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) 1996-2026, PostgreSQL Global Development Group
38 : * Portions Copyright (c) 1994, Regents of the University of California
39 : *
40 : *
41 : * IDENTIFICATION
42 : * src/backend/storage/ipc/procarray.c
43 : *
44 : *-------------------------------------------------------------------------
45 : */
46 : #include "postgres.h"
47 :
48 : #include <signal.h>
49 :
50 : #include "access/subtrans.h"
51 : #include "access/transam.h"
52 : #include "access/twophase.h"
53 : #include "access/xact.h"
54 : #include "access/xlogutils.h"
55 : #include "catalog/catalog.h"
56 : #include "catalog/pg_authid.h"
57 : #include "miscadmin.h"
58 : #include "pgstat.h"
59 : #include "postmaster/bgworker.h"
60 : #include "port/pg_lfind.h"
61 : #include "storage/proc.h"
62 : #include "storage/procarray.h"
63 : #include "storage/procsignal.h"
64 : #include "utils/acl.h"
65 : #include "utils/builtins.h"
66 : #include "utils/injection_point.h"
67 : #include "utils/lsyscache.h"
68 : #include "utils/rel.h"
69 : #include "utils/snapmgr.h"
70 :
71 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
72 :
73 : /* Our shared memory area */
74 : typedef struct ProcArrayStruct
75 : {
76 : int numProcs; /* number of valid procs entries */
77 : int maxProcs; /* allocated size of procs array */
78 :
79 : /*
80 : * Known assigned XIDs handling
81 : */
82 : int maxKnownAssignedXids; /* allocated size of array */
83 : int numKnownAssignedXids; /* current # of valid entries */
84 : int tailKnownAssignedXids; /* index of oldest valid element */
85 : int headKnownAssignedXids; /* index of newest element, + 1 */
86 :
87 : /*
88 : * Highest subxid that has been removed from KnownAssignedXids array to
89 : * prevent overflow; or InvalidTransactionId if none. We track this for
90 : * similar reasons to tracking overflowing cached subxids in PGPROC
91 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
92 : * lock to read it.
93 : */
94 : TransactionId lastOverflowedXid;
95 :
96 : /* oldest xmin of any replication slot */
97 : TransactionId replication_slot_xmin;
98 : /* oldest catalog xmin of any replication slot */
99 : TransactionId replication_slot_catalog_xmin;
100 :
101 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
102 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
103 : } ProcArrayStruct;
104 :
105 : /*
106 : * State for the GlobalVisTest* family of functions. Those functions can
107 : * e.g. be used to decide if a deleted row can be removed without violating
108 : * MVCC semantics: If the deleted row's xmax is not considered to be running
109 : * by anyone, the row can be removed.
110 : *
111 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
112 : * cutoff XID while building a snapshot (looking at the frequently changing
113 : * xmins scales badly). Instead we compute two boundaries while building the
114 : * snapshot:
115 : *
116 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
117 : * definitely_needed are definitely still visible.
118 : *
119 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
120 : * definitely be removed
121 : *
122 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
123 : * && XID < definitely_needed), the boundaries can be recomputed (using
124 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
125 : * maintaining an accurate value all the time.
126 : *
127 : * As it is not cheap to compute accurate boundaries, we limit the number of
128 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
129 : *
130 : *
131 : * There are three backend lifetime instances of this struct, optimized for
132 : * different types of relations. As e.g. a normal user defined table in one
133 : * database is inaccessible to backends connected to another database, a test
134 : * specific to a relation can be more aggressive than a test for a shared
135 : * relation. Currently we track four different states:
136 : *
137 : * 1) GlobalVisSharedRels, which only considers an XID's
138 : * effects visible-to-everyone if neither snapshots in any database, nor a
139 : * replication slot's xmin, nor a replication slot's catalog_xmin might
140 : * still consider XID as running.
141 : *
142 : * 2) GlobalVisCatalogRels, which only considers an XID's
143 : * effects visible-to-everyone if neither snapshots in the current
144 : * database, nor a replication slot's xmin, nor a replication slot's
145 : * catalog_xmin might still consider XID as running.
146 : *
147 : * I.e. the difference to GlobalVisSharedRels is that
148 : * snapshot in other databases are ignored.
149 : *
150 : * 3) GlobalVisDataRels, which only considers an XID's
151 : * effects visible-to-everyone if neither snapshots in the current
152 : * database, nor a replication slot's xmin consider XID as running.
