : * xid, since if xid is a subtransaction its parent will always have a
1562 : * lower value. Note we will collect both main and subXIDs here, but
1563 : * there's no help for it.
1564 : */
1565 0 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1566 0 : nxids = KnownAssignedXidsGet(xids, xid);
1567 : }
1568 :
1569 16494 : LWLockRelease(ProcArrayLock);
1570 :
1571 : /*
1572 : * If none of the relevant caches overflowed, we know the Xid is not
1573 : * running without even looking at pg_subtrans.
1574 : */
1575 16494 : if (nxids == 0)
1576 : {
1577 : xc_no_overflow_inc();
1578 16180 : cachedXidIsNotInProgress = xid;
1579 16180 : return false;
1580 : }
1581 :
1582 : /*
1583 : * Step 4: have to check pg_subtrans.
1584 : *
1585 : * At this point, we know it's either a subtransaction of one of the Xids
1586 : * in xids[], or it's not running. If it's an already-failed
1587 : * subtransaction, we want to say "not running" even though its parent may
1588 : * still be running. So first, check pg_xact to see if it's been aborted.
1589 : */
1590 : xc_slow_answer_inc();
1591 :
1592 314 : if (TransactionIdDidAbort(xid))
1593 : {
1594 0 : cachedXidIsNotInProgress = xid;
1595 0 : return false;
1596 : }
1597 :
1598 : /*
1599 : * It isn't aborted, so check whether the transaction tree it belongs to
1600 : * is still running (or, more precisely, whether it was running when we
1601 : * held ProcArrayLock).
1602 : */
1603 314 : topxid = SubTransGetTopmostTransaction(xid);
1604 : Assert(TransactionIdIsValid(topxid));
1605 628 : if (!TransactionIdEquals(topxid, xid) &&
1606 314 : pg_lfind32(topxid, xids, nxids))
1607 314 : return true;
1608 :
1609 0 : cachedXidIsNotInProgress = xid;
1610 0 : return false;
1611 : }
1612 :
1613 : /*
1614 : * TransactionIdIsActive -- is xid the top-level XID of an active backend?
1615 : *
1616 : * This differs from TransactionIdIsInProgress in that it ignores prepared
1617 : * transactions, as well as transactions running on the primary if we're in
1618 : * hot standby. Also, we ignore subtransactions since that's not needed
1619 : * for current uses.
1620 : */
1621 : bool
1622 0 : TransactionIdIsActive(TransactionId xid)
1623 : {
1624 0 : bool result = false;
1625 0 : ProcArrayStruct *arrayP = procArray;
1626 0 : TransactionId *other_xids = ProcGlobal->xids;
1627 : int i;
1628 :
1629 : /*
1630 : * Don't bother checking a transaction older than RecentXmin; it could not
1631 : * possibly still be running.
1632 : */
1633 0 : if (TransactionIdPrecedes(xid, RecentXmin))
1634 0 : return false;
1635 :
1636 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1637 :
1638 0 : for (i = 0; i < arrayP->numProcs; i++)
1639 : {
1640 0 : int pgprocno = arrayP->pgprocnos[i];
1641 0 : PGPROC *proc = &allProcs[pgprocno];
1642 : TransactionId pxid;
1643 :
1644 : /* Fetch xid just once - see GetNewTransactionId */
1645 0 : pxid = UINT32_ACCESS_ONCE(other_xids[i]);
1646 :
1647 0 : if (!TransactionIdIsValid(pxid))
1648 0 : continue;
1649 :
1650 0 : if (proc->pid == 0)
1651 0 : continue; /* ignore prepared transactions */
1652 :
1653 0 : if (TransactionIdEquals(pxid, xid))
1654 : {
1655 0 : result = true;
1656 0 : break;
1657 : }
1658 : }
1659 :
1660 0 : LWLockRelease(ProcArrayLock);
1661 :
1662 0 : return result;
1663 : }
1664 :
1665 :
1666 : /*
1667 : * Determine XID horizons.
1668 : *
1669 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1670 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1671 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1672 : * GlobalVisState).
1673 : *
1674 : * See the definition of ComputeXidHorizonsResult for the various computed
1675 : * horizons.
1676 : *
1677 : * For VACUUM separate horizons (used to decide which deleted tuples must
1678 : * be preserved), for shared and non-shared tables are computed. For shared
1679 : * relations backends in all databases must be considered, but for non-shared
1680 : * relations that's not required, since only backends in my own database could
1681 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1682 : * VACUUMs because (a) they must be working on other tables, and (b) they
1683 : * don't need to do snapshot-based lookups.
1684 : *
1685 : * This also computes a horizon used to truncate pg_subtrans. For that
1686 : * backends in all databases have to be considered, and concurrently running
1687 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1688 : * accesses.
1689 : *
1690 : * Note: we include all currently running xids in the set of considered xids.
1691 : * This ensures that if a just-started xact has not yet set its snapshot,
1692 : * when it does set the snapshot it cannot set xmin less than what we compute.
1693 : * See notes in src/backend/access/transam/README.
1694 : *
1695 : * Note: despite the above, it's possible for the calculated values to move
1696 : * backwards on repeated calls. The calculated values are conservative, so
1697 : * that anything older is definitely not considered as running by anyone
1698 : * anymore, but the exact values calculated depend on a number of things. For
1699 : * example, if there are no transactions running in the current database, the
1700 : * horizon for normal tables will be latestCompletedXid. If a transaction
1701 : * begins after that, its xmin will include in-progress transactions in other
1702 : * databases that started earlier, so another call will return a lower value.
1703 : * Nonetheless it is safe to vacuum a table in the current database with the
1704 : * first result. There are also replication-related effects: a walsender
1705 : * process can set its xmin based on transactions that are no longer running
1706 : * on the primary but are still being replayed on the standby, thus possibly
1707 : * making the values go backwards. In this case there is a possibility that
1708 : * we lose data that the standby would like to have, but unless the standby
1709 : * uses a replication slot to make its xmin persistent there is little we can
1710 : * do about that --- data is only protected if the walsender runs continuously
1711 : * while queries are executed on the standby. (The Hot Standby code deals
1712 : * with such cases by failing standby queries that needed to access
1713 : * already-removed data, so there's no integrity bug.)
1714 : *
1715 : * Note: the approximate horizons (see definition of GlobalVisState) are
1716 : * updated by the computations done here. That's currently required for
1717 : * correctness and a small optimization. Without doing so it's possible that
1718 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1719 : * horizon than later when deciding which tuples can be removed - which the
1720 : * code doesn't expect (breaking HOT).
1721 : */
1722 : static void
1723 173356 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1724 : {
1725 173356 : ProcArrayStruct *arrayP = procArray;
1726 : TransactionId kaxmin;
1727 173356 : bool in_recovery = RecoveryInProgress();
1728 173356 : TransactionId *other_xids = ProcGlobal->xids;
1729 :
1730 : /* inferred after ProcArrayLock is released */
1731 173356 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1732 :
1733 173356 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1734 :
1735 173356 : h->latest_completed = TransamVariables->latestCompletedXid;
1736 :
1737 : /*
1738 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1739 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1740 : * and so protects us against overestimating the result due to future
1741 : * additions.
1742 : */
1743 : {
1744 : TransactionId initial;
1745 :
1746 173356 : initial = XidFromFullTransactionId(h->latest_completed);
1747 : Assert(TransactionIdIsValid(initial));
1748 173356 : TransactionIdAdvance(initial);
1749 :
1750 173356 : h->oldest_considered_running = initial;
1751 173356 : h->shared_oldest_nonremovable = initial;
1752 173356 : h->data_oldest_nonremovable = initial;
1753 :
1754 : /*
1755 : * Only modifications made by this backend affect the horizon for
1756 : * temporary relations. Instead of a check in each iteration of the
1757 : * loop over all PGPROCs it is cheaper to just initialize to the
1758 : * current top-level xid any.
1759 : *
1760 : * Without an assigned xid we could use a horizon as aggressive as
1761 : * GetNewTransactionId(), but we can get away with the much cheaper
1762 : * latestCompletedXid + 1: If this backend has no xid there, by
1763 : * definition, can't be any newer changes in the temp table than
1764 : * latestCompletedXid.
1765 : */
1766 173356 : if (TransactionIdIsValid(MyProc->xid))
1767 60746 : h->temp_oldest_nonremovable = MyProc->xid;
1768 : else
1769 112610 : h->temp_oldest_nonremovable = initial;
1770 : }
1771 :
1772 : /*
1773 : * Fetch slot horizons while ProcArrayLock is held - the
1774 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1775 : * the lock.
1776 : */
1777 173356 : h->slot_xmin = procArray->replication_slot_xmin;
1778 173356 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1779 :
1780 975020 : for (int index = 0; index < arrayP->numProcs; index++)
1781 : {
1782 801664 : int pgprocno = arrayP->pgprocnos[index];
1783 801664 : PGPROC *proc = &allProcs[pgprocno];
1784 801664 : int8 statusFlags = ProcGlobal->statusFlags[index];
1785 : TransactionId xid;
1786 : TransactionId xmin;
1787 :
1788 : /* Fetch xid just once - see GetNewTransactionId */
1789 801664 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1790 801664 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1791 :
1792 : /*
1793 : * Consider both the transaction's Xmin, and its Xid.
1794 : *
1795 : * We must check both because a transaction might have an Xmin but not
1796 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1797 : * some not-yet-set Xmin.
1798 : */
1799 801664 : xmin = TransactionIdOlder(xmin, xid);
1800 :
1801 : /* if neither is set, this proc doesn't influence the horizon */
1802 801664 : if (!TransactionIdIsValid(xmin))
1803 431198 : continue;
1804 :
1805 : /*
1806 : * Don't ignore any procs when determining which transactions might be
1807 : * considered running. While slots should ensure logical decoding
1808 : * backends are protected even without this check, it can't hurt to
1809 : * include them here as well..
1810 : */
1811 370466 : h->oldest_considered_running =
1812 370466 : TransactionIdOlder(h->oldest_considered_running, xmin);
1813 :
1814 : /*
1815 : * Skip over backends either vacuuming (which is ok with rows being
1816 : * removed, as long as pg_subtrans is not truncated) or doing logical
1817 : * decoding (which manages xmin separately, check below).
1818 : */
1819 370466 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1820 21038 : continue;
1821 :
1822 : /* shared tables need to take backends in all databases into account */
1823 349428 : h->shared_oldest_nonremovable =
1824 349428 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1825 :
1826 : /*
1827 : * Normally sessions in other databases are ignored for anything but
1828 : * the shared horizon.
1829 : *
1830 : * However, include them when MyDatabaseId is not (yet) set. A
1831 : * backend in the process of starting up must not compute a "too
1832 : * aggressive" horizon, otherwise we could end up using it to prune
1833 : * still-needed data away. If the current backend never connects to a
1834 : * database this is harmless, because data_oldest_nonremovable will
1835 : * never be utilized.
1836 : *
1837 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1838 : * be included. (This flag is used for hot standby feedback, which
1839 : * can't be tied to a specific database.)
1840 : *
1841 : * Also, while in recovery we cannot compute an accurate per-database
1842 : * horizon, as all xids are managed via the KnownAssignedXids
1843 : * machinery.
1844 : */
1845 349428 : if (proc->databaseId == MyDatabaseId ||
1846 9400 : MyDatabaseId == InvalidOid ||
1847 610 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1848 : in_recovery)
1849 : {
1850 348818 : h->data_oldest_nonremovable =
1851 348818 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1852 : }
1853 : }
1854 :
1855 : /*
1856 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1857 : * after lock is released.
1858 : */
1859 173356 : if (in_recovery)
1860 356 : kaxmin = KnownAssignedXidsGetOldestXmin();
1861 :
1862 : /*
1863 : * No other information from shared state is needed, release the lock
1864 : * immediately. The rest of the computations can be done without a lock.
1865 : */
1866 173356 : LWLockRelease(ProcArrayLock);
1867 :
1868 173356 : if (in_recovery)
1869 : {
1870 356 : h->oldest_considered_running =
1871 356 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1872 356 : h->shared_oldest_nonremovable =
1873 356 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1874 356 : h->data_oldest_nonremovable =
1875 356 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1876 : /* temp relations cannot be accessed in recovery */
1877 : }
1878 :
1879 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1880 : h->shared_oldest_nonremovable));
1881 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1882 : h->data_oldest_nonremovable));
1883 :
1884 : /*
1885 : * Check whether there are replication slots requiring an older xmin.
1886 : */
1887 173356 : h->shared_oldest_nonremovable =
1888 173356 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1889 173356 : h->data_oldest_nonremovable =
1890 173356 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1891 :
1892 : /*
1893 : * The only difference between catalog / data horizons is that the slot's
1894 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1895 : * for logical decoding). Initialize with data horizon, and then back up
1896 : * further if necessary. Have to back up the shared horizon as well, since
1897 : * that also can contain catalogs.
1898 : */
1899 173356 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1900 173356 : h->shared_oldest_nonremovable =
1901 173356 : TransactionIdOlder(h->shared_oldest_nonremovable,
1902 : h->slot_catalog_xmin);
1903 173356 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1904 173356 : h->catalog_oldest_nonremovable =
1905 173356 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1906 : h->slot_catalog_xmin);
1907 :
1908 : /*
1909 : * It's possible that slots backed up the horizons further than
1910 : * oldest_considered_running. Fix.
1911 : */
1912 173356 : h->oldest_considered_running =
1913 173356 : TransactionIdOlder(h->oldest_considered_running,
1914 : h->shared_oldest_nonremovable);
1915 173356 : h->oldest_considered_running =
1916 173356 : TransactionIdOlder(h->oldest_considered_running,
1917 : h->catalog_oldest_nonremovable);
1918 173356 : h->oldest_considered_running =
1919 173356 : TransactionIdOlder(h->oldest_considered_running,
1920 : h->data_oldest_nonremovable);
1921 :
1922 : /*
1923 : * shared horizons have to be at least as old as the oldest visible in
1924 : * current db
1925 : */
1926 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1927 : h->data_oldest_nonremovable));
1928 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1929 : h->catalog_oldest_nonremovable));
1930 :
1931 : /*
1932 : * Horizons need to ensure that pg_subtrans access is still possible for
1933 : * the relevant backends.
1934 : */
1935 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1936 : h->shared_oldest_nonremovable));
1937 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1938 : h->catalog_oldest_nonremovable));
1939 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1940 : h->data_oldest_nonremovable));
1941 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1942 : h->temp_oldest_nonremovable));
1943 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1944 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1945 : h->slot_xmin));
1946 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1947 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1948 : h->slot_catalog_xmin));
1949 :
1950 : /* update approximate horizons with the computed horizons */
1951 173356 : GlobalVisUpdateApply(h);
1952 173356 : }
1953 :
1954 : /*
1955 : * Determine what kind of visibility horizon needs to be used for a
1956 : * relation. If rel is NULL, the most conservative horizon is used.
1957 : */
1958 : static inline GlobalVisHorizonKind
1959 27536300 : GlobalVisHorizonKindForRel(Relation rel)
1960 : {
1961 : /*
1962 : * Other relkinds currently don't contain xids, nor always the necessary
1963 : * logical decoding markers.
1964 : */
1965 : Assert(!rel ||
1966 : rel->rd_rel->relkind == RELKIND_RELATION ||
1967 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1968 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1969 :
1970 27536300 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1971 107796 : return VISHORIZON_SHARED;
1972 27428504 : else if (IsCatalogRelation(rel) ||
1973 22685038 : RelationIsAccessibleInLogicalDecoding(rel))
1974 4743474 : return VISHORIZON_CATALOG;
1975 22685030 : else if (!RELATION_IS_LOCAL(rel))
1976 22578582 : return VISHORIZON_DATA;
1977 : else
1978 106448 : return VISHORIZON_TEMP;
1979 : }
1980 :
1981 : /*
1982 : * Return the oldest XID for which deleted tuples must be preserved in the
1983 : * passed table.