153 : *
154 : * I.e. the difference to GlobalVisCatalogRels is that
155 : * replication slot's catalog_xmin is not taken into account.
156 : *
157 : * 4) GlobalVisTempRels, which only considers the current session, as temp
158 : * tables are not visible to other sessions.
159 : *
160 : * GlobalVisTestFor(relation) returns the appropriate state
161 : * for the relation.
162 : *
163 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
164 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
165 : * prevent maybe_needed to become old enough after the GetSnapshotData()
166 : * call.
167 : *
168 : * The typedef is in the header.
169 : */
170 : struct GlobalVisState
171 : {
172 : /* XIDs >= are considered running by some backend */
173 : FullTransactionId definitely_needed;
174 :
175 : /* XIDs < are not considered to be running by any backend */
176 : FullTransactionId maybe_needed;
177 : };
178 :
179 : /*
180 : * Result of ComputeXidHorizons().
181 : */
182 : typedef struct ComputeXidHorizonsResult
183 : {
184 : /*
185 : * The value of TransamVariables->latestCompletedXid when
186 : * ComputeXidHorizons() held ProcArrayLock.
187 : */
188 : FullTransactionId latest_completed;
189 :
190 : /*
191 : * The same for procArray->replication_slot_xmin and
192 : * procArray->replication_slot_catalog_xmin.
193 : */
194 : TransactionId slot_xmin;
195 : TransactionId slot_catalog_xmin;
196 :
197 : /*
198 : * Oldest xid that any backend might still consider running. This needs to
199 : * include processes running VACUUM, in contrast to the normal visibility
200 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
201 : * determining visibility, but doesn't care about rows above its xmin to
202 : * be removed.
203 : *
204 : * This likely should only be needed to determine whether pg_subtrans can
205 : * be truncated. It currently includes the effects of replication slots,
206 : * for historical reasons. But that could likely be changed.
207 : */
208 : TransactionId oldest_considered_running;
209 :
210 : /*
211 : * Oldest xid for which deleted tuples need to be retained in shared
212 : * tables.
213 : *
214 : * This includes the effects of replication slots. If that's not desired,
215 : * look at shared_oldest_nonremovable_raw;
216 : */
217 : TransactionId shared_oldest_nonremovable;
218 :
219 : /*
220 : * Oldest xid that may be necessary to retain in shared tables. This is
221 : * the same as shared_oldest_nonremovable, except that is not affected by
222 : * replication slot's catalog_xmin.
223 : *
224 : * This is mainly useful to be able to send the catalog_xmin to upstream
225 : * streaming replication servers via hot_standby_feedback, so they can
226 : * apply the limit only when accessing catalog tables.
227 : */
228 : TransactionId shared_oldest_nonremovable_raw;
229 :
230 : /*
231 : * Oldest xid for which deleted tuples need to be retained in non-shared
232 : * catalog tables.
233 : */
234 : TransactionId catalog_oldest_nonremovable;
235 :
236 : /*
237 : * Oldest xid for which deleted tuples need to be retained in normal user
238 : * defined tables.
239 : */
240 : TransactionId data_oldest_nonremovable;
241 :
242 : /*
243 : * Oldest xid for which deleted tuples need to be retained in this
244 : * session's temporary tables.
245 : */
246 : TransactionId temp_oldest_nonremovable;
247 : } ComputeXidHorizonsResult;
248 :
249 : /*
250 : * Return value for GlobalVisHorizonKindForRel().
251 : */
252 : typedef enum GlobalVisHorizonKind
253 : {
254 : VISHORIZON_SHARED,
255 : VISHORIZON_CATALOG,
256 : VISHORIZON_DATA,
257 : VISHORIZON_TEMP,
258 : } GlobalVisHorizonKind;
259 :
260 : /*
261 : * Reason codes for KnownAssignedXidsCompress().
262 : */
263 : typedef enum KAXCompressReason
264 : {
265 : KAX_NO_SPACE, /* need to free up space at array end */
266 : KAX_PRUNE, /* we just pruned old entries */
267 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
268 : KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
269 : } KAXCompressReason;
270 :
271 :
272 : static ProcArrayStruct *procArray;
273 :
274 : static PGPROC *allProcs;
275 :
276 : /*
277 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
278 : */
279 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
280 :
281 : /*
282 : * Bookkeeping for tracking emulated transactions in recovery
283 : */
284 : static TransactionId *KnownAssignedXids;
285 : static bool *KnownAssignedXidsValid;
286 : static TransactionId latestObservedXid = InvalidTransactionId;
287 :
288 : /*
289 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
290 : * the highest xid that might still be running that we don't have in
291 : * KnownAssignedXids.