1984 : *
1985 : * If rel is not NULL the horizon may be considerably more recent than
1986 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1987 : * that is correct (but not optimal) for all relations will be returned.
1988 : *
1989 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1990 : * the passed in table.
1991 : */
1992 : TransactionId
1993 72324 : GetOldestNonRemovableTransactionId(Relation rel)
1994 : {
1995 : ComputeXidHorizonsResult horizons;
1996 :
1997 72324 : ComputeXidHorizons(&horizons);
1998 :
1999 72324 : switch (GlobalVisHorizonKindForRel(rel))
2000 : {
2001 6754 : case VISHORIZON_SHARED:
2002 6754 : return horizons.shared_oldest_nonremovable;
2003 21448 : case VISHORIZON_CATALOG:
2004 21448 : return horizons.catalog_oldest_nonremovable;
2005 21316 : case VISHORIZON_DATA:
2006 21316 : return horizons.data_oldest_nonremovable;
2007 22806 : case VISHORIZON_TEMP:
2008 22806 : return horizons.temp_oldest_nonremovable;
2009 : }
2010 :
2011 : /* just to prevent compiler warnings */
2012 0 : return InvalidTransactionId;
2013 : }
2014 :
2015 : /*
2016 : * Return the oldest transaction id any currently running backend might still
2017 : * consider running. This should not be used for visibility / pruning
2018 : * determinations (see GetOldestNonRemovableTransactionId()), but for
2019 : * decisions like up to where pg_subtrans can be truncated.
2020 : */
2021 : TransactionId
2022 1648 : GetOldestTransactionIdConsideredRunning(void)
2023 : {
2024 : ComputeXidHorizonsResult horizons;
2025 :
2026 1648 : ComputeXidHorizons(&horizons);
2027 :
2028 1648 : return horizons.oldest_considered_running;
2029 : }
2030 :
2031 : /*
2032 : * Return the visibility horizons for a hot standby feedback message.
2033 : */
2034 : void
2035 64 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
2036 : {
2037 : ComputeXidHorizonsResult horizons;
2038 :
2039 64 : ComputeXidHorizons(&horizons);
2040 :
2041 : /*
2042 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2043 : * effect of replication slot's catalog_xmin. We want to send a separate
2044 : * feedback for the catalog horizon, so the primary can remove data table
2045 : * contents more aggressively.
2046 : */
2047 64 : *xmin = horizons.shared_oldest_nonremovable_raw;
2048 64 : *catalog_xmin = horizons.slot_catalog_xmin;
2049 64 : }
2050 :
2051 : /*
2052 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2053 : *
2054 : * We have to export this for use by snapmgr.c.
2055 : */
2056 : int
2057 52698 : GetMaxSnapshotXidCount(void)
2058 : {
2059 52698 : return procArray->maxProcs;
2060 : }
2061 :
2062 : /*
2063 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2064 : *
2065 : * We have to export this for use by snapmgr.c.
2066 : */
2067 : int
2068 52372 : GetMaxSnapshotSubxidCount(void)
2069 : {
2070 52372 : return TOTAL_MAX_CACHED_SUBXIDS;
2071 : }
2072 :
2073 : /*
2074 : * Helper function for GetSnapshotData() that checks if the bulk of the
2075 : * visibility information in the snapshot is still valid. If so, it updates
2076 : * the fields that need to change and returns true. Otherwise it returns
2077 : * false.
2078 : *
2079 : * This very likely can be evolved to not need ProcArrayLock held (at very
2080 : * least in the case we already hold a snapshot), but that's for another day.
2081 : */
2082 : static bool
2083 3209366 : GetSnapshotDataReuse(Snapshot snapshot)
2084 : {
2085 : uint64 curXactCompletionCount;
2086 :
2087 : Assert(LWLockHeldByMe(ProcArrayLock));
2088 :
2089 3209366 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2090 52330 : return false;
2091 :
2092 3157036 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2093 3157036 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2094 602346 : return false;
2095 :
2096 : /*
2097 : * If the current xactCompletionCount is still the same as it was at the
2098 : * time the snapshot was built, we can be sure that rebuilding the
2099 : * contents of the snapshot the hard way would result in the same snapshot
2100 : * contents:
2101 : *
2102 : * As explained in transam/README, the set of xids considered running by
2103 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2104 : * contents only depend on transactions with xids and xactCompletionCount
2105 : * is incremented whenever a transaction with an xid finishes (while
2106 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2107 : * ensures we would detect if the snapshot would have changed.
2108 : *
2109 : * As the snapshot contents are the same as it was before, it is safe to
2110 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2111 : * visible under the snapshot could already have been removed (that'd
2112 : * require the set of running transactions to change) and it fulfills the
2113 : * requirement that concurrent GetSnapshotData() calls yield the same
2114 : * xmin.
2115 : */
2116 2554690 : if (!TransactionIdIsValid(MyProc->xmin))
2117 907774 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2118 :
2119 2554690 : RecentXmin = snapshot->xmin;
2120 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2121 :
2122 2554690 : snapshot->curcid = GetCurrentCommandId(false);
2123 2554690 : snapshot->active_count = 0;
2124 2554690 : snapshot->regd_count = 0;
2125 2554690 : snapshot->copied = false;
2126 2554690 : snapshot->lsn = InvalidXLogRecPtr;
2127 2554690 : snapshot->whenTaken = 0;
2128 :
2129 2554690 : return true;
2130 : }
2131 :
2132 : /*
2133 : * GetSnapshotData -- returns information about running transactions.
2134 : *
2135 : * The returned snapshot includes xmin (lowest still-running xact ID),
2136 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2137 : * in the range xmin <= xid < xmax. It is used as follows:
2138 : * All xact IDs < xmin are considered finished.
2139 : * All xact IDs >= xmax are considered still running.
2140 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2141 : * it is considered running or not.
2142 : * This ensures that the set of transactions seen as "running" by the
2143 : * current xact will not change after it takes the snapshot.
2144 : *
2145 : * All running top-level XIDs are included in the snapshot, except for lazy
2146 : * VACUUM processes. We also try to include running subtransaction XIDs,
2147 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2148 : * information may not be available. If we find any overflowed subxid arrays,
2149 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2150 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2151 : *
2152 : * We also update the following backend-global variables:
2153 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2154 : * current transaction (this is the same as MyProc->xmin).
2155 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2156 : * older than this are known not running any more.
2157 : *
2158 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2159 : * for the benefit of the GlobalVisTest* family of functions.
2160 : *
2161 : * Note: this function should probably not be called with an argument that's
2162 : * not statically allocated (see xip allocation below).
2163 : */
2164 : Snapshot
2165 3209366 : GetSnapshotData(Snapshot snapshot)
2166 : {
2167 3209366 : ProcArrayStruct *arrayP = procArray;
2168 3209366 : TransactionId *other_xids = ProcGlobal->xids;
2169 : TransactionId xmin;
2170 : TransactionId xmax;
2171 3209366 : int count = 0;
2172 3209366 : int subcount = 0;
2173 3209366 : bool suboverflowed = false;
2174 : FullTransactionId latest_completed;
2175 : TransactionId oldestxid;
2176 : int mypgxactoff;
2177 : TransactionId myxid;
2178 : uint64 curXactCompletionCount;
2179 :
2180 3209366 : TransactionId replication_slot_xmin = InvalidTransactionId;
2181 3209366 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2182 :
2183 : Assert(snapshot != NULL);
2184 :
2185 : /*
2186 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2187 : * be sufficient. But it seems better to do the malloc while not holding
2188 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2189 : * more subxip storage than is probably needed.
2190 : *
2191 : * This does open a possibility for avoiding repeated malloc/free: since
2192 : * maxProcs does not change at runtime, we can simply reuse the previous
2193 : * xip arrays if any. (This relies on the fact that all callers pass
2194 : * static SnapshotData structs.)
2195 : */
2196 3209366 : if (snapshot->xip == NULL)
2197 : {
2198 : /*
2199 : * First call for this snapshot. Snapshot is same size whether or not
2200 : * we are in recovery, see later comments.
2201 : */
2202 52324 : snapshot->xip = (TransactionId *)
2203 52324 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2204 52324 : if (snapshot->xip == NULL)
2205 0 : ereport(ERROR,
2206 : (errcode(ERRCODE_OUT_OF_MEMORY),
2207 : errmsg("out of memory")));
2208 : Assert(snapshot->subxip == NULL);
2209 52324 : snapshot->subxip = (TransactionId *)
2210 52324 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2211 52324 : if (snapshot->subxip == NULL)
2212 0 : ereport(ERROR,
2213 : (errcode(ERRCODE_OUT_OF_MEMORY),
2214 : errmsg("out of memory")));
2215 : }
2216 :
2217 : /*
2218 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2219 : * going to set MyProc->xmin.
2220 : */
2221 3209366 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2222 :
2223 3209366 : if (GetSnapshotDataReuse(snapshot))
2224 : {
2225 2554690 : LWLockRelease(ProcArrayLock);
2226 2554690 : return snapshot;
2227 : }
2228 :
2229 654676 : latest_completed = TransamVariables->latestCompletedXid;
2230 654676 : mypgxactoff = MyProc->pgxactoff;
2231 654676 : myxid = other_xids[mypgxactoff];
2232 : Assert(myxid == MyProc->xid);
2233 :
2234 654676 : oldestxid = TransamVariables->oldestXid;
2235 654676 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2236 :
2237 : /* xmax is always latestCompletedXid + 1 */
2238 654676 : xmax = XidFromFullTransactionId(latest_completed);
2239 654676 : TransactionIdAdvance(xmax);
2240 : Assert(TransactionIdIsNormal(xmax));
2241 :
2242 : /* initialize xmin calculation with xmax */
2243 654676 : xmin = xmax;
2244 :
2245 : /* take own xid into account, saves a check inside the loop */
2246 654676 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2247 34368 : xmin = myxid;
2248 :
2249 654676 : snapshot->takenDuringRecovery = RecoveryInProgress();
2250 :
2251 654676 : if (!snapshot->takenDuringRecovery)
2252 : {
2253 652688 : int numProcs = arrayP->numProcs;
2254 652688 : TransactionId *xip = snapshot->xip;
2255 652688 : int *pgprocnos = arrayP->pgprocnos;
2256 652688 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2257 652688 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2258 :
2259 : /*
2260 : * First collect set of pgxactoff/xids that need to be included in the
2261 : * snapshot.
2262 : */
2263 4769946 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2264 : {
2265 : /* Fetch xid just once - see GetNewTransactionId */
2266 4117258 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2267 : uint8 statusFlags;
2268 :
2269 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2270 :
2271 : /*
2272 : * If the transaction has no XID assigned, we can skip it; it
2273 : * won't have sub-XIDs either.
2274 : */
2275 4117258 : if (likely(xid == InvalidTransactionId))
2276 3418534 : continue;
2277 :
2278 : /*
2279 : * We don't include our own XIDs (if any) in the snapshot. It
2280 : * needs to be included in the xmin computation, but we did so
2281 : * outside the loop.
2282 : */
2283 698724 : if (pgxactoff == mypgxactoff)
2284 72328 : continue;
2285 :
2286 : /*
2287 : * The only way we are able to get here with a non-normal xid is
2288 : * during bootstrap - with this backend using
2289 : * BootstrapTransactionId. But the above test should filter that
2290 : * out.
2291 : */
2292 : Assert(TransactionIdIsNormal(xid));
2293 :
2294 : /*
2295 : * If the XID is >= xmax, we can skip it; such transactions will
2296 : * be treated as running anyway (and any sub-XIDs will also be >=
2297 : * xmax).
2298 : */
2299 626396 : if (!NormalTransactionIdPrecedes(xid, xmax))
2300 143394 : continue;
2301 :
2302 : /*
2303 : * Skip over backends doing logical decoding which manages xmin
2304 : * separately (check below) and ones running LAZY VACUUM.
2305 : */
2306 483002 : statusFlags = allStatusFlags[pgxactoff];
2307 483002 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2308 76 : continue;
2309 :
2310 482926 : if (NormalTransactionIdPrecedes(xid, xmin))
2311 348360 : xmin = xid;
2312 :
2313 : /* Add XID to snapshot. */
2314 482926 : xip[count++] = xid;
2315 :
2316 : /*
2317 : * Save subtransaction XIDs if possible (if we've already
2318 : * overflowed, there's no point). Note that the subxact XIDs must
2319 : * be later than their parent, so no need to check them against
2320 : * xmin. We could filter against xmax, but it seems better not to
2321 : * do that much work while holding the ProcArrayLock.
2322 : *
2323 : * The other backend can add more subxids concurrently, but cannot
2324 : * remove any. Hence it's important to fetch nxids just once.
2325 : * Should be safe to use memcpy, though. (We needn't worry about
2326 : * missing any xids added concurrently, because they must postdate
2327 : * xmax.)
2328 : *
2329 : * Again, our own XIDs are not included in the snapshot.
2330 : */
2331 482926 : if (!suboverflowed)
2332 : {
2333 :
2334 482918 : if (subxidStates[pgxactoff].overflowed)
2335 58 : suboverflowed = true;
2336 : else
2337 : {
2338 482860 : int nsubxids = subxidStates[pgxactoff].count;
2339 :
2340 482860 : if (nsubxids > 0)
2341 : {
2342 6980 : int pgprocno = pgprocnos[pgxactoff];
2343 6980 : PGPROC *proc = &allProcs[pgprocno];
2344 :
2345 6980 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2346 :
2347 6980 : memcpy(snapshot->subxip + subcount,
2348 6980 : proc->subxids.xids,
2349 : nsubxids * sizeof(TransactionId));
2350 6980 : subcount += nsubxids;
2351 : }
2352 : }
2353 : }
2354 : }
2355 : }
2356 : else
2357 : {
2358 : /*
2359 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2360 : *
2361 : * We store all xids directly into subxip[]. Here's why:
2362 : *
2363 : * In recovery we don't know which xids are top-level and which are
2364 : * subxacts, a design choice that greatly simplifies xid processing.
2365 : *
2366 : * It seems like we would want to try to put xids into xip[] only, but
2367 : * that is fairly small. We would either need to make that bigger or
2368 : * to increase the rate at which we WAL-log xid assignment; neither is
2369 : * an appealing choice.
2370 : *
2371 : * We could try to store xids into xip[] first and then into subxip[]
2372 : * if there are too many xids. That only works if the snapshot doesn't
2373 : * overflow because we do not search subxip[] in that case. A simpler
2374 : * way is to just store all xids in the subxip array because this is
2375 : * by far the bigger array. We just leave the xip array empty.
2376 : *
2377 : * Either way we need to change the way XidInMVCCSnapshot() works
2378 : * depending upon when the snapshot was taken, or change normal
2379 : * snapshot processing so it matches.
2380 : *
2381 : * Note: It is possible for recovery to end before we finish taking
2382 : * the snapshot, and for newly assigned transaction ids to be added to
2383 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2384 : * those newly added transaction ids would be filtered away, so we
2385 : * need not be concerned about them.
2386 : */
2387 1988 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2388 : xmax);
2389 :
2390 1988 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2391 8 : suboverflowed = true;
2392 : }
2393 :
2394 :
2395 : /*
2396 : * Fetch into local variable while ProcArrayLock is held - the
2397 : * LWLockRelease below is a barrier, ensuring this happens inside the
2398 : * lock.
2399 : */
2400 654676 : replication_slot_xmin = procArray->replication_slot_xmin;
2401 654676 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2402 :
2403 654676 : if (!TransactionIdIsValid(MyProc->xmin))
2404 338906 : MyProc->xmin = TransactionXmin = xmin;
2405 :
2406 654676 : LWLockRelease(ProcArrayLock);
2407 :
2408 : /* maintain state for GlobalVis* */
2409 : {
2410 : TransactionId def_vis_xid;
2411 : TransactionId def_vis_xid_data;
2412 : FullTransactionId def_vis_fxid;
2413 : FullTransactionId def_vis_fxid_data;
2414 : FullTransactionId oldestfxid;
2415 :
2416 : /*
2417 : * Converting oldestXid is only safe when xid horizon cannot advance,
2418 : * i.e. holding locks. While we don't hold the lock anymore, all the
2419 : * necessary data has been gathered with lock held.
2420 : */
2421 654676 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2422 :
2423 : /* Check whether there's a replication slot requiring an older xmin. */
2424 : def_vis_xid_data =
2425 654676 : TransactionIdOlder(xmin, replication_slot_xmin);
2426 :
2427 : /*
2428 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2429 : * if older than this xid.
2430 : */
2431 654676 : def_vis_xid = def_vis_xid_data;
2432 :
2433 : /*
2434 : * Check whether there's a replication slot requiring an older catalog
2435 : * xmin.
2436 : */
2437 : def_vis_xid =
2438 654676 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2439 :
2440 654676 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2441 654676 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2442 :
2443 : /*
2444 : * Check if we can increase upper bound. As a previous
2445 : * GlobalVisUpdate() might have computed more aggressive values, don't
2446 : * overwrite them if so.
2447 : */
2448 : GlobalVisSharedRels.definitely_needed =
2449 654676 : FullTransactionIdNewer(def_vis_fxid,
2450 : GlobalVisSharedRels.definitely_needed);
2451 : GlobalVisCatalogRels.definitely_needed =
2452 654676 : FullTransactionIdNewer(def_vis_fxid,
2453 : GlobalVisCatalogRels.definitely_needed);
2454 : GlobalVisDataRels.definitely_needed =
2455 654676 : FullTransactionIdNewer(def_vis_fxid_data,
2456 : GlobalVisDataRels.definitely_needed);
2457 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2458 654676 : if (TransactionIdIsNormal(myxid))
2459 : GlobalVisTempRels.definitely_needed =
2460 72168 : FullXidRelativeTo(latest_completed, myxid);
2461 : else
2462 : {
2463 582508 : GlobalVisTempRels.definitely_needed = latest_completed;
2464 582508 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2465 : }
2466 :
2467 : /*
2468 : * Check if we know that we can initialize or increase the lower
2469 : * bound. Currently the only cheap way to do so is to use
2470 : * TransamVariables->oldestXid as input.
2471 : *
2472 : * We should definitely be able to do better. We could e.g. put a
2473 : * global lower bound value into TransamVariables.
2474 : */
2475 : GlobalVisSharedRels.maybe_needed =
2476 654676 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2477 : oldestfxid);
2478 : GlobalVisCatalogRels.maybe_needed =
2479 654676 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2480 : oldestfxid);
2481 : GlobalVisDataRels.maybe_needed =
2482 654676 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2483 : oldestfxid);
2484 : /* accurate value known */
2485 654676 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2486 : }
2487 :
2488 654676 : RecentXmin = xmin;
2489 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2490 :
2491 654676 : snapshot->xmin = xmin;
2492 654676 : snapshot->xmax = xmax;
2493 654676 : snapshot->xcnt = count;
2494 654676 : snapshot->subxcnt = subcount;
2495 654676 : snapshot->suboverflowed = suboverflowed;
2496 654676 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2497 :
2498 654676 : snapshot->curcid = GetCurrentCommandId(false);
2499 :
2500 : /*
2501 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2502 : * not copied in persistent memory.
2503 : */
2504 654676 : snapshot->active_count = 0;
2505 654676 : snapshot->regd_count = 0;
2506 654676 : snapshot->copied = false;
2507 654676 : snapshot->lsn = InvalidXLogRecPtr;
2508 654676 : snapshot->whenTaken = 0;
2509 :
2510 654676 : return snapshot;
2511 : }
2512 :
2513 : /*
2514 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2515 : *
2516 : * This is called when installing a snapshot imported from another
2517 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2518 : * check that the source transaction is still running, and we'd better do
2519 : * that atomically with installing the new xmin.
2520 : *
2521 : * Returns true if successful, false if source xact is no longer running.
2522 : */
2523 : bool
2524 32 : ProcArrayInstallImportedXmin(TransactionId xmin,
2525 : VirtualTransactionId *sourcevxid)
2526 : {
2527 32 : bool result = false;
2528 32 : ProcArrayStruct *arrayP = procArray;
2529 : int index;
2530 :
2531 : Assert(TransactionIdIsNormal(xmin));
2532 32 : if (!sourcevxid)
2533 0 : return false;
2534 :
2535 : /* Get lock so source xact can't end while we're doing this */
2536 32 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2537 :
2538 : /*
2539 : * Find the PGPROC entry of the source transaction. (This could use
2540 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2541 : * performance critical.)
2542 : */
2543 32 : for (index = 0; index < arrayP->numProcs; index++)
2544 : {
2545 32 : int pgprocno = arrayP->pgprocnos[index];
2546 32 : PGPROC *proc = &allProcs[pgprocno];
2547 32 : int statusFlags = ProcGlobal->statusFlags[index];
2548 : TransactionId xid;
2549 :
2550 : /* Ignore procs running LAZY VACUUM */
2551 32 : if (statusFlags & PROC_IN_VACUUM)
2552 0 : continue;
2553 :
2554 : /* We are only interested in the specific virtual transaction. */
2555 32 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2556 0 : continue;
2557 32 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2558 0 : continue;
2559 :
2560 : /*
2561 : * We check the transaction's database ID for paranoia's sake: if it's
2562 : * in another DB then its xmin does not cover us. Caller should have
2563 : * detected this already, so we just treat any funny cases as
2564 : * "transaction not found".
2565 : */
2566 32 : if (proc->databaseId != MyDatabaseId)
2567 0 : continue;
2568 :
2569 : /*
2570 : * Likewise, let's just make real sure its xmin does cover us.
2571 : */
2572 32 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2573 32 : if (!TransactionIdIsNormal(xid) ||
2574 32 : !TransactionIdPrecedesOrEquals(xid, xmin))
2575 0 : continue;
2576 :
2577 : /*
2578 : * We're good. Install the new xmin. As in GetSnapshotData, set
2579 : * TransactionXmin too. (Note that because snapmgr.c called
2580 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2581 : * we don't check that.)
2582 : */
2583 32 : MyProc->xmin = TransactionXmin = xmin;
2584 :
2585 32 : result = true;
2586 32 : break;
2587 : }
2588 :
2589 32 : LWLockRelease(ProcArrayLock);
2590 :
2591 32 : return result;
2592 : }
2593 :
2594 : /*
2595 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2596 : *
2597 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2598 : * PGPROC of the transaction from which we imported the snapshot, rather than
2599 : * an XID.
2600 : *
2601 : * Note that this function also copies statusFlags from the source `proc` in
2602 : * order to avoid the case where MyProc's xmin needs to be skipped for
2603 : * computing xid horizon.
2604 : *
2605 : * Returns true if successful, false if source xact is no longer running.
2606 : */
2607 : bool
2608 2982 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2609 : {
2610 2982 : bool result = false;
2611 : TransactionId xid;
2612 :
2613 : Assert(TransactionIdIsNormal(xmin));
2614 : Assert(proc != NULL);
2615 :
2616 : /*
2617 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2618 : */
2619 2982 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2620 :
2621 : /*
2622 : * Be certain that the referenced PGPROC has an advertised xmin which is
2623 : * no later than the one we're installing, so that the system-wide xmin
2624 : * can't go backwards. Also, make sure it's running in the same database,
2625 : * so that the per-database xmin cannot go backwards.
2626 : */
2627 2982 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2628 2982 : if (proc->databaseId == MyDatabaseId &&
2629 2982 : TransactionIdIsNormal(xid) &&
2630 2982 : TransactionIdPrecedesOrEquals(xid, xmin))
2631 : {
2632 : /*
2633 : * Install xmin and propagate the statusFlags that affect how the
2634 : * value is interpreted by vacuum.
2635 : */
2636 2982 : MyProc->xmin = TransactionXmin = xmin;
2637 2982 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2638 2982 : (proc->statusFlags & PROC_XMIN_FLAGS);
2639 2982 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2640 :
2641 2982 : result = true;
2642 : }
2643 :
2644 2982 : LWLockRelease(ProcArrayLock);
2645 :
2646 2982 : return result;
2647 : }
2648 :
2649 : /*
2650 : * GetRunningTransactionData -- returns information about running transactions.
2651 : *
2652 : * Similar to GetSnapshotData but returns more information. We include
2653 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2654 : * prepared transactions.
2655 : *
2656 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2657 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2658 : * array until the caller has WAL-logged this snapshot, and releases the
2659 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2660 : * lock is released.
2661 : *
2662 : * The returned data structure is statically allocated; caller should not
2663 : * modify it, and must not assume it is valid past the next call.
2664 : *
2665 : * This is never executed during recovery so there is no need to look at
2666 : * KnownAssignedXids.
2667 : *
2668 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2669 : * may return entries with duplicated TransactionId values coming from
2670 : * transaction finishing to prepare. Nothing is done about duplicated
2671 : * entries here to not hold on ProcArrayLock more than necessary.
2672 : *
2673 : * We don't worry about updating other counters, we want to keep this as
2674 : * simple as possible and leave GetSnapshotData() as the primary code for
2675 : * that bookkeeping.
2676 : *
2677 : * Note that if any transaction has overflowed its cached subtransactions
2678 : * then there is no real need include any subtransactions.
2679 : */
2680 : RunningTransactions
2681 1484 : GetRunningTransactionData(void)
2682 : {
2683 : /* result workspace */
2684 : static RunningTransactionsData CurrentRunningXactsData;
2685 :
2686 1484 : ProcArrayStruct *arrayP = procArray;
2687 1484 : TransactionId *other_xids = ProcGlobal->xids;
2688 1484 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2689 : TransactionId latestCompletedXid;
2690 : TransactionId oldestRunningXid;
2691 : TransactionId oldestDatabaseRunningXid;
2692 : TransactionId *xids;
2693 : int index;
2694 : int count;
2695 : int subcount;
2696 : bool suboverflowed;
2697 :
2698 : Assert(!RecoveryInProgress());
2699 :
2700 : /*
2701 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2702 : * be sufficient. But it seems better to do the malloc while not holding
2703 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2704 : * more subxip storage than is probably needed.
2705 : *
2706 : * Should only be allocated in bgwriter, since only ever executed during
2707 : * checkpoints.
2708 : */
2709 1484 : if (CurrentRunningXacts->xids == NULL)
2710 : {
2711 : /*
2712 : * First call
2713 : */
2714 892 : CurrentRunningXacts->xids = (TransactionId *)
2715 892 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2716 892 : if (CurrentRunningXacts->xids == NULL)
2717 0 : ereport(ERROR,
2718 : (errcode(ERRCODE_OUT_OF_MEMORY),
2719 : errmsg("out of memory")));
2720 : }
2721 :
2722 1484 : xids = CurrentRunningXacts->xids;
2723 :
2724 1484 : count = subcount = 0;
2725 1484 : suboverflowed = false;
2726 :
2727 : /*
2728 : * Ensure that no xids enter or leave the procarray while we obtain
2729 : * snapshot.
2730 : */
2731 1484 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2732 1484 : LWLockAcquire(XidGenLock, LW_SHARED);
2733 :
2734 1484 : latestCompletedXid =
2735 1484 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2736 1484 : oldestDatabaseRunningXid = oldestRunningXid =
2737 1484 : XidFromFullTransactionId(TransamVariables->nextXid);
2738 :
2739 : /*
2740 : * Spin over procArray collecting all xids
2741 : */
2742 6416 : for (index = 0; index < arrayP->numProcs; index++)
2743 : {
2744 4932 : int pgprocno = arrayP->pgprocnos[index];
2745 4932 : PGPROC *proc = &allProcs[pgprocno];
2746 : TransactionId xid;
2747 :
2748 : /* Fetch xid just once - see GetNewTransactionId */
2749 4932 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2750 :
2751 : /*
2752 : * We don't need to store transactions that don't have a TransactionId
2753 : * yet because they will not show as running on a standby server.
2754 : */
2755 4932 : if (!TransactionIdIsValid(xid))
2756 4414 : continue;
2757 :
2758 : /*
2759 : * Be careful not to exclude any xids before calculating the values of
2760 : * oldestRunningXid and suboverflowed, since these are used to clean
2761 : * up transaction information held on standbys.
2762 : */
2763 518 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2764 484 : oldestRunningXid = xid;
2765 :
2766 : /*
2767 : * Also, update the oldest running xid within the current database.
2768 : */
2769 854 : if (proc->databaseId == MyDatabaseId &&
2770 336 : TransactionIdPrecedes(xid, oldestRunningXid))
2771 0 : oldestDatabaseRunningXid = xid;
2772 :
2773 518 : if (ProcGlobal->subxidStates[index].overflowed)
2774 4 : suboverflowed = true;
2775 :
2776 : /*
2777 : * If we wished to exclude xids this would be the right place for it.
2778 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2779 : * but they do during truncation at the end when they get the lock and
2780 : * truncate, so it is not much of a problem to include them if they
2781 : * are seen and it is cleaner to include them.
2782 : */
2783 :
2784 518 : xids[count++] = xid;
2785 : }
2786 :
2787 : /*
2788 : * Spin over procArray collecting all subxids, but only if there hasn't
2789 : * been a suboverflow.
2790 : */
2791 1484 : if (!suboverflowed)
2792 : {
2793 1480 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2794 :
2795 6404 : for (index = 0; index < arrayP->numProcs; index++)
2796 : {
2797 4924 : int pgprocno = arrayP->pgprocnos[index];
2798 4924 : PGPROC *proc = &allProcs[pgprocno];
2799 : int nsubxids;
2800 :
2801 : /*
2802 : * Save subtransaction XIDs. Other backends can't add or remove
2803 : * entries while we're holding XidGenLock.
2804 : */
2805 4924 : nsubxids = other_subxidstates[index].count;
2806 4924 : if (nsubxids > 0)
2807 : {
2808 : /* barrier not really required, as XidGenLock is held, but ... */
2809 16 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2810 :
2811 16 : memcpy(&xids[count], proc->subxids.xids,
2812 : nsubxids * sizeof(TransactionId));
2813 16 : count += nsubxids;
2814 16 : subcount += nsubxids;
2815 :
2816 : /*
2817 : * Top-level XID of a transaction is always less than any of
2818 : * its subxids, so we don't need to check if any of the
2819 : * subxids are smaller than oldestRunningXid
2820 : */
2821 : }
2822 : }
2823 : }
2824 :
2825 : /*
2826 : * It's important *not* to include the limits set by slots here because
2827 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2828 : * were to be included here the initial value could never increase because
2829 : * of a circular dependency where slots only increase their limits when
2830 : * running xacts increases oldestRunningXid and running xacts only
2831 : * increases if slots do.
2832 : */
2833 :
2834 1484 : CurrentRunningXacts->xcnt = count - subcount;
2835 1484 : CurrentRunningXacts->subxcnt = subcount;
2836 1484 : CurrentRunningXacts->subxid_overflow = suboverflowed;
2837 1484 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2838 1484 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2839 1484 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2840 1484 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2841 :
2842 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2843 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2844 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2845 :
2846 : /* We don't release the locks here, the caller is responsible for that */
2847 :
2848 1484 : return CurrentRunningXacts;
2849 : }
2850 :
2851 : /*
2852 : * GetOldestActiveTransactionId()
2853 : *
2854 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2855 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2856 : * We look at all databases, though there is no need to include WALSender
2857 : * since this has no effect on hot standby conflicts.
2858 : *
2859 : * This is never executed during recovery so there is no need to look at
2860 : * KnownAssignedXids.