292 : */
293 : static TransactionId standbySnapshotPendingXmin;
294 :
295 : /*
296 : * State for visibility checks on different types of relations. See struct
297 : * GlobalVisState for details. As shared, catalog, normal and temporary
298 : * relations can have different horizons, one such state exists for each.
299 : */
300 : static GlobalVisState GlobalVisSharedRels;
301 : static GlobalVisState GlobalVisCatalogRels;
302 : static GlobalVisState GlobalVisDataRels;
303 : static GlobalVisState GlobalVisTempRels;
304 :
305 : /*
306 : * This backend's RecentXmin at the last time the accurate xmin horizon was
307 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
308 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
309 : */
310 : static TransactionId ComputeXidHorizonsResultLastXmin;
311 :
312 : #ifdef XIDCACHE_DEBUG
313 :
314 : /* counters for XidCache measurement */
315 : static long xc_by_recent_xmin = 0;
316 : static long xc_by_known_xact = 0;
317 : static long xc_by_my_xact = 0;
318 : static long xc_by_latest_xid = 0;
319 : static long xc_by_main_xid = 0;
320 : static long xc_by_child_xid = 0;
321 : static long xc_by_known_assigned = 0;
322 : static long xc_no_overflow = 0;
323 : static long xc_slow_answer = 0;
324 :
325 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
326 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
327 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
328 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
329 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
330 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
331 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
332 : #define xc_no_overflow_inc() (xc_no_overflow++)
333 : #define xc_slow_answer_inc() (xc_slow_answer++)
334 :
335 : static void DisplayXidCache(void);
336 : #else /* !XIDCACHE_DEBUG */
337 :
338 : #define xc_by_recent_xmin_inc() ((void) 0)
339 : #define xc_by_known_xact_inc() ((void) 0)
340 : #define xc_by_my_xact_inc() ((void) 0)
341 : #define xc_by_latest_xid_inc() ((void) 0)
342 : #define xc_by_main_xid_inc() ((void) 0)
343 : #define xc_by_child_xid_inc() ((void) 0)
344 : #define xc_by_known_assigned_inc() ((void) 0)
345 : #define xc_no_overflow_inc() ((void) 0)
346 : #define xc_slow_answer_inc() ((void) 0)
347 : #endif /* XIDCACHE_DEBUG */
348 :
349 : /* Primitives for KnownAssignedXids array handling for standby */
350 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
351 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
352 : bool exclusive_lock);
353 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
354 : static bool KnownAssignedXidExists(TransactionId xid);
355 : static void KnownAssignedXidsRemove(TransactionId xid);
356 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
357 : TransactionId *subxids);
358 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
359 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
360 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
361 : TransactionId *xmin,
362 : TransactionId xmax);
363 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
364 : static void KnownAssignedXidsDisplay(int trace_level);
365 : static void KnownAssignedXidsReset(void);
366 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
367 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
368 : static void MaintainLatestCompletedXid(TransactionId latestXid);
369 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
370 :
371 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
372 : TransactionId xid);
373 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
374 :
375 : /*
376 : * Report shared-memory space needed by ProcArrayShmemInit
377 : */
378 : Size
379 2127 : ProcArrayShmemSize(void)
380 : {
381 : Size size;
382 :
383 : /* Size of the ProcArray structure itself */
384 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
385 :
386 2127 : size = offsetof(ProcArrayStruct, pgprocnos);
387 2127 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
388 :
389 : /*
390 : * During Hot Standby processing we have a data structure called
391 : * KnownAssignedXids, created in shared memory. Local data structures are
392 : * also created in various backends during GetSnapshotData(),
393 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
394 : * main structures created in those functions must be identically sized,
395 : * since we may at times copy the whole of the data structures around. We
396 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
397 : *
398 : * Ideally we'd only create this structure if we were actually doing hot
399 : * standby in the current run, but we don't know that yet at the time
400 : * shared memory is being set up.