2861 : *
2862 : * We don't worry about updating other counters, we want to keep this as
2863 : * simple as possible and leave GetSnapshotData() as the primary code for
2864 : * that bookkeeping.
2865 : */
2866 : TransactionId
2867 614 : GetOldestActiveTransactionId(void)
2868 : {
2869 614 : ProcArrayStruct *arrayP = procArray;
2870 614 : TransactionId *other_xids = ProcGlobal->xids;
2871 : TransactionId oldestRunningXid;
2872 : int index;
2873 :
2874 : Assert(!RecoveryInProgress());
2875 :
2876 : /*
2877 : * Read nextXid, as the upper bound of what's still active.
2878 : *
2879 : * Reading a TransactionId is atomic, but we must grab the lock to make
2880 : * sure that all XIDs < nextXid are already present in the proc array (or
2881 : * have already completed), when we spin over it.
2882 : */
2883 614 : LWLockAcquire(XidGenLock, LW_SHARED);
2884 614 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2885 614 : LWLockRelease(XidGenLock);
2886 :
2887 : /*
2888 : * Spin over procArray collecting all xids and subxids.
2889 : */
2890 614 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2891 2090 : for (index = 0; index < arrayP->numProcs; index++)
2892 : {
2893 : TransactionId xid;
2894 :
2895 : /* Fetch xid just once - see GetNewTransactionId */
2896 1476 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2897 :
2898 1476 : if (!TransactionIdIsNormal(xid))
2899 1040 : continue;
2900 :
2901 436 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2902 404 : oldestRunningXid = xid;
2903 :
2904 : /*
2905 : * Top-level XID of a transaction is always less than any of its
2906 : * subxids, so we don't need to check if any of the subxids are
2907 : * smaller than oldestRunningXid
2908 : */
2909 : }
2910 614 : LWLockRelease(ProcArrayLock);
2911 :
2912 614 : return oldestRunningXid;
2913 : }
2914 :
2915 : /*
2916 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2917 : *
2918 : * Returns the oldest xid that we can guarantee not to have been affected by
2919 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2920 : * transaction aborted. Note that the value can (and most of the time will) be
2921 : * much more conservative than what really has been affected by vacuum, but we
2922 : * currently don't have better data available.
2923 : *
2924 : * This is useful to initialize the cutoff xid after which a new changeset
2925 : * extraction replication slot can start decoding changes.
2926 : *
2927 : * Must be called with ProcArrayLock held either shared or exclusively,
2928 : * although most callers will want to use exclusive mode since it is expected
2929 : * that the caller will immediately use the xid to peg the xmin horizon.
2930 : */
2931 : TransactionId
2932 1146 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2933 : {
2934 1146 : ProcArrayStruct *arrayP = procArray;
2935 : TransactionId oldestSafeXid;
2936 : int index;
2937 1146 : bool recovery_in_progress = RecoveryInProgress();
2938 :
2939 : Assert(LWLockHeldByMe(ProcArrayLock));
2940 :
2941 : /*
2942 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2943 : * running. If no transaction with xid were running concurrently a new xid
2944 : * could influence the RecentXmin et al.
2945 : *
2946 : * We initialize the computation to nextXid since that's guaranteed to be
2947 : * a safe, albeit pessimal, value.
2948 : */
2949 1146 : LWLockAcquire(XidGenLock, LW_SHARED);
2950 1146 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2951 :
2952 : /*
2953 : * If there's already a slot pegging the xmin horizon, we can start with
2954 : * that value, it's guaranteed to be safe since it's computed by this
2955 : * routine initially and has been enforced since. We can always use the
2956 : * slot's general xmin horizon, but the catalog horizon is only usable
2957 : * when only catalog data is going to be looked at.
2958 : */
2959 1508 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2960 362 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2961 : oldestSafeXid))
2962 16 : oldestSafeXid = procArray->replication_slot_xmin;
2963 :
2964 1146 : if (catalogOnly &&
2965 584 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2966 122 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2967 : oldestSafeXid))
2968 44 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2969 :
2970 : /*
2971 : * If we're not in recovery, we walk over the procarray and collect the
2972 : * lowest xid. Since we're called with ProcArrayLock held and have
2973 : * acquired XidGenLock, no entries can vanish concurrently, since
2974 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2975 : * with ProcArrayLock held.
2976 : *
2977 : * In recovery we can't lower the safe value besides what we've computed
2978 : * above, so we'll have to wait a bit longer there. We unfortunately can
2979 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2980 : * machinery can miss values and return an older value than is safe.
2981 : */
2982 1146 : if (!recovery_in_progress)
2983 : {
2984 1094 : TransactionId *other_xids = ProcGlobal->xids;
2985 :
2986 : /*
2987 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2988 : */
2989 5454 : for (index = 0; index < arrayP->numProcs; index++)
2990 : {
2991 : TransactionId xid;
2992 :
2993 : /* Fetch xid just once - see GetNewTransactionId */
2994 4360 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2995 :
2996 4360 : if (!TransactionIdIsNormal(xid))
2997 4344 : continue;
2998 :
2999 16 : if (TransactionIdPrecedes(xid, oldestSafeXid))
3000 16 : oldestSafeXid = xid;
3001 : }
3002 : }
3003 :
3004 1146 : LWLockRelease(XidGenLock);
3005 :
3006 1146 : return oldestSafeXid;
3007 : }
3008 :
3009 : /*
3010 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
3011 : * delaying checkpoint because they have critical actions in progress.
3012 : *
3013 : * Constructs an array of VXIDs of transactions that are currently in commit
3014 : * critical sections, as shown by having specified delayChkptFlags bits set
3015 : * in their PGPROC.
3016 : *
3017 : * Returns a palloc'd array that should be freed by the caller.
3018 : * *nvxids is the number of valid entries.
3019 : *
3020 : * Note that because backends set or clear delayChkptFlags without holding any
3021 : * lock, the result is somewhat indeterminate, but we don't really care. Even
3022 : * in a multiprocessor with delayed writes to shared memory, it should be
3023 : * certain that setting of delayChkptFlags will propagate to shared memory
3024 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
3025 : * delayChkptFlags if it's already inserted its commit record. Whether it
3026 : * takes a little while for clearing of delayChkptFlags to propagate is
3027 : * unimportant for correctness.
3028 : */
3029 : VirtualTransactionId *
3030 3256 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3031 : {
3032 : VirtualTransactionId *vxids;
3033 3256 : ProcArrayStruct *arrayP = procArray;
3034 3256 : int count = 0;
3035 : int index;
3036 :
3037 : Assert(type != 0);
3038 :
3039 : /* allocate what's certainly enough result space */
3040 : vxids = (VirtualTransactionId *)
3041 3256 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3042 :
3043 3256 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3044 :
3045 7338 : for (index = 0; index < arrayP->numProcs; index++)
3046 : {
3047 4082 : int pgprocno = arrayP->pgprocnos[index];
3048 4082 : PGPROC *proc = &allProcs[pgprocno];
3049 :
3050 4082 : if ((proc->delayChkptFlags & type) != 0)
3051 : {
3052 : VirtualTransactionId vxid;
3053 :
3054 20 : GET_VXID_FROM_PGPROC(vxid, *proc);
3055 20 : if (VirtualTransactionIdIsValid(vxid))
3056 20 : vxids[count++] = vxid;
3057 : }
3058 : }
3059 :
3060 3256 : LWLockRelease(ProcArrayLock);
3061 :
3062 3256 : *nvxids = count;
3063 3256 : return vxids;
3064 : }
3065 :
3066 : /*
3067 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3068 : *
3069 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3070 : * of the specified VXIDs are still in critical sections of code.
3071 : *
3072 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3073 : * those numbers should be small enough for it not to be a problem.
3074 : */
3075 : bool
3076 24 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3077 : {
3078 24 : bool result = false;
3079 24 : ProcArrayStruct *arrayP = procArray;
3080 : int index;
3081 :
3082 : Assert(type != 0);
3083 :
3084 24 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3085 :
3086 198 : for (index = 0; index < arrayP->numProcs; index++)
3087 : {
3088 180 : int pgprocno = arrayP->pgprocnos[index];
3089 180 : PGPROC *proc = &allProcs[pgprocno];
3090 : VirtualTransactionId vxid;
3091 :
3092 180 : GET_VXID_FROM_PGPROC(vxid, *proc);
3093 :
3094 180 : if ((proc->delayChkptFlags & type) != 0 &&
3095 12 : VirtualTransactionIdIsValid(vxid))
3096 : {
3097 : int i;
3098 :
3099 18 : for (i = 0; i < nvxids; i++)
3100 : {
3101 12 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3102 : {
3103 6 : result = true;
3104 6 : break;
3105 : }
3106 : }
3107 12 : if (result)
3108 6 : break;
3109 : }
3110 : }
3111 :
3112 24 : LWLockRelease(ProcArrayLock);
3113 :
3114 24 : return result;
3115 : }
3116 :
3117 : /*
3118 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3119 : *
3120 : * The result may be out of date arbitrarily quickly, so the caller
3121 : * must be careful about how this information is used. NULL is
3122 : * returned if the backend is not active.
3123 : */
3124 : PGPROC *
3125 1154 : ProcNumberGetProc(ProcNumber procNumber)
3126 : {
3127 : PGPROC *result;
3128 :
3129 1154 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3130 2 : return NULL;
3131 1152 : result = GetPGProcByNumber(procNumber);
3132 :
3133 1152 : if (result->pid == 0)
3134 10 : return NULL;
3135 :
3136 1142 : return result;
3137 : }
3138 :
3139 : /*
3140 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3141 : *
3142 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3143 : * result may be out of date arbitrarily quickly, so the caller must be
3144 : * careful about how this information is used.
3145 : */
3146 : void
3147 9352 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3148 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3149 : {
3150 : PGPROC *proc;
3151 :
3152 9352 : *xid = InvalidTransactionId;
3153 9352 : *xmin = InvalidTransactionId;
3154 9352 : *nsubxid = 0;
3155 9352 : *overflowed = false;
3156 :
3157 9352 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3158 0 : return;
3159 9352 : proc = GetPGProcByNumber(procNumber);
3160 :
3161 : /* Need to lock out additions/removals of backends */
3162 9352 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3163 :
3164 9352 : if (proc->pid != 0)
3165 : {
3166 9352 : *xid = proc->xid;
3167 9352 : *xmin = proc->xmin;
3168 9352 : *nsubxid = proc->subxidStatus.count;
3169 9352 : *overflowed = proc->subxidStatus.overflowed;
3170 : }
3171 :
3172 9352 : LWLockRelease(ProcArrayLock);
3173 : }
3174 :
3175 : /*
3176 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3177 : *
3178 : * Returns NULL if not found. Note that it is up to the caller to be
3179 : * sure that the question remains meaningful for long enough for the
3180 : * answer to be used ...
3181 : */
3182 : PGPROC *
3183 9222 : BackendPidGetProc(int pid)
3184 : {
3185 : PGPROC *result;
3186 :
3187 9222 : if (pid == 0) /* never match dummy PGPROCs */
3188 0 : return NULL;
3189 :
3190 9222 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3191 :
3192 9222 : result = BackendPidGetProcWithLock(pid);
3193 :
3194 9222 : LWLockRelease(ProcArrayLock);
3195 :
3196 9222 : return result;
3197 : }
3198 :
3199 : /*
3200 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3201 : *
3202 : * Same as above, except caller must be holding ProcArrayLock. The found
3203 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3204 : */
3205 : PGPROC *
3206 12204 : BackendPidGetProcWithLock(int pid)
3207 : {
3208 12204 : PGPROC *result = NULL;
3209 12204 : ProcArrayStruct *arrayP = procArray;
3210 : int index;
3211 :
3212 12204 : if (pid == 0) /* never match dummy PGPROCs */
3213 0 : return NULL;
3214 :
3215 42988 : for (index = 0; index < arrayP->numProcs; index++)
3216 : {
3217 39178 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3218 :
3219 39178 : if (proc->pid == pid)
3220 : {
3221 8394 : result = proc;
3222 8394 : break;
3223 : }
3224 : }
3225 :
3226 12204 : return result;
3227 : }
3228 :
3229 : /*
3230 : * BackendXidGetPid -- get a backend's pid given its XID
3231 : *
3232 : * Returns 0 if not found or it's a prepared transaction. Note that
3233 : * it is up to the caller to be sure that the question remains
3234 : * meaningful for long enough for the answer to be used ...
3235 : *
3236 : * Only main transaction Ids are considered. This function is mainly
3237 : * useful for determining what backend owns a lock.
3238 : *
3239 : * Beware that not every xact has an XID assigned. However, as long as you
3240 : * only call this using an XID found on disk, you're safe.
3241 : */
3242 : int
3243 60 : BackendXidGetPid(TransactionId xid)
3244 : {
3245 60 : int result = 0;
3246 60 : ProcArrayStruct *arrayP = procArray;
3247 60 : TransactionId *other_xids = ProcGlobal->xids;
3248 : int index;
3249 :
3250 60 : if (xid == InvalidTransactionId) /* never match invalid xid */
3251 0 : return 0;
3252 :
3253 60 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3254 :
3255 184 : for (index = 0; index < arrayP->numProcs; index++)
3256 : {
3257 168 : if (other_xids[index] == xid)
3258 : {
3259 44 : int pgprocno = arrayP->pgprocnos[index];
3260 44 : PGPROC *proc = &allProcs[pgprocno];
3261 :
3262 44 : result = proc->pid;
3263 44 : break;
3264 : }
3265 : }
3266 :
3267 60 : LWLockRelease(ProcArrayLock);
3268 :
3269 60 : return result;
3270 : }
3271 :
3272 : /*
3273 : * IsBackendPid -- is a given pid a running backend
3274 : *
3275 : * This is not called by the backend, but is called by external modules.
3276 : */
3277 : bool
3278 4 : IsBackendPid(int pid)
3279 : {
3280 4 : return (BackendPidGetProc(pid) != NULL);
3281 : }
3282 :
3283 :
3284 : /*
3285 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3286 : *
3287 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3288 : *
3289 : * The arguments allow filtering the set of VXIDs returned. Our own process
3290 : * is always skipped. In addition:
3291 : * If limitXmin is not InvalidTransactionId, skip processes with
3292 : * xmin > limitXmin.
3293 : * If excludeXmin0 is true, skip processes with xmin = 0.
3294 : * If allDbs is false, skip processes attached to other databases.
3295 : * If excludeVacuum isn't zero, skip processes for which
3296 : * (statusFlags & excludeVacuum) is not zero.
3297 : *
3298 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3299 : * allow skipping backends whose oldest live snapshot is no older than
3300 : * some snapshot we have. Since we examine the procarray with only shared
3301 : * lock, there are race conditions: a backend could set its xmin just after
3302 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3303 : * other backend could have set its xmin *before* we look. We know however
3304 : * that such a backend must have held shared ProcArrayLock overlapping our
3305 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3306 : * any snapshot the other backend is taking concurrently with our scan cannot
3307 : * consider any transactions as still running that we think are committed
3308 : * (since backends must hold ProcArrayLock exclusive to commit).
3309 : */
3310 : VirtualTransactionId *
3311 692 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3312 : bool allDbs, int excludeVacuum,
3313 : int *nvxids)
3314 : {
3315 : VirtualTransactionId *vxids;
3316 692 : ProcArrayStruct *arrayP = procArray;
3317 692 : int count = 0;
3318 : int index;
3319 :
3320 : /* allocate what's certainly enough result space */
3321 : vxids = (VirtualTransactionId *)
3322 692 : palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs);
3323 :
3324 692 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3325 :
3326 4200 : for (index = 0; index < arrayP->numProcs; index++)
3327 : {
3328 3508 : int pgprocno = arrayP->pgprocnos[index];
3329 3508 : PGPROC *proc = &allProcs[pgprocno];
3330 3508 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3331 :
3332 3508 : if (proc == MyProc)
3333 692 : continue;
3334 :
3335 2816 : if (excludeVacuum & statusFlags)
3336 16 : continue;
3337 :
3338 2800 : if (allDbs || proc->databaseId == MyDatabaseId)
3339 : {
3340 : /* Fetch xmin just once - might change on us */
3341 1278 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3342 :
3343 1278 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3344 780 : continue;
3345 :
3346 : /*
3347 : * InvalidTransactionId precedes all other XIDs, so a proc that
3348 : * hasn't set xmin yet will not be rejected by this test.