401 : */
402 : #define TOTAL_MAX_CACHED_SUBXIDS \
403 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
404 :
405 2127 : if (EnableHotStandby)
406 : {
407 2115 : size = add_size(size,
408 : mul_size(sizeof(TransactionId),
409 2115 : TOTAL_MAX_CACHED_SUBXIDS));
410 2115 : size = add_size(size,
411 2115 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
412 : }
413 :
414 2127 : return size;
415 : }
416 :
417 : /*
418 : * Initialize the shared PGPROC array during postmaster startup.
419 : */
420 : void
421 1140 : ProcArrayShmemInit(void)
422 : {
423 : bool found;
424 :
425 : /* Create or attach to the ProcArray shared structure */
426 1140 : procArray = (ProcArrayStruct *)
427 1140 : ShmemInitStruct("Proc Array",
428 : add_size(offsetof(ProcArrayStruct, pgprocnos),
429 : mul_size(sizeof(int),
430 1140 : PROCARRAY_MAXPROCS)),
431 : &found);
432 :
433 1140 : if (!found)
434 : {
435 : /*
436 : * We're the first - initialize.
437 : */
438 1140 : procArray->numProcs = 0;
439 1140 : procArray->maxProcs = PROCARRAY_MAXPROCS;
440 1140 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
441 1140 : procArray->numKnownAssignedXids = 0;
442 1140 : procArray->tailKnownAssignedXids = 0;
443 1140 : procArray->headKnownAssignedXids = 0;
444 1140 : procArray->lastOverflowedXid = InvalidTransactionId;
445 1140 : procArray->replication_slot_xmin = InvalidTransactionId;
446 1140 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
447 1140 : TransamVariables->xactCompletionCount = 1;
448 : }
449 :
450 1140 : allProcs = ProcGlobal->allProcs;
451 :
452 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
453 1140 : if (EnableHotStandby)
454 : {
455 1134 : KnownAssignedXids = (TransactionId *)
456 1134 : ShmemInitStruct("KnownAssignedXids",
457 : mul_size(sizeof(TransactionId),
458 1134 : TOTAL_MAX_CACHED_SUBXIDS),
459 : &found);
460 1134 : KnownAssignedXidsValid = (bool *)
461 1134 : ShmemInitStruct("KnownAssignedXidsValid",
462 1134 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
463 : &found);
464 : }
465 1140 : }
466 :
467 : /*
468 : * Add the specified PGPROC to the shared array.
469 : */
470 : void
471 18778 : ProcArrayAdd(PGPROC *proc)
472 : {
473 18778 : int pgprocno = GetNumberFromPGProc(proc);
474 18778 : ProcArrayStruct *arrayP = procArray;
475 : int index;
476 : int movecount;
477 :
478 : /* See ProcGlobal comment explaining why both locks are held */
479 18778 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
480 18778 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
481 :
482 18778 : if (arrayP->numProcs >= arrayP->maxProcs)
483 : {
484 : /*
485 : * Oops, no room. (This really shouldn't happen, since there is a
486 : * fixed supply of PGPROC structs too, and so we should have failed
487 : * earlier.)
488 : */
489 0 : ereport(FATAL,
490 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
491 : errmsg("sorry, too many clients already")));
492 : }
493 :
494 : /*
495 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
496 : * locality of references much better. This is useful while traversing the
497 : * ProcArray because there is an increased likelihood of finding the next
498 : * PGPROC structure in the cache.