3349 : */
3350 996 : if (!TransactionIdIsValid(limitXmin) ||
3351 498 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3352 : {
3353 : VirtualTransactionId vxid;
3354 :
3355 482 : GET_VXID_FROM_PGPROC(vxid, *proc);
3356 482 : if (VirtualTransactionIdIsValid(vxid))
3357 482 : vxids[count++] = vxid;
3358 : }
3359 : }
3360 : }
3361 :
3362 692 : LWLockRelease(ProcArrayLock);
3363 :
3364 692 : *nvxids = count;
3365 692 : return vxids;
3366 : }
3367 :
3368 : /*
3369 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3370 : *
3371 : * Usage is limited to conflict resolution during recovery on standby servers.
3372 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3373 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3374 : * determine a safe snapshotConflictHorizon value.
3375 : *
3376 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3377 : * so we're not worried if they have a snapshot or not, nor does it really
3378 : * matter what type of lock we hold. Caller must avoid calling here with
3379 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3380 : * during original execution, since that actually indicates that there is
3381 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3382 : * convention for InvalidTransactionId values is the opposite of our own!).
3383 : *
3384 : * All callers that are checking xmins always now supply a valid and useful
3385 : * value for limitXmin. The limitXmin is always lower than the lowest
3386 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3387 : * because we only care about cleanup records that are cleaning up tuple
3388 : * versions from committed transactions. In that case they will only occur
3389 : * at the point where the record is less than the lowest running xid. That
3390 : * allows us to say that if any backend takes a snapshot concurrently with
3391 : * us then the conflict assessment made here would never include the snapshot
3392 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3393 : * concurrent snapshots when limitXmin is valid. We might think about adding
3394 : * Assert(limitXmin < lowest(KnownAssignedXids))
3395 : * but that would not be true in the case of FATAL errors lagging in array,
3396 : * but we already know those are bogus anyway, so we skip that test.
3397 : *
3398 : * If dbOid is valid we skip backends attached to other databases.
3399 : *
3400 : * Be careful to *not* pfree the result from this function. We reuse
3401 : * this array sufficiently often that we use malloc for the result.
3402 : */
3403 : VirtualTransactionId *
3404 18218 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3405 : {
3406 : static VirtualTransactionId *vxids;
3407 18218 : ProcArrayStruct *arrayP = procArray;
3408 18218 : int count = 0;
3409 : int index;
3410 :
3411 : /*
3412 : * If first time through, get workspace to remember main XIDs in. We
3413 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3414 : * result space, remembering room for a terminator.
3415 : */
3416 18218 : if (vxids == NULL)
3417 : {
3418 36 : vxids = (VirtualTransactionId *)
3419 36 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3420 36 : if (vxids == NULL)
3421 0 : ereport(ERROR,
3422 : (errcode(ERRCODE_OUT_OF_MEMORY),
3423 : errmsg("out of memory")));
3424 : }
3425 :
3426 18218 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3427 :
3428 18512 : for (index = 0; index < arrayP->numProcs; index++)
3429 : {
3430 294 : int pgprocno = arrayP->pgprocnos[index];
3431 294 : PGPROC *proc = &allProcs[pgprocno];
3432 :
3433 : /* Exclude prepared transactions */
3434 294 : if (proc->pid == 0)
3435 0 : continue;
3436 :
3437 294 : if (!OidIsValid(dbOid) ||
3438 282 : proc->databaseId == dbOid)
3439 : {
3440 : /* Fetch xmin just once - can't change on us, but good coding */
3441 32 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3442 :
3443 : /*
3444 : * We ignore an invalid pxmin because this means that backend has
3445 : * no snapshot currently. We hold a Share lock to avoid contention
3446 : * with users taking snapshots. That is not a problem because the
3447 : * current xmin is always at least one higher than the latest
3448 : * removed xid, so any new snapshot would never conflict with the
3449 : * test here.
3450 : */
3451 32 : if (!TransactionIdIsValid(limitXmin) ||
3452 6 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3453 : {
3454 : VirtualTransactionId vxid;
3455 :
3456 4 : GET_VXID_FROM_PGPROC(vxid, *proc);
3457 4 : if (VirtualTransactionIdIsValid(vxid))
3458 4 : vxids[count++] = vxid;
3459 : }
3460 : }
3461 : }
3462 :
3463 18218 : LWLockRelease(ProcArrayLock);
3464 :
3465 : /* add the terminator */
3466 18218 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3467 18218 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3468 :
3469 18218 : return vxids;
3470 : }
3471 :
3472 : /*
3473 : * CancelVirtualTransaction - used in recovery conflict processing
3474 : *
3475 : * Returns pid of the process signaled, or 0 if not found.
3476 : */
3477 : pid_t
3478 6 : CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
3479 : {
3480 6 : return SignalVirtualTransaction(vxid, sigmode, true);
3481 : }
3482 :
3483 : pid_t
3484 10 : SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode,
3485 : bool conflictPending)
3486 : {
3487 10 : ProcArrayStruct *arrayP = procArray;
3488 : int index;
3489 10 : pid_t pid = 0;
3490 :
3491 10 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3492 :
3493 10 : for (index = 0; index < arrayP->numProcs; index++)
3494 : {
3495 10 : int pgprocno = arrayP->pgprocnos[index];
3496 10 : PGPROC *proc = &allProcs[pgprocno];
3497 : VirtualTransactionId procvxid;
3498 :
3499 10 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3500 :
3501 10 : if (procvxid.procNumber == vxid.procNumber &&
3502 10 : procvxid.localTransactionId == vxid.localTransactionId)
3503 : {
3504 10 : proc->recoveryConflictPending = conflictPending;
3505 10 : pid = proc->pid;
3506 10 : if (pid != 0)
3507 : {
3508 : /*
3509 : * Kill the pid if it's still here. If not, that's what we
3510 : * wanted so ignore any errors.
3511 : */
3512 10 : (void) SendProcSignal(pid, sigmode, vxid.procNumber);
3513 : }
3514 10 : break;
3515 : }
3516 : }
3517 :
3518 10 : LWLockRelease(ProcArrayLock);
3519 :
3520 10 : return pid;
3521 : }
3522 :
3523 : /*
3524 : * MinimumActiveBackends --- count backends (other than myself) that are
3525 : * in active transactions. Return true if the count exceeds the
3526 : * minimum threshold passed. This is used as a heuristic to decide if
3527 : * a pre-XLOG-flush delay is worthwhile during commit.
3528 : *
3529 : * Do not count backends that are blocked waiting for locks, since they are
3530 : * not going to get to run until someone else commits.
3531 : */
3532 : bool
3533 0 : MinimumActiveBackends(int min)
3534 : {
3535 0 : ProcArrayStruct *arrayP = procArray;
3536 0 : int count = 0;
3537 : int index;
3538 :
3539 : /* Quick short-circuit if no minimum is specified */
3540 0 : if (min == 0)
3541 0 : return true;
3542 :
3543 : /*
3544 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3545 : * bogus, but since we are only testing fields for zero or nonzero, it
3546 : * should be OK. The result is only used for heuristic purposes anyway...
3547 : */
3548 0 : for (index = 0; index < arrayP->numProcs; index++)
3549 : {
3550 0 : int pgprocno = arrayP->pgprocnos[index];
3551 0 : PGPROC *proc = &allProcs[pgprocno];
3552 :
3553 : /*
3554 : * Since we're not holding a lock, need to be prepared to deal with
3555 : * garbage, as someone could have incremented numProcs but not yet
3556 : * filled the structure.
3557 : *
3558 : * If someone just decremented numProcs, 'proc' could also point to a
3559 : * PGPROC entry that's no longer in the array. It still points to a
3560 : * PGPROC struct, though, because freed PGPROC entries just go to the
3561 : * free list and are recycled. Its contents are nonsense in that case,
3562 : * but that's acceptable for this function.
3563 : */
3564 0 : if (pgprocno == -1)
3565 0 : continue; /* do not count deleted entries */
3566 0 : if (proc == MyProc)
3567 0 : continue; /* do not count myself */
3568 0 : if (proc->xid == InvalidTransactionId)
3569 0 : continue; /* do not count if no XID assigned */
3570 0 : if (proc->pid == 0)
3571 0 : continue; /* do not count prepared xacts */
3572 0 : if (proc->waitLock != NULL)
3573 0 : continue; /* do not count if blocked on a lock */
3574 0 : count++;
3575 0 : if (count >= min)
3576 0 : break;
3577 : }
3578 :
3579 0 : return count >= min;
3580 : }
3581 :
3582 : /*
3583 : * CountDBBackends --- count backends that are using specified database
3584 : */
3585 : int
3586 26 : CountDBBackends(Oid databaseid)
3587 : {
3588 26 : ProcArrayStruct *arrayP = procArray;
3589 26 : int count = 0;
3590 : int index;
3591 :
3592 26 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3593 :
3594 42 : for (index = 0; index < arrayP->numProcs; index++)
3595 : {
3596 16 : int pgprocno = arrayP->pgprocnos[index];
3597 16 : PGPROC *proc = &allProcs[pgprocno];
3598 :
3599 16 : if (proc->pid == 0)
3600 0 : continue; /* do not count prepared xacts */
3601 16 : if (!OidIsValid(databaseid) ||
3602 16 : proc->databaseId == databaseid)
3603 4 : count++;
3604 : }
3605 :
3606 26 : LWLockRelease(ProcArrayLock);
3607 :
3608 26 : return count;
3609 : }
3610 :
3611 : /*
3612 : * CountDBConnections --- counts database backends ignoring any background
3613 : * worker processes
3614 : */
3615 : int
3616 0 : CountDBConnections(Oid databaseid)
3617 : {
3618 0 : ProcArrayStruct *arrayP = procArray;
3619 0 : int count = 0;
3620 : int index;
3621 :
3622 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3623 :
3624 0 : for (index = 0; index < arrayP->numProcs; index++)
3625 : {
3626 0 : int pgprocno = arrayP->pgprocnos[index];
3627 0 : PGPROC *proc = &allProcs[pgprocno];
3628 :
3629 0 : if (proc->pid == 0)
3630 0 : continue; /* do not count prepared xacts */
3631 0 : if (proc->isBackgroundWorker)
3632 0 : continue; /* do not count background workers */
3633 0 : if (!OidIsValid(databaseid) ||
3634 0 : proc->databaseId == databaseid)
3635 0 : count++;
3636 : }
3637 :
3638 0 : LWLockRelease(ProcArrayLock);
3639 :
3640 0 : return count;
3641 : }
3642 :
3643 : /*
3644 : * CancelDBBackends --- cancel backends that are using specified database
3645 : */
3646 : void
3647 20 : CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3648 : {
3649 20 : ProcArrayStruct *arrayP = procArray;
3650 : int index;
3651 :
3652 : /* tell all backends to die */
3653 20 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3654 :
3655 40 : for (index = 0; index < arrayP->numProcs; index++)
3656 : {
3657 20 : int pgprocno = arrayP->pgprocnos[index];
3658 20 : PGPROC *proc = &allProcs[pgprocno];
3659 :
3660 20 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3661 : {
3662 : VirtualTransactionId procvxid;
3663 : pid_t pid;
3664 :
3665 20 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3666 :
3667 20 : proc->recoveryConflictPending = conflictPending;
3668 20 : pid = proc->pid;
3669 20 : if (pid != 0)
3670 : {
3671 : /*
3672 : * Kill the pid if it's still here. If not, that's what we
3673 : * wanted so ignore any errors.
3674 : */
3675 20 : (void) SendProcSignal(pid, sigmode, procvxid.procNumber);
3676 : }
3677 : }
3678 : }
3679 :
3680 20 : LWLockRelease(ProcArrayLock);
3681 20 : }
3682 :
3683 : /*
3684 : * CountUserBackends --- count backends that are used by specified user
3685 : */
3686 : int
3687 0 : CountUserBackends(Oid roleid)
3688 : {
3689 0 : ProcArrayStruct *arrayP = procArray;
3690 0 : int count = 0;
3691 : int index;
3692 :
3693 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3694 :
3695 0 : for (index = 0; index < arrayP->numProcs; index++)
3696 : {
3697 0 : int pgprocno = arrayP->pgprocnos[index];
3698 0 : PGPROC *proc = &allProcs[pgprocno];
3699 :
3700 0 : if (proc->pid == 0)
3701 0 : continue; /* do not count prepared xacts */
3702 0 : if (proc->isBackgroundWorker)
3703 0 : continue; /* do not count background workers */
3704 0 : if (proc->roleId == roleid)
3705 0 : count++;
3706 : }
3707 :
3708 0 : LWLockRelease(ProcArrayLock);
3709 :
3710 0 : return count;
3711 : }
3712 :
3713 : /*
3714 : * CountOtherDBBackends -- check for other backends running in the given DB
3715 : *
3716 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3717 : * for them to exit. Autovacuum backends are encouraged to exit early by
3718 : * sending them SIGTERM, but normal user backends are just waited for.
3719 : *
3720 : * The current backend is always ignored; it is caller's responsibility to
3721 : * check whether the current backend uses the given DB, if it's important.
3722 : *
3723 : * Returns true if there are (still) other backends in the DB, false if not.
3724 : * Also, *nbackends and *nprepared are set to the number of other backends
3725 : * and prepared transactions in the DB, respectively.
3726 : *
3727 : * This function is used to interlock DROP DATABASE and related commands
3728 : * against there being any active backends in the target DB --- dropping the
3729 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3730 : * detect here the possibility of a newly-started backend that is trying to
3731 : * connect to the doomed database, so additional interlocking is needed during
3732 : * backend startup. The caller should normally hold an exclusive lock on the
3733 : * target DB before calling this, which is one reason we mustn't wait
3734 : * indefinitely.
3735 : */
3736 : bool
3737 676 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3738 : {
3739 676 : ProcArrayStruct *arrayP = procArray;
3740 :
3741 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3742 : int autovac_pids[MAXAUTOVACPIDS];
3743 : int tries;
3744 :
3745 : /* 50 tries with 100ms sleep between tries makes 5 sec total wait */
3746 676 : for (tries = 0; tries < 50; tries++)
3747 : {
3748 676 : int nautovacs = 0;
3749 676 : bool found = false;
3750 : int index;
3751 :
3752 676 : CHECK_FOR_INTERRUPTS();
3753 :
3754 676 : *nbackends = *nprepared = 0;
3755 :
3756 676 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3757 :
3758 2410 : for (index = 0; index < arrayP->numProcs; index++)
3759 : {
3760 1734 : int pgprocno = arrayP->pgprocnos[index];
3761 1734 : PGPROC *proc = &allProcs[pgprocno];
3762 1734 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3763 :
3764 1734 : if (proc->databaseId != databaseId)
3765 1582 : continue;
3766 152 : if (proc == MyProc)
3767 152 : continue;
3768 :
3769 0 : found = true;
3770 :
3771 0 : if (proc->pid == 0)
3772 0 : (*nprepared)++;
3773 : else
3774 : {
3775 0 : (*nbackends)++;
3776 0 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3777 : nautovacs < MAXAUTOVACPIDS)
3778 0 : autovac_pids[nautovacs++] = proc->pid;
3779 : }
3780 : }
3781 :
3782 676 : LWLockRelease(ProcArrayLock);
3783 :
3784 676 : if (!found)
3785 676 : return false; /* no conflicting backends, so done */
3786 :
3787 : /*
3788 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3789 : * postpone this step until after the loop because we don't want to
3790 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3791 : * block kill() inside the kernel...