499 : *
500 : * Since the occurrence of adding/removing a proc is much lower than the
501 : * access to the ProcArray itself, the overhead should be marginal
502 : */
503 46158 : for (index = 0; index < arrayP->numProcs; index++)
504 : {
505 41609 : int this_procno = arrayP->pgprocnos[index];
506 :
507 : Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
508 : Assert(allProcs[this_procno].pgxactoff == index);
509 :
510 : /* If we have found our right position in the array, break */
511 41609 : if (this_procno > pgprocno)
512 14229 : break;
513 : }
514 :
515 18778 : movecount = arrayP->numProcs - index;
516 18778 : memmove(&arrayP->pgprocnos[index + 1],
517 18778 : &arrayP->pgprocnos[index],
518 : movecount * sizeof(*arrayP->pgprocnos));
519 18778 : memmove(&ProcGlobal->xids[index + 1],
520 18778 : &ProcGlobal->xids[index],
521 : movecount * sizeof(*ProcGlobal->xids));
522 18778 : memmove(&ProcGlobal->subxidStates[index + 1],
523 18778 : &ProcGlobal->subxidStates[index],
524 : movecount * sizeof(*ProcGlobal->subxidStates));
525 18778 : memmove(&ProcGlobal->statusFlags[index + 1],
526 18778 : &ProcGlobal->statusFlags[index],
527 : movecount * sizeof(*ProcGlobal->statusFlags));
528 :
529 18778 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
530 18778 : proc->pgxactoff = index;
531 18778 : ProcGlobal->xids[index] = proc->xid;
532 18778 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
533 18778 : ProcGlobal->statusFlags[index] = proc->statusFlags;
534 :
535 18778 : arrayP->numProcs++;
536 :
537 : /* adjust pgxactoff for all following PGPROCs */
538 18778 : index++;
539 51514 : for (; index < arrayP->numProcs; index++)
540 : {
541 32736 : int procno = arrayP->pgprocnos[index];
542 :
543 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
544 : Assert(allProcs[procno].pgxactoff == index - 1);
545 :
546 32736 : allProcs[procno].pgxactoff = index;
547 : }
548 :
549 : /*
550 : * Release in reversed acquisition order, to reduce frequency of having to
551 : * wait for XidGenLock while holding ProcArrayLock.
552 : */
553 18778 : LWLockRelease(XidGenLock);
554 18778 : LWLockRelease(ProcArrayLock);
555 18778 : }
556 :
557 : /*
558 : * Remove the specified PGPROC from the shared array.
559 : *
560 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
561 : * array, and thus causing it to appear as "not running" anymore. In this
562 : * case we must advance latestCompletedXid. (This is essentially the same
563 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
564 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
565 : * twophase.c depends on the latter.)
566 : */
567 : void
568 18755 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
569 : {
570 18755 : ProcArrayStruct *arrayP = procArray;
571 : int myoff;
572 : int movecount;
573 :
574 : #ifdef XIDCACHE_DEBUG
575 : /* dump stats at backend shutdown, but not prepared-xact end */
576 : if (proc->pid != 0)
577 : DisplayXidCache();
578 : #endif
579 :
580 : /* See ProcGlobal comment explaining why both locks are held */
581 18755 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
582 18755 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
583 :
584 18755 : myoff = proc->pgxactoff;
585 :
586 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
587 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
588 :
589 18755 : if (TransactionIdIsValid(latestXid))
590 : {
591 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
592 :
593 : /* Advance global latestCompletedXid while holding the lock */
594 316 : MaintainLatestCompletedXid(latestXid);
595 :
596 : /* Same with xactCompletionCount */
597 316 : TransamVariables->xactCompletionCount++;
598 :
599 316 : ProcGlobal->xids[myoff] = InvalidTransactionId;
600 316 : ProcGlobal->subxidStates[myoff].overflowed = false;
601 316 : ProcGlobal->subxidStates[myoff].count = 0;
602 : }
603 : else
604 : {
605 : /* Shouldn't be trying to remove a live transaction here */
606 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
607 : }
608 :
609 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
610 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
611 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
612 :
613 18755 : ProcGlobal->statusFlags[myoff] = 0;
614 :
615 : /* Keep the PGPROC array sorted. See notes above */
616 18755 : movecount = arrayP->numProcs - myoff - 1;
617 18755 : memmove(&arrayP->pgprocnos[myoff],
618 18755 : &arrayP->pgprocnos[myoff + 1],
619 : movecount * sizeof(*arrayP->pgprocnos));
620 18755 : memmove(&ProcGlobal->xids[myoff],
621 18755 : &ProcGlobal->xids[myoff + 1],
622 : movecount * sizeof(*ProcGlobal->xids));
623 18755 : memmove(&ProcGlobal->subxidStates[myoff],
624 18755 : &ProcGlobal->subxidStates[myoff + 1],
625 : movecount * sizeof(*ProcGlobal->subxidStates));
626 18755 : memmove(&ProcGlobal->statusFlags[myoff],
627 18755 : &ProcGlobal->statusFlags[myoff + 1],
628 : movecount * sizeof(*ProcGlobal->statusFlags));
629 :
630 18755 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
631 18755 : arrayP->numProcs--;
632 :
633 : /*
634 : * Adjust pgxactoff of following procs for removed PGPROC (note that
635 : * numProcs already has been decremented).
636 : */
637 55267 : for (int index = myoff; index < arrayP->numProcs; index++)
638 : {
639 36512 : int procno = arrayP->pgprocnos[index];
640 :
641 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
642 : Assert(allProcs[procno].pgxactoff - 1 == index);
643 :
644 36512 : allProcs[procno].pgxactoff = index;
645 : }
646 :
647 : /*
648 : * Release in reversed acquisition order, to reduce frequency of having to
649 : * wait for XidGenLock while holding ProcArrayLock.
650 : */
651 18755 : LWLockRelease(XidGenLock);
652 18755 : LWLockRelease(ProcArrayLock);
653 18755 : }
654 :
655 :
656 : /*
657 : * ProcArrayEndTransaction -- mark a transaction as no longer running
658 : *
659 : * This is used interchangeably for commit and abort cases. The transaction
660 : * commit/abort must already be reported to WAL and pg_xact.
661 : *
662 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
663 : * latestXid is the latest Xid among the transaction's main XID and
664 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
665 : * the caller to pass latestXid, instead of computing it from the PGPROC's
666 : * contents, because the subxid information in the PGPROC might be
667 : * incomplete.)
668 : */
669 : void
670 552233 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
671 : {
672 552233 : if (TransactionIdIsValid(latestXid))
673 : {
674 : /*
675 : * We must lock ProcArrayLock while clearing our advertised XID, so
676 : * that we do not exit the set of "running" transactions while someone
677 : * else is taking a snapshot. See discussion in
678 : * src/backend/access/transam/README.
679 : */
680 : Assert(TransactionIdIsValid(proc->xid));
681 :
682 : /*
683 : * If we can immediately acquire ProcArrayLock, we clear our own XID
684 : * and release the lock. If not, use group XID clearing to improve
685 : * efficiency.
686 : */
687 138972 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
688 : {
689 138860 : ProcArrayEndTransactionInternal(proc, latestXid);
690 138860 : LWLockRelease(ProcArrayLock);
691 : }
692 : else
693 112 : ProcArrayGroupClearXid(proc, latestXid);
694 : }
695 : else
696 : {
697 : /*
698 : * If we have no XID, we don't need to lock, since we won't affect
699 : * anyone else's calculation of a snapshot. We might change their
700 : * estimate of global xmin, but that's OK.
701 : */
702 : Assert(!TransactionIdIsValid(proc->xid));
703 : Assert(proc->subxidStatus.count == 0);
704 : Assert(!proc->subxidStatus.overflowed);
705 :
706 413261 : proc->vxid.lxid = InvalidLocalTransactionId;
707 413261 : proc->xmin = InvalidTransactionId;
708 :
709 : /* be sure this is cleared in abort */
710 413261 : proc->delayChkptFlags = 0;
711 :
712 413261 : pg_atomic_write_u32(&proc->pendingRecoveryConflicts, 0);
713 :
714 : /* must be cleared with xid/xmin: */
715 : /* avoid unnecessarily dirtying shared cachelines */
716 413261 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
717 : {
718 : Assert(!LWLockHeldByMe(ProcArrayLock));
719 116764 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
720 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
721 116764 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
722 116764 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
723 116764 : LWLockRelease(ProcArrayLock);
724 : }
725 : }
726 552233 : }
727 :
728 : /*
729 : * Mark a write transaction as no longer running.
730 : *
731 : * We don't do any locking here; caller must handle that.
732 : */
733 : static inline void
734 138972 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
735 : {
736 138972 : int pgxactoff = proc->pgxactoff;
737 :
738 : /*
739 : * Note: we need exclusive lock here because we're going to change other
740 : * processes' PGPROC entries.