3792 : */
3793 0 : for (index = 0; index < nautovacs; index++)
3794 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3795 :
3796 : /* sleep, then try again */
3797 0 : pg_usleep(100 * 1000L); /* 100ms */
3798 : }
3799 :
3800 0 : return true; /* timed out, still conflicts */
3801 : }
3802 :
3803 : /*
3804 : * Terminate existing connections to the specified database. This routine
3805 : * is used by the DROP DATABASE command when user has asked to forcefully
3806 : * drop the database.
3807 : *
3808 : * The current backend is always ignored; it is caller's responsibility to
3809 : * check whether the current backend uses the given DB, if it's important.
3810 : *
3811 : * It doesn't allow to terminate the connections even if there is a one
3812 : * backend with the prepared transaction in the target database.
3813 : */
3814 : void
3815 2 : TerminateOtherDBBackends(Oid databaseId)
3816 : {
3817 2 : ProcArrayStruct *arrayP = procArray;
3818 2 : List *pids = NIL;
3819 2 : int nprepared = 0;
3820 : int i;
3821 :
3822 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3823 :
3824 8 : for (i = 0; i < procArray->numProcs; i++)
3825 : {
3826 6 : int pgprocno = arrayP->pgprocnos[i];
3827 6 : PGPROC *proc = &allProcs[pgprocno];
3828 :
3829 6 : if (proc->databaseId != databaseId)
3830 6 : continue;
3831 0 : if (proc == MyProc)
3832 0 : continue;
3833 :
3834 0 : if (proc->pid != 0)
3835 0 : pids = lappend_int(pids, proc->pid);
3836 : else
3837 0 : nprepared++;
3838 : }
3839 :
3840 2 : LWLockRelease(ProcArrayLock);
3841 :
3842 2 : if (nprepared > 0)
3843 0 : ereport(ERROR,
3844 : (errcode(ERRCODE_OBJECT_IN_USE),
3845 : errmsg("database \"%s\" is being used by prepared transactions",
3846 : get_database_name(databaseId)),
3847 : errdetail_plural("There is %d prepared transaction using the database.",
3848 : "There are %d prepared transactions using the database.",
3849 : nprepared,
3850 : nprepared)));
3851 :
3852 2 : if (pids)
3853 : {
3854 : ListCell *lc;
3855 :
3856 : /*
3857 : * Check whether we have the necessary rights to terminate other
3858 : * sessions. We don't terminate any session until we ensure that we
3859 : * have rights on all the sessions to be terminated. These checks are
3860 : * the same as we do in pg_terminate_backend.
3861 : *
3862 : * In this case we don't raise some warnings - like "PID %d is not a
3863 : * PostgreSQL server process", because for us already finished session
3864 : * is not a problem.
3865 : */
3866 0 : foreach(lc, pids)
3867 : {
3868 0 : int pid = lfirst_int(lc);
3869 0 : PGPROC *proc = BackendPidGetProc(pid);
3870 :
3871 0 : if (proc != NULL)
3872 : {
3873 : /* Only allow superusers to signal superuser-owned backends. */
3874 0 : if (superuser_arg(proc->roleId) && !superuser())
3875 0 : ereport(ERROR,
3876 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3877 : errmsg("permission denied to terminate process"),
3878 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3879 : "SUPERUSER", "SUPERUSER")));
3880 :
3881 : /* Users can signal backends they have role membership in. */
3882 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3883 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3884 0 : ereport(ERROR,
3885 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3886 : errmsg("permission denied to terminate process"),
3887 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3888 : "pg_signal_backend")));
3889 : }
3890 : }
3891 :
3892 : /*
3893 : * There's a race condition here: once we release the ProcArrayLock,
3894 : * it's possible for the session to exit before we issue kill. That
3895 : * race condition possibility seems too unlikely to worry about. See
3896 : * pg_signal_backend.
3897 : */
3898 0 : foreach(lc, pids)
3899 : {
3900 0 : int pid = lfirst_int(lc);
3901 0 : PGPROC *proc = BackendPidGetProc(pid);
3902 :
3903 0 : if (proc != NULL)
3904 : {
3905 : /*
3906 : * If we have setsid(), signal the backend's whole process
3907 : * group
3908 : */
3909 : #ifdef HAVE_SETSID
3910 0 : (void) kill(-pid, SIGTERM);
3911 : #else
3912 : (void) kill(pid, SIGTERM);
3913 : #endif
3914 : }
3915 : }
3916 : }
3917 2 : }
3918 :
3919 : /*
3920 : * ProcArraySetReplicationSlotXmin
3921 : *
3922 : * Install limits to future computations of the xmin horizon to prevent vacuum
3923 : * and HOT pruning from removing affected rows still needed by clients with
3924 : * replication slots.
3925 : */
3926 : void
3927 3896 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3928 : bool already_locked)
3929 : {
3930 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3931 :
3932 3896 : if (!already_locked)
3933 3092 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3934 :
3935 3896 : procArray->replication_slot_xmin = xmin;
3936 3896 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3937 :
3938 3896 : if (!already_locked)
3939 3092 : LWLockRelease(ProcArrayLock);
3940 :
3941 3896 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3942 : xmin, catalog_xmin);
3943 3896 : }
3944 :
3945 : /*
3946 : * ProcArrayGetReplicationSlotXmin
3947 : *
3948 : * Return the current slot xmin limits. That's useful to be able to remove
3949 : * data that's older than those limits.
3950 : */
3951 : void
3952 44 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3953 : TransactionId *catalog_xmin)
3954 : {
3955 44 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3956 :
3957 44 : if (xmin != NULL)
3958 0 : *xmin = procArray->replication_slot_xmin;
3959 :
3960 44 : if (catalog_xmin != NULL)
3961 44 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3962 :
3963 44 : LWLockRelease(ProcArrayLock);
3964 44 : }
3965 :
3966 : /*
3967 : * XidCacheRemoveRunningXids
3968 : *
3969 : * Remove a bunch of TransactionIds from the list of known-running
3970 : * subtransactions for my backend. Both the specified xid and those in
3971 : * the xids[] array (of length nxids) are removed from the subxids cache.
3972 : * latestXid must be the latest XID among the group.
3973 : */
3974 : void
3975 1292 : XidCacheRemoveRunningXids(TransactionId xid,
3976 : int nxids, const TransactionId *xids,
3977 : TransactionId latestXid)
3978 : {
3979 : int i,
3980 : j;
3981 : XidCacheStatus *mysubxidstat;
3982 :
3983 : Assert(TransactionIdIsValid(xid));
3984 :
3985 : /*
3986 : * We must hold ProcArrayLock exclusively in order to remove transactions
3987 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
3988 : * possible this could be relaxed since we know this routine is only used
3989 : * to abort subtransactions, but pending closer analysis we'd best be
3990 : * conservative.
3991 : *
3992 : * Note that we do not have to be careful about memory ordering of our own
3993 : * reads wrt. GetNewTransactionId() here - only this process can modify
3994 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
3995 : * careful about our own writes being well ordered.
3996 : */
3997 1292 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3998 :
3999 1292 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4000 :
4001 : /*
4002 : * Under normal circumstances xid and xids[] will be in increasing order,
4003 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4004 : * behavior when removing a lot of xids.
4005 : */
4006 1352 : for (i = nxids - 1; i >= 0; i--)
4007 : {
4008 60 : TransactionId anxid = xids[i];
4009 :
4010 60 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4011 : {
4012 60 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4013 : {
4014 60 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4015 60 : pg_write_barrier();
4016 60 : mysubxidstat->count--;
4017 60 : MyProc->subxidStatus.count--;
4018 60 : break;
4019 : }
4020 : }
4021 :
4022 : /*
4023 : * Ordinarily we should have found it, unless the cache has
4024 : * overflowed. However it's also possible for this routine to be
4025 : * invoked multiple times for the same subtransaction, in case of an
4026 : * error during AbortSubTransaction. So instead of Assert, emit a
4027 : * debug warning.
4028 : */
4029 60 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4030 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4031 : }
4032 :
4033 1292 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4034 : {
4035 1292 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4036 : {
4037 1292 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4038 1292 : pg_write_barrier();
4039 1292 : mysubxidstat->count--;
4040 1292 : MyProc->subxidStatus.count--;
4041 1292 : break;
4042 : }
4043 : }
4044 : /* Ordinarily we should have found it, unless the cache has overflowed */
4045 1292 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4046 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4047 :
4048 : /* Also advance global latestCompletedXid while holding the lock */
4049 1292 : MaintainLatestCompletedXid(latestXid);
4050 :
4051 : /* ... and xactCompletionCount */
4052 1292 : TransamVariables->xactCompletionCount++;
4053 :
4054 1292 : LWLockRelease(ProcArrayLock);
4055 1292 : }
4056 :
4057 : #ifdef XIDCACHE_DEBUG
4058 :
4059 : /*
4060 : * Print stats about effectiveness of XID cache
4061 : */
4062 : static void
4063 : DisplayXidCache(void)
4064 : {
4065 : fprintf(stderr,
4066 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4067 : xc_by_recent_xmin,
4068 : xc_by_known_xact,
4069 : xc_by_my_xact,
4070 : xc_by_latest_xid,
4071 : xc_by_main_xid,
4072 : xc_by_child_xid,
4073 : xc_by_known_assigned,
4074 : xc_no_overflow,
4075 : xc_slow_answer);
4076 : }
4077 : #endif /* XIDCACHE_DEBUG */
4078 :
4079 : /*
4080 : * If rel != NULL, return test state appropriate for relation, otherwise
4081 : * return state usable for all relations. The latter may consider XIDs as
4082 : * not-yet-visible-to-everyone that a state for a specific relation would
4083 : * already consider visible-to-everyone.
4084 : *
4085 : * This needs to be called while a snapshot is active or registered, otherwise
4086 : * there are wraparound and other dangers.
4087 : *
4088 : * See comment for GlobalVisState for details.
4089 : */
4090 : GlobalVisState *
4091 27463976 : GlobalVisTestFor(Relation rel)
4092 : {
4093 27463976 : GlobalVisState *state = NULL;
4094 :
4095 : /* XXX: we should assert that a snapshot is pushed or registered */
4096 : Assert(RecentXmin);
4097 :
4098 27463976 : switch (GlobalVisHorizonKindForRel(rel))
4099 : {
4100 101042 : case VISHORIZON_SHARED:
4101 101042 : state = &GlobalVisSharedRels;
4102 101042 : break;
4103 4722026 : case VISHORIZON_CATALOG:
4104 4722026 : state = &GlobalVisCatalogRels;
4105 4722026 : break;
4106 22557266 : case VISHORIZON_DATA:
4107 22557266 : state = &GlobalVisDataRels;
4108 22557266 : break;
4109 83642 : case VISHORIZON_TEMP:
4110 83642 : state = &GlobalVisTempRels;
4111 83642 : break;
4112 : }
4113 :
4114 27463976 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4115 : FullTransactionIdIsValid(state->maybe_needed));
4116 :
4117 27463976 : return state;
4118 : }
4119 :
4120 : /*
4121 : * Return true if it's worth updating the accurate maybe_needed boundary.
4122 : *
4123 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4124 : * repeatedly do so when there is a low likelihood of it being beneficial.
4125 : *
4126 : * The current heuristic is that we update only if RecentXmin has changed
4127 : * since the last update. If the oldest currently running transaction has not
4128 : * finished, it is unlikely that recomputing the horizon would be useful.
4129 : */
4130 : static bool
4131 902762 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4132 : {
4133 : /* hasn't been updated yet */
4134 902762 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4135 14742 : return true;
4136 :
4137 : /*
4138 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4139 : * unlikely to be beneficial to refresh boundaries.
4140 : */
4141 888020 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4142 : state->definitely_needed))
4143 0 : return false;
4144 :
4145 : /* does the last snapshot built have a different xmin? */
4146 888020 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4147 : }
4148 :
4149 : static void
4150 173356 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4151 : {
4152 : GlobalVisSharedRels.maybe_needed =
4153 173356 : FullXidRelativeTo(horizons->latest_completed,
4154 : horizons->shared_oldest_nonremovable);
4155 : GlobalVisCatalogRels.maybe_needed =
4156 173356 : FullXidRelativeTo(horizons->latest_completed,
4157 : horizons->catalog_oldest_nonremovable);
4158 : GlobalVisDataRels.maybe_needed =
4159 173356 : FullXidRelativeTo(horizons->latest_completed,
4160 : horizons->data_oldest_nonremovable);
4161 : GlobalVisTempRels.maybe_needed =
4162 173356 : FullXidRelativeTo(horizons->latest_completed,
4163 : horizons->temp_oldest_nonremovable);
4164 :
4165 : /*
4166 : * In longer running transactions it's possible that transactions we
4167 : * previously needed to treat as running aren't around anymore. So update
4168 : * definitely_needed to not be earlier than maybe_needed.
4169 : */
4170 : GlobalVisSharedRels.definitely_needed =
4171 173356 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4172 : GlobalVisSharedRels.definitely_needed);
4173 : GlobalVisCatalogRels.definitely_needed =
4174 173356 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4175 : GlobalVisCatalogRels.definitely_needed);
4176 : GlobalVisDataRels.definitely_needed =
4177 173356 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4178 : GlobalVisDataRels.definitely_needed);
4179 173356 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4180 :
4181 173356 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4182 173356 : }
4183 :
4184 : /*
4185 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4186 : * using ComputeXidHorizons().
4187 : */
4188 : static void
4189 99320 : GlobalVisUpdate(void)
4190 : {
4191 : ComputeXidHorizonsResult horizons;
4192 :
4193 : /* updates the horizons as a side-effect */
4194 99320 : ComputeXidHorizons(&horizons);
4195 99320 : }
4196 :
4197 : /*
4198 : * Return true if no snapshot still considers fxid to be running.
4199 : *
4200 : * The state passed needs to have been initialized for the relation fxid is
4201 : * from (NULL is also OK), otherwise the result may not be correct.
4202 : *
4203 : * See comment for GlobalVisState for details.
4204 : */
4205 : bool
4206 19105144 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4207 : FullTransactionId fxid)
4208 : {
4209 : /*
4210 : * If fxid is older than maybe_needed bound, it definitely is visible to
4211 : * everyone.
4212 : */
4213 19105144 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4214 5077156 : return true;
4215 :
4216 : /*
4217 : * If fxid is >= definitely_needed bound, it is very likely to still be
4218 : * considered running.
4219 : */
4220 14027988 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4221 13125226 : return false;
4222 :
4223 : /*
4224 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4225 : * might not exist a snapshot considering fxid running. If it makes sense,
4226 : * update boundaries and recheck.
4227 : */
4228 902762 : if (GlobalVisTestShouldUpdate(state))
4229 : {
4230 99320 : GlobalVisUpdate();
4231 :
4232 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4233 :
4234 99320 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4235 : }
4236 : else
4237 803442 : return false;
4238 : }
4239 :
4240 : /*
4241 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4242 : *
4243 : * It is crucial that this only gets called for xids from a source that
4244 : * protects against xid wraparounds (e.g. from a table and thus protected by
4245 : * relfrozenxid).
4246 : */
4247 : bool
4248 19104820 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4249 : {
4250 : FullTransactionId fxid;
4251 :
4252 : /*
4253 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4254 : * because we know the xid has to, at the very least, be between
4255 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4256 : * lock to determine either, we can just compare with
4257 : * state->definitely_needed, which was based on those value at the time
4258 : * the current snapshot was built.
4259 : */
4260 19104820 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4261 :
4262 19104820 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4263 : }
4264 :
4265 : /*
4266 : * Convenience wrapper around GlobalVisTestFor() and
4267 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4268 : */
4269 : bool
4270 324 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4271 : {
4272 : GlobalVisState *state;
4273 :
4274 324 : state = GlobalVisTestFor(rel);
4275 :
4276 324 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4277 : }
4278 :
4279 : /*
4280 : * Convenience wrapper around GlobalVisTestFor() and
4281 : * GlobalVisTestIsRemovableXid(), see their comments.
4282 : */
4283 : bool
4284 12 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4285 : {
4286 : GlobalVisState *state;
4287 :
4288 12 : state = GlobalVisTestFor(rel);
4289 :
4290 12 : return GlobalVisTestIsRemovableXid(state, xid);
4291 : }
4292 :
4293 : /*
4294 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4295 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4296 : *
4297 : * Be very careful about when to use this function. It can only safely be used
4298 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4299 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4300 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4301 : * the xid has to be within that range), or if xid is from the procarray and
4302 : * prevents xid wraparound that way.
4303 : */
4304 : static inline FullTransactionId
4305 22082082 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4306 : {
4307 22082082 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4308 :
4309 : Assert(TransactionIdIsValid(xid));
4310 : Assert(TransactionIdIsValid(rel_xid));
4311 :
4312 : /* not guaranteed to find issues, but likely to catch mistakes */
4313 : AssertTransactionIdInAllowableRange(xid);
4314 :
4315 44164164 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4316 22082082 : + (int32) (xid - rel_xid));
4317 : }
4318 :
4319 :
4320 : /* ----------------------------------------------
4321 : * KnownAssignedTransactionIds sub-module
4322 : * ----------------------------------------------
4323 : */
4324 :
4325 : /*
4326 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4327 : * running on the primary at the current point in WAL. These XIDs must be
4328 : * treated as running by standby transactions, even though they are not in
4329 : * the standby server's PGPROC array.
4330 : *
4331 : * We record all XIDs that we know have been assigned. That includes all the
4332 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4333 : * been assigned. We can deduce the existence of unobserved XIDs because we
4334 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4335 : * list expands as new XIDs are observed or inferred, and contracts when
4336 : * transaction completion records arrive.
4337 : *
4338 : * During hot standby we do not fret too much about the distinction between
4339 : * top-level XIDs and subtransaction XIDs. We store both together in the
4340 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4341 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4342 : * doesn't care about the distinction either. Subtransaction XIDs are
4343 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4344 : * links are *not* maintained (which does not affect visibility).
4345 : *
4346 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4347 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4348 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4349 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4350 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4351 : * and then remove them from KnownAssignedXids. This prevents overflow of
4352 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4353 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4354 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4355 : * though it should be complete for top-level XIDs; this is the same situation
4356 : * that holds with respect to the PGPROC entries in normal running.
4357 : *
4358 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4359 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4360 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4361 : * As long as that is within the range of interesting XIDs, we have to assume
4362 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4363 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4364 : * subXID arrives - that is not an error.)
4365 : *
4366 : * Should a backend on primary somehow disappear before it can write an abort
4367 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4368 : * aborted but we think they were running; the distinction is irrelevant
4369 : * because either way any changes done by the transaction are not visible to
4370 : * backends in the standby. We prune KnownAssignedXids when
4371 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4372 : * array due to such dead XIDs.
4373 : */
4374 :
4375 : /*
4376 : * RecordKnownAssignedTransactionIds
4377 : * Record the given XID in KnownAssignedXids, as well as any preceding
4378 : * unobserved XIDs.
4379 : *
4380 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4381 : * associated with a transaction. Must be called for each record after we
4382 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4383 : *
4384 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4385 : */
4386 : void
4387 4737326 : RecordKnownAssignedTransactionIds(TransactionId xid)
4388 : {
4389 : Assert(standbyState >= STANDBY_INITIALIZED);
4390 : Assert(TransactionIdIsValid(xid));
4391 : Assert(TransactionIdIsValid(latestObservedXid));
4392 :
4393 4737326 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4394 : xid, latestObservedXid);
4395 :
4396 : /*
4397 : * When a newly observed xid arrives, it is frequently the case that it is
4398 : * *not* the next xid in sequence. When this occurs, we must treat the
4399 : * intervening xids as running also.
4400 : */
4401 4737326 : if (TransactionIdFollows(xid, latestObservedXid))
4402 : {
4403 : TransactionId next_expected_xid;
4404 :
4405 : /*
4406 : * Extend subtrans like we do in GetNewTransactionId() during normal
4407 : * operation using individual extend steps. Note that we do not need
4408 : * to extend clog since its extensions are WAL logged.
4409 : *
4410 : * This part has to be done regardless of standbyState since we
4411 : * immediately start assigning subtransactions to their toplevel
4412 : * transactions.
4413 : */
4414 42458 : next_expected_xid = latestObservedXid;
4415 85406 : while (TransactionIdPrecedes(next_expected_xid, xid))
4416 : {
4417 42948 : TransactionIdAdvance(next_expected_xid);
4418 42948 : ExtendSUBTRANS(next_expected_xid);
4419 : }
4420 : Assert(next_expected_xid == xid);
4421 :
4422 : /*
4423 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4424 : * more to do since we don't track assigned xids yet.
4425 : */
4426 42458 : if (standbyState <= STANDBY_INITIALIZED)
4427 : {
4428 0 : latestObservedXid = xid;
4429 0 : return;
4430 : }
4431 :
4432 : /*
4433 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4434 : */
4435 42458 : next_expected_xid = latestObservedXid;
4436 42458 : TransactionIdAdvance(next_expected_xid);
4437 42458 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4438 :
4439 : /*
4440 : * Now we can advance latestObservedXid
4441 : */
4442 42458 : latestObservedXid = xid;
4443 :
4444 : /* TransamVariables->nextXid must be beyond any observed xid */
4445 42458 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4446 : }
4447 : }
4448 :
4449 : /*
4450 : * ExpireTreeKnownAssignedTransactionIds
4451 : * Remove the given XIDs from KnownAssignedXids.
4452 : *
4453 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4454 : */
4455 : void
4456 39890 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4457 : TransactionId *subxids, TransactionId max_xid)
4458 : {
4459 : Assert(standbyState >= STANDBY_INITIALIZED);
4460 :
4461 : /*
4462 : * Uses same locking as transaction commit
4463 : */
4464 39890 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4465 :
4466 39890 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4467 :
4468 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4469 39890 : MaintainLatestCompletedXidRecovery(max_xid);
4470 :
4471 : /* ... and xactCompletionCount */
4472 39890 : TransamVariables->xactCompletionCount++;
4473 :
4474 39890 : LWLockRelease(ProcArrayLock);
4475 39890 : }
4476 :
4477 : /*
4478 : * ExpireAllKnownAssignedTransactionIds
4479 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4480 : */
4481 : void
4482 186 : ExpireAllKnownAssignedTransactionIds(void)
4483 : {
4484 186 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4485 186 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4486 :
4487 : /*
4488 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4489 : * the call of this function. But do this for unification with what
4490 : * ExpireOldKnownAssignedTransactionIds() do.
4491 : */
4492 186 : procArray->lastOverflowedXid = InvalidTransactionId;
4493 186 : LWLockRelease(ProcArrayLock);
4494 186 : }
4495 :
4496 : /*
4497 : * ExpireOldKnownAssignedTransactionIds
4498 : * Remove KnownAssignedXids entries preceding the given XID and
4499 : * potentially reset lastOverflowedXid.
4500 : */
4501 : void
4502 506 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4503 : {
4504 506 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4505 :
4506 : /*
4507 : * Reset lastOverflowedXid if we know all transactions that have been
4508 : * possibly running are being gone. Not doing so could cause an incorrect
4509 : * lastOverflowedXid value, which makes extra snapshots be marked as
4510 : * suboverflowed.
4511 : */
4512 506 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4513 498 : procArray->lastOverflowedXid = InvalidTransactionId;
4514 506 : KnownAssignedXidsRemovePreceding(xid);
4515 506 : LWLockRelease(ProcArrayLock);
4516 506 : }
4517 :
4518 : /*
4519 : * KnownAssignedTransactionIdsIdleMaintenance
4520 : * Opportunistically do maintenance work when the startup process
4521 : * is about to go idle.
4522 : */
4523 : void
4524 29842 : KnownAssignedTransactionIdsIdleMaintenance(void)
4525 : {
4526 29842 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4527 29842 : }
4528 :
4529 :
4530 : /*
4531 : * Private module functions to manipulate KnownAssignedXids
4532 : *
4533 : * There are 5 main uses of the KnownAssignedXids data structure:
4534 : *
4535 : * * backends taking snapshots - all valid XIDs need to be copied out
4536 : * * backends seeking to determine presence of a specific XID
4537 : * * startup process adding new known-assigned XIDs
4538 : * * startup process removing specific XIDs as transactions end
4539 : * * startup process pruning array when special WAL records arrive
4540 : *
4541 : * This data structure is known to be a hot spot during Hot Standby, so we
4542 : * go to some lengths to make these operations as efficient and as concurrent
4543 : * as possible.
4544 : *
4545 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4546 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4547 : * in general TransactionIdPrecedes would not provide a total order, but
4548 : * we know that the entries present at any instant should not extend across
4549 : * a large enough fraction of XID space to wrap around (the primary would
4550 : * shut down for fear of XID wrap long before that happens). So it's OK to
4551 : * use TransactionIdPrecedes as a binary-search comparator.
4552 : *
4553 : * It's cheap to maintain the sortedness during insertions, since new known
4554 : * XIDs are always reported in XID order; we just append them at the right.
4555 : *
4556 : * To keep individual deletions cheap, we need to allow gaps in the array.
4557 : * This is implemented by marking array elements as valid or invalid using
4558 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4559 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4560 : * XID entry itself. This preserves the property that the XID entries are
4561 : * sorted, so we can do binary searches easily. Periodically we compress
4562 : * out the unused entries; that's much cheaper than having to compress the
4563 : * array immediately on every deletion.
4564 : *
4565 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4566 : * are those with indexes tail <= i < head; items outside this subscript range
4567 : * have unspecified contents. When head reaches the end of the array, we
4568 : * force compression of unused entries rather than wrapping around, since
4569 : * allowing wraparound would greatly complicate the search logic. We maintain
4570 : * an explicit tail pointer so that pruning of old XIDs can be done without
4571 : * immediately moving the array contents. In most cases only a small fraction
4572 : * of the array contains valid entries at any instant.
4573 : *
4574 : * Although only the startup process can ever change the KnownAssignedXids
4575 : * data structure, we still need interlocking so that standby backends will
4576 : * not observe invalid intermediate states. The convention is that backends
4577 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4578 : * the array, the startup process must hold ProcArrayLock exclusively, for
4579 : * the usual transactional reasons (compare commit/abort of a transaction
4580 : * during normal running). Compressing unused entries out of the array
4581 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4582 : * them into slots to the right of the head pointer and then advance the head
4583 : * pointer. This doesn't require any lock at all, but on machines with weak
4584 : * memory ordering, we need to be careful that other processors see the array
4585 : * element changes before they see the head pointer change. We handle this by
4586 : * using memory barriers when reading or writing the head/tail pointers (unless
4587 : * the caller holds ProcArrayLock exclusively).
4588 : *
4589 : * Algorithmic analysis:
4590 : *
4591 : * If we have a maximum of M slots, with N XIDs currently spread across
4592 : * S elements then we have N <= S <= M always.
4593 : *
4594 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4595 : * happen)
4596 : * * Compressing the array is O(S) and requires exclusive lock
4597 : * * Removing an XID is O(logS) and requires exclusive lock
4598 : * * Taking a snapshot is O(S) and requires shared lock
4599 : * * Checking for an XID is O(logS) and requires shared lock
4600 : *
4601 : * In comparison, using a hash table for KnownAssignedXids would mean that
4602 : * taking snapshots would be O(M). If we can maintain S << M then the
4603 : * sorted array technique will deliver significantly faster snapshots.
4604 : * If we try to keep S too small then we will spend too much time compressing,
4605 : * so there is an optimal point for any workload mix. We use a heuristic to
4606 : * decide when to compress the array, though trimming also helps reduce
4607 : * frequency of compressing. The heuristic requires us to track the number of
4608 : * currently valid XIDs in the array (N). Except in special cases, we'll
4609 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4610 : * taking a snapshot to be O(N), which it would have to be anyway.
4611 : */
4612 :
4613 :
4614 : /*
4615 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4616 : * array, removing any gaps.
4617 : *
4618 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4619 : * we do it only if a heuristic indicates it's a good time to do it.
4620 : *
4621 : * Compression requires holding ProcArrayLock in exclusive mode.
4622 : * Caller must pass haveLock = true if it already holds the lock.
4623 : */
4624 : static void
4625 70280 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4626 : {
4627 70280 : ProcArrayStruct *pArray = procArray;
4628 : int head,
4629 : tail,
4630 : nelements;
4631 : int compress_index;
4632 : int i;
4633 :
4634 : /* Counters for compression heuristics */
4635 : static unsigned int transactionEndsCounter;
4636 : static TimestampTz lastCompressTs;
4637 :
4638 : /* Tuning constants */
4639 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4640 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4641 :
4642 : /*
4643 : * Since only the startup process modifies the head/tail pointers, we
4644 : * don't need a lock to read them here.
4645 : */
4646 70280 : head = pArray->headKnownAssignedXids;
4647 70280 : tail = pArray->tailKnownAssignedXids;
4648 70280 : nelements = head - tail;
4649 :
4650 : /*
4651 : * If we can choose whether to compress, use a heuristic to avoid
4652 : * compressing too often or not often enough. "Compress" here simply
4653 : * means moving the values to the beginning of the array, so it is not as
4654 : * complex or costly as typical data compression algorithms.
4655 : */
4656 70280 : if (nelements == pArray->numKnownAssignedXids)
4657 : {
4658 : /*
4659 : * When there are no gaps between head and tail, don't bother to
4660 : * compress, except in the KAX_NO_SPACE case where we must compress to
4661 : * create some space after the head.
4662 : */
4663 36094 : if (reason != KAX_NO_SPACE)
4664 36094 : return;
4665 : }
4666 34186 : else if (reason == KAX_TRANSACTION_END)
4667 : {
4668 : /*
4669 : * Consider compressing only once every so many commits. Frequency
4670 : * determined by benchmarks.
4671 : */
4672 22082 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4673 21892 : return;
4674 :
4675 : /*
4676 : * Furthermore, compress only if the used part of the array is less
4677 : * than 50% full (see comments above).
4678 : */
4679 190 : if (nelements < 2 * pArray->numKnownAssignedXids)
4680 8 : return;
4681 : }
4682 12104 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4683 : {
4684 : /*
4685 : * We're about to go idle for lack of new WAL, so we might as well
4686 : * compress. But not too often, to avoid ProcArray lock contention
4687 : * with readers.
4688 : */
4689 12082 : if (lastCompressTs != 0)
4690 : {
4691 : TimestampTz compress_after;
4692 :
4693 12082 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4694 : KAX_COMPRESS_IDLE_INTERVAL);
4695 12082 : if (GetCurrentTimestamp() < compress_after)
4696 12058 : return;
4697 : }
4698 : }
4699 :
4700 : /* Need to compress, so get the lock if we don't have it. */
4701 228 : if (!haveLock)
4702 24 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4703 :
4704 : /*
4705 : * We compress the array by reading the valid values from tail to head,
4706 : * re-aligning data to 0th element.
4707 : */
4708 228 : compress_index = 0;
4709 10624 : for (i = tail; i < head; i++)
4710 : {
4711 10396 : if (KnownAssignedXidsValid[i])
4712 : {
4713 536 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4714 536 : KnownAssignedXidsValid[compress_index] = true;
4715 536 : compress_index++;
4716 : }
4717 : }
4718 : Assert(compress_index == pArray->numKnownAssignedXids);
4719 :
4720 228 : pArray->tailKnownAssignedXids = 0;
4721 228 : pArray->headKnownAssignedXids = compress_index;
4722 :
4723 228 : if (!haveLock)
4724 24 : LWLockRelease(ProcArrayLock);
4725 :
4726 : /* Update timestamp for maintenance. No need to hold lock for this. */
4727 228 : lastCompressTs = GetCurrentTimestamp();
4728 : }
4729 :
4730 : /*
4731 : * Add xids into KnownAssignedXids at the head of the array.