741 : */
742 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
743 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
744 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
745 :
746 138972 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
747 138972 : proc->xid = InvalidTransactionId;
748 138972 : proc->vxid.lxid = InvalidLocalTransactionId;
749 138972 : proc->xmin = InvalidTransactionId;
750 :
751 : /* be sure this is cleared in abort */
752 138972 : proc->delayChkptFlags = 0;
753 :
754 138972 : pg_atomic_write_u32(&proc->pendingRecoveryConflicts, 0);
755 :
756 : /* must be cleared with xid/xmin: */
757 : /* avoid unnecessarily dirtying shared cachelines */
758 138972 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
759 : {
760 732 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
761 732 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
762 : }
763 :
764 : /* Clear the subtransaction-XID cache too while holding the lock */
765 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
766 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
767 138972 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
768 : {
769 597 : ProcGlobal->subxidStates[pgxactoff].count = 0;
770 597 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
771 597 : proc->subxidStatus.count = 0;
772 597 : proc->subxidStatus.overflowed = false;
773 : }
774 :
775 : /* Also advance global latestCompletedXid while holding the lock */
776 138972 : MaintainLatestCompletedXid(latestXid);
777 :
778 : /* Same with xactCompletionCount */
779 138972 : TransamVariables->xactCompletionCount++;
780 138972 : }
781 :
782 : /*
783 : * ProcArrayGroupClearXid -- group XID clearing
784 : *
785 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
786 : * commit time, add ourselves to a list of processes that need their XIDs
787 : * cleared. The first process to add itself to the list will acquire
788 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
789 : * on behalf of all group members. This avoids a great deal of contention
790 : * around ProcArrayLock when many processes are trying to commit at once,
791 : * since the lock need not be repeatedly handed off from one committing
792 : * process to the next.
793 : */
794 : static void
795 112 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
796 : {
797 112 : int pgprocno = GetNumberFromPGProc(proc);
798 112 : PROC_HDR *procglobal = ProcGlobal;
799 : uint32 nextidx;
800 : uint32 wakeidx;
801 :
802 : /* We should definitely have an XID to clear. */
803 : Assert(TransactionIdIsValid(proc->xid));
804 :
805 : /* Add ourselves to the list of processes needing a group XID clear. */
806 112 : proc->procArrayGroupMember = true;
807 112 : proc->procArrayGroupMemberXid = latestXid;
808 112 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
809 : while (true)
810 : {
811 112 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
812 :
813 112 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
814 : &nextidx,
815 : (uint32) pgprocno))
816 112 : break;
817 : }
818 :
819 : /*
820 : * If the list was not empty, the leader will clear our XID. It is
821 : * impossible to have followers without a leader because the first process
822 : * that has added itself to the list will always have nextidx as
823 : * INVALID_PROC_NUMBER.
824 : */
825 112 : if (nextidx != INVALID_PROC_NUMBER)
826 : {
827 9 : int extraWaits = 0;
828 :
829 : /* Sleep until the leader clears our XID. */
830 9 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
831 : for (;;)
832 : {
833 : /* acts as a read barrier */
834 9 : PGSemaphoreLock(proc->sem);
835 9 : if (!proc->procArrayGroupMember)
836 9 : break;
837 0 : extraWaits++;
838 : }
839 9 : pgstat_report_wait_end();
840 :
841 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER);
842 :
843 : /* Fix semaphore count for any absorbed wakeups */
844 9 : while (extraWaits-- > 0)
845 0 : PGSemaphoreUnlock(proc->sem);
846 9 : return;
847 : }
848 :
849 : /* We are the leader. Acquire the lock on behalf of everyone. */
850 103 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
851 :
852 : /*
853 : * Now that we've got the lock, clear the list of processes waiting for
854 : * group XID clearing, saving a pointer to the head of the list. Trying
855 : * to pop elements one at a time could lead to an ABA problem.
856 : */
857 103 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
858 : INVALID_PROC_NUMBER);
859 :
860 : /* Remember head of list so we can perform wakeups after dropping lock. */
861 103 : wakeidx = nextidx;
862 :
863 : /* Walk the list and clear all XIDs. */
864 215 : while (nextidx != INVALID_PROC_NUMBER)
865 : {
866 112 : PGPROC *nextproc = &allProcs[nextidx];
867 :
868 112 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
869 :
870 : /* Move to next proc in list. */
871 112 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
872 : }
873 :
874 : /* We're done with the lock now. */
875 103 : LWLockRelease(ProcArrayLock);
876 :
877 : /*
878 : * Now that we've released the lock, go back and wake everybody up. We
879 : * don't do this under the lock so as to keep lock hold times to a
880 : * minimum. The system calls we need to perform to wake other processes
881 : * up are probably much slower than the simple memory writes we did while
882 : * holding the lock.