4732 : *
4733 : * xids from from_xid to to_xid, inclusive, are added to the array.
4734 : *
4735 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4736 : * exclusive mode, so we need no extra locking here. Else caller holds no
4737 : * lock, so we need to be sure we maintain sufficient interlocks against
4738 : * concurrent readers. (Only the startup process ever calls this, so no need
4739 : * to worry about concurrent writers.)
4740 : */
4741 : static void
4742 42464 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4743 : bool exclusive_lock)
4744 : {
4745 42464 : ProcArrayStruct *pArray = procArray;
4746 : TransactionId next_xid;
4747 : int head,
4748 : tail;
4749 : int nxids;
4750 : int i;
4751 :
4752 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4753 :
4754 : /*
4755 : * Calculate how many array slots we'll need. Normally this is cheap; in
4756 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4757 : * way.
4758 : */
4759 42464 : if (to_xid >= from_xid)
4760 42464 : nxids = to_xid - from_xid + 1;
4761 : else
4762 : {
4763 0 : nxids = 1;
4764 0 : next_xid = from_xid;
4765 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4766 : {
4767 0 : nxids++;
4768 0 : TransactionIdAdvance(next_xid);
4769 : }
4770 : }
4771 :
4772 : /*
4773 : * Since only the startup process modifies the head/tail pointers, we
4774 : * don't need a lock to read them here.
4775 : */
4776 42464 : head = pArray->headKnownAssignedXids;
4777 42464 : tail = pArray->tailKnownAssignedXids;
4778 :
4779 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4780 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4781 :
4782 : /*
4783 : * Verify that insertions occur in TransactionId sequence. Note that even
4784 : * if the last existing element is marked invalid, it must still have a
4785 : * correctly sequenced XID value.
4786 : */
4787 69282 : if (head > tail &&
4788 26818 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4789 : {
4790 0 : KnownAssignedXidsDisplay(LOG);
4791 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4792 : }
4793 :
4794 : /*
4795 : * If our xids won't fit in the remaining space, compress out free space
4796 : */
4797 42464 : if (head + nxids > pArray->maxKnownAssignedXids)
4798 : {
4799 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4800 :
4801 0 : head = pArray->headKnownAssignedXids;
4802 : /* note: we no longer care about the tail pointer */
4803 :
4804 : /*
4805 : * If it still won't fit then we're out of memory
4806 : */
4807 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4808 0 : elog(ERROR, "too many KnownAssignedXids");
4809 : }
4810 :
4811 : /* Now we can insert the xids into the space starting at head */
4812 42464 : next_xid = from_xid;
4813 85418 : for (i = 0; i < nxids; i++)
4814 : {
4815 42954 : KnownAssignedXids[head] = next_xid;
4816 42954 : KnownAssignedXidsValid[head] = true;
4817 42954 : TransactionIdAdvance(next_xid);
4818 42954 : head++;
4819 : }
4820 :
4821 : /* Adjust count of number of valid entries */
4822 42464 : pArray->numKnownAssignedXids += nxids;
4823 :
4824 : /*
4825 : * Now update the head pointer. We use a write barrier to ensure that
4826 : * other processors see the above array updates before they see the head
4827 : * pointer change. The barrier isn't required if we're holding
4828 : * ProcArrayLock exclusively.
4829 : */
4830 42464 : if (!exclusive_lock)
4831 42458 : pg_write_barrier();
4832 :
4833 42464 : pArray->headKnownAssignedXids = head;
4834 42464 : }
4835 :
4836 : /*
4837 : * KnownAssignedXidsSearch
4838 : *
4839 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4840 : * Returns true if it was found, false if not.
4841 : *
4842 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4843 : * Exclusive lock must be held for remove = true.
4844 : */
4845 : static bool
4846 45226 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4847 : {
4848 45226 : ProcArrayStruct *pArray = procArray;
4849 : int first,
4850 : last;
4851 : int head;
4852 : int tail;
4853 45226 : int result_index = -1;
4854 :
4855 45226 : tail = pArray->tailKnownAssignedXids;
4856 45226 : head = pArray->headKnownAssignedXids;
4857 :
4858 : /*
4859 : * Only the startup process removes entries, so we don't need the read
4860 : * barrier in that case.
4861 : */
4862 45226 : if (!remove)
4863 0 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4864 :
4865 : /*
4866 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4867 : * array here, since even invalid entries will contain sorted XIDs.
4868 : */
4869 45226 : first = tail;
4870 45226 : last = head - 1;
4871 148624 : while (first <= last)
4872 : {
4873 : int mid_index;
4874 : TransactionId mid_xid;
4875 :
4876 146278 : mid_index = (first + last) / 2;
4877 146278 : mid_xid = KnownAssignedXids[mid_index];
4878 :
4879 146278 : if (xid == mid_xid)
4880 : {
4881 42880 : result_index = mid_index;
4882 42880 : break;
4883 : }
4884 103398 : else if (TransactionIdPrecedes(xid, mid_xid))
4885 21418 : last = mid_index - 1;
4886 : else
4887 81980 : first = mid_index + 1;
4888 : }
4889 :
4890 45226 : if (result_index < 0)
4891 2346 : return false; /* not in array */
4892 :
4893 42880 : if (!KnownAssignedXidsValid[result_index])
4894 90 : return false; /* in array, but invalid */
4895 :
4896 42790 : if (remove)
4897 : {
4898 42790 : KnownAssignedXidsValid[result_index] = false;
4899 :
4900 42790 : pArray->numKnownAssignedXids--;
4901 : Assert(pArray->numKnownAssignedXids >= 0);
4902 :
4903 : /*
4904 : * If we're removing the tail element then advance tail pointer over
4905 : * any invalid elements. This will speed future searches.
4906 : */
4907 42790 : if (result_index == tail)
4908 : {
4909 18770 : tail++;
4910 32930 : while (tail < head && !KnownAssignedXidsValid[tail])
4911 14160 : tail++;
4912 18770 : if (tail >= head)
4913 : {
4914 : /* Array is empty, so we can reset both pointers */
4915 15632 : pArray->headKnownAssignedXids = 0;
4916 15632 : pArray->tailKnownAssignedXids = 0;
4917 : }
4918 : else
4919 : {
4920 3138 : pArray->tailKnownAssignedXids = tail;
4921 : }
4922 : }
4923 : }
4924 :
4925 42790 : return true;
4926 : }
4927 :
4928 : /*
4929 : * Is the specified XID present in KnownAssignedXids[]?
4930 : *
4931 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4932 : */
4933 : static bool
4934 0 : KnownAssignedXidExists(TransactionId xid)
4935 : {
4936 : Assert(TransactionIdIsValid(xid));
4937 :
4938 0 : return KnownAssignedXidsSearch(xid, false);
4939 : }
4940 :
4941 : /*
4942 : * Remove the specified XID from KnownAssignedXids[].
4943 : *
4944 : * Caller must hold ProcArrayLock in exclusive mode.
4945 : */
4946 : static void
4947 45226 : KnownAssignedXidsRemove(TransactionId xid)
4948 : {
4949 : Assert(TransactionIdIsValid(xid));
4950 :
4951 45226 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
4952 :
4953 : /*
4954 : * Note: we cannot consider it an error to remove an XID that's not
4955 : * present. We intentionally remove subxact IDs while processing
4956 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
4957 : * removed again when the top-level xact commits or aborts.
4958 : *
4959 : * It might be possible to track such XIDs to distinguish this case from
4960 : * actual errors, but it would be complicated and probably not worth it.
4961 : * So, just ignore the search result.
4962 : */
4963 45226 : (void) KnownAssignedXidsSearch(xid, true);
4964 45226 : }
4965 :
4966 : /*
4967 : * KnownAssignedXidsRemoveTree
4968 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
4969 : *
4970 : * Caller must hold ProcArrayLock in exclusive mode.
4971 : */
4972 : static void
4973 39932 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
4974 : TransactionId *subxids)
4975 : {
4976 : int i;
4977 :
4978 39932 : if (TransactionIdIsValid(xid))
4979 39890 : KnownAssignedXidsRemove(xid);
4980 :
4981 45268 : for (i = 0; i < nsubxids; i++)
4982 5336 : KnownAssignedXidsRemove(subxids[i]);
4983 :
4984 : /* Opportunistically compress the array */
4985 39932 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
4986 39932 : }
4987 :
4988 : /*
4989 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
4990 : * then clear the whole table.
4991 : *
4992 : * Caller must hold ProcArrayLock in exclusive mode.
4993 : */
4994 : static void
4995 692 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
4996 : {
4997 692 : ProcArrayStruct *pArray = procArray;
4998 692 : int count = 0;
4999 : int head,
5000 : tail,
5001 : i;
5002 :
5003 692 : if (!TransactionIdIsValid(removeXid))
5004 : {
5005 186 : elog(DEBUG4, "removing all KnownAssignedXids");
5006 186 : pArray->numKnownAssignedXids = 0;
5007 186 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5008 186 : return;
5009 : }
5010 :
5011 506 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5012 :
5013 : /*
5014 : * Mark entries invalid starting at the tail. Since array is sorted, we
5015 : * can stop as soon as we reach an entry >= removeXid.
5016 : */
5017 506 : tail = pArray->tailKnownAssignedXids;
5018 506 : head = pArray->headKnownAssignedXids;
5019 :
5020 506 : for (i = tail; i < head; i++)
5021 : {
5022 90 : if (KnownAssignedXidsValid[i])
5023 : {
5024 90 : TransactionId knownXid = KnownAssignedXids[i];
5025 :
5026 90 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5027 90 : break;
5028 :
5029 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5030 : {
5031 0 : KnownAssignedXidsValid[i] = false;
5032 0 : count++;
5033 : }
5034 : }
5035 : }
5036 :
5037 506 : pArray->numKnownAssignedXids -= count;
5038 : Assert(pArray->numKnownAssignedXids >= 0);
5039 :
5040 : /*
5041 : * Advance the tail pointer if we've marked the tail item invalid.
5042 : */
5043 506 : for (i = tail; i < head; i++)
5044 : {
5045 90 : if (KnownAssignedXidsValid[i])
5046 90 : break;
5047 : }
5048 506 : if (i >= head)
5049 : {
5050 : /* Array is empty, so we can reset both pointers */
5051 416 : pArray->headKnownAssignedXids = 0;
5052 416 : pArray->tailKnownAssignedXids = 0;
5053 : }
5054 : else
5055 : {
5056 90 : pArray->tailKnownAssignedXids = i;
5057 : }
5058 :
5059 : /* Opportunistically compress the array */
5060 506 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5061 : }
5062 :
5063 : /*
5064 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5065 : * We filter out anything >= xmax.
5066 : *
5067 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5068 : * that array is large enough.
5069 : *
5070 : * Caller must hold ProcArrayLock in (at least) shared mode.
5071 : */
5072 : static int
5073 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5074 : {
5075 0 : TransactionId xtmp = InvalidTransactionId;
5076 :
5077 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5078 : }
5079 :
5080 : /*
5081 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5082 : * we reduce *xmin to the lowest xid value seen if not already lower.
5083 : *
5084 : * Caller must hold ProcArrayLock in (at least) shared mode.
5085 : */
5086 : static int
5087 1988 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5088 : TransactionId xmax)
5089 : {
5090 1988 : int count = 0;
5091 : int head,
5092 : tail;
5093 : int i;
5094 :
5095 : /*
5096 : * Fetch head just once, since it may change while we loop. We can stop
5097 : * once we reach the initially seen head, since we are certain that an xid
5098 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5099 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5100 : * anyway.
5101 : */
5102 1988 : tail = procArray->tailKnownAssignedXids;
5103 1988 : head = procArray->headKnownAssignedXids;
5104 :
5105 1988 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5106 :
5107 2032 : for (i = tail; i < head; i++)
5108 : {
5109 : /* Skip any gaps in the array */
5110 140 : if (KnownAssignedXidsValid[i])
5111 : {
5112 120 : TransactionId knownXid = KnownAssignedXids[i];
5113 :
5114 : /*
5115 : * Update xmin if required. Only the first XID need be checked,
5116 : * since the array is sorted.
5117 : */
5118 240 : if (count == 0 &&
5119 120 : TransactionIdPrecedes(knownXid, *xmin))
5120 24 : *xmin = knownXid;
5121 :
5122 : /*
5123 : * Filter out anything >= xmax, again relying on sorted property
5124 : * of array.
5125 : */
5126 240 : if (TransactionIdIsValid(xmax) &&
5127 120 : TransactionIdFollowsOrEquals(knownXid, xmax))
5128 96 : break;
5129 :
5130 : /* Add knownXid into output array */
5131 24 : xarray[count++] = knownXid;
5132 : }
5133 : }
5134 :
5135 1988 : return count;
5136 : }
5137 :
5138 : /*
5139 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5140 : * if nothing there.
5141 : */
5142 : static TransactionId
5143 356 : KnownAssignedXidsGetOldestXmin(void)
5144 : {
5145 : int head,
5146 : tail;
5147 : int i;
5148 :
5149 : /*
5150 : * Fetch head just once, since it may change while we loop.
5151 : */
5152 356 : tail = procArray->tailKnownAssignedXids;
5153 356 : head = procArray->headKnownAssignedXids;
5154 :
5155 356 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5156 :
5157 356 : for (i = tail; i < head; i++)
5158 : {
5159 : /* Skip any gaps in the array */
5160 12 : if (KnownAssignedXidsValid[i])
5161 12 : return KnownAssignedXids[i];
5162 : }
5163 :
5164 344 : return InvalidTransactionId;
5165 : }
5166 :
5167 : /*
5168 : * Display KnownAssignedXids to provide debug trail
5169 : *
5170 : * Currently this is only called within startup process, so we need no
5171 : * special locking.
5172 : *
5173 : * Note this is pretty expensive, and much of the expense will be incurred
5174 : * even if the elog message will get discarded. It's not currently called
5175 : * in any performance-critical places, however, so no need to be tenser.
5176 : */
5177 : static void
5178 192 : KnownAssignedXidsDisplay(int trace_level)
5179 : {
5180 192 : ProcArrayStruct *pArray = procArray;
5181 : StringInfoData buf;
5182 : int head,
5183 : tail,
5184 : i;
5185 192 : int nxids = 0;
5186 :
5187 192 : tail = pArray->tailKnownAssignedXids;
5188 192 : head = pArray->headKnownAssignedXids;
5189 :
5190 192 : initStringInfo(&buf);
5191 :
5192 204 : for (i = tail; i < head; i++)
5193 : {
5194 12 : if (KnownAssignedXidsValid[i])
5195 : {
5196 12 : nxids++;
5197 12 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5198 : }
5199 : }
5200 :
5201 192 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5202 : nxids,
5203 : pArray->numKnownAssignedXids,
5204 : pArray->tailKnownAssignedXids,
5205 : pArray->headKnownAssignedXids,
5206 : buf.data);
5207 :
5208 192 : pfree(buf.data);
5209 192 : }
5210 :
5211 : /*
5212 : * KnownAssignedXidsReset
5213 : * Resets KnownAssignedXids to be empty
5214 : */
5215 : static void
5216 0 : KnownAssignedXidsReset(void)
5217 : {
5218 0 : ProcArrayStruct *pArray = procArray;
5219 :
5220 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5221 :
5222 0 : pArray->numKnownAssignedXids = 0;
5223 0 : pArray->tailKnownAssignedXids = 0;
5224 0 : pArray->headKnownAssignedXids = 0;
5225 :
5226 0 : LWLockRelease(ProcArrayLock);
5227 0 : }
|