883 : */
884 215 : while (wakeidx != INVALID_PROC_NUMBER)
885 : {
886 112 : PGPROC *nextproc = &allProcs[wakeidx];
887 :
888 112 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
889 112 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
890 :
891 : /* ensure all previous writes are visible before follower continues. */
892 112 : pg_write_barrier();
893 :
894 112 : nextproc->procArrayGroupMember = false;
895 :
896 112 : if (nextproc != MyProc)
897 9 : PGSemaphoreUnlock(nextproc->sem);
898 : }
899 : }
900 :
901 : /*
902 : * ProcArrayClearTransaction -- clear the transaction fields
903 : *
904 : * This is used after successfully preparing a 2-phase transaction. We are
905 : * not actually reporting the transaction's XID as no longer running --- it
906 : * will still appear as running because the 2PC's gxact is in the ProcArray
907 : * too. We just have to clear out our own PGPROC.
908 : */
909 : void
910 306 : ProcArrayClearTransaction(PGPROC *proc)
911 : {
912 : int pgxactoff;
913 :
914 : /*
915 : * Currently we need to lock ProcArrayLock exclusively here, as we
916 : * increment xactCompletionCount below. We also need it at least in shared
917 : * mode for pgproc->pgxactoff to stay the same below.
918 : *
919 : * We could however, as this action does not actually change anyone's view
920 : * of the set of running XIDs (our entry is duplicate with the gxact that
921 : * has already been inserted into the ProcArray), lower the lock level to
922 : * shared if we were to make xactCompletionCount an atomic variable. But
923 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
924 : * enough for this not to be the bottleneck. If it ever becomes a
925 : * bottleneck it may also be worth considering to combine this with the
926 : * subsequent ProcArrayRemove()
927 : */
928 306 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
929 :
930 306 : pgxactoff = proc->pgxactoff;
931 :
932 306 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
933 306 : proc->xid = InvalidTransactionId;
934 :
935 306 : proc->vxid.lxid = InvalidLocalTransactionId;
936 306 : proc->xmin = InvalidTransactionId;
937 306 : pg_atomic_write_u32(&proc->pendingRecoveryConflicts, 0);
938 :
939 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
940 : Assert(!proc->delayChkptFlags);
941 :
942 : /*
943 : * Need to increment completion count even though transaction hasn't
944 : * really committed yet. The reason for that is that GetSnapshotData()
945 : * omits the xid of the current transaction, thus without the increment we
946 : * otherwise could end up reusing the snapshot later. Which would be bad,
947 : * because it might not count the prepared transaction as running.
948 : */
949 306 : TransamVariables->xactCompletionCount++;
950 :
951 : /* Clear the subtransaction-XID cache too */
952 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
953 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
954 306 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
955 : {
956 101 : ProcGlobal->subxidStates[pgxactoff].count = 0;
957 101 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
958 101 : proc->subxidStatus.count = 0;
959 101 : proc->subxidStatus.overflowed = false;
960 : }
961 :
962 306 : LWLockRelease(ProcArrayLock);
963 306 : }
964 :
965 : /*
966 : * Update TransamVariables->latestCompletedXid to point to latestXid if
967 : * currently older.
968 : */
969 : static void
970 139955 : MaintainLatestCompletedXid(TransactionId latestXid)
971 : {
972 139955 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
973 :
974 : Assert(FullTransactionIdIsValid(cur_latest));
975 : Assert(!RecoveryInProgress());
976 : Assert(LWLockHeldByMe(ProcArrayLock));
977 :
978 139955 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
979 : {
980 126221 : TransamVariables->latestCompletedXid =
981 126221 : FullXidRelativeTo(cur_latest, latestXid);
982 : }
983 :
984 : Assert(IsBootstrapProcessingMode() ||
985 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
986 139955 : }
987 :
988 : /*
989 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
990 : */
991 : static void
992 23315 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
993 : {
994 23315 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
995 : FullTransactionId rel;
996 :
997 : Assert(AmStartupProcess() || !IsUnderPostmaster);
998 : Assert(LWLockHeldByMe(ProcArrayLock));
999 :
1000 : /*
1001 : * Need a FullTransactionId to compare latestXid with. Can't rely on
1002 : * latestCompletedXid to be initialized in recovery. But in recovery it's
1003 : * safe to access nextXid without a lock for the startup process.
1004 : */
1005 23315 : rel = TransamVariables->nextXid;
1006 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1007 :
1008 46516 : if (!FullTransactionIdIsValid(cur_latest) ||
1009 23201 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1010 : {
1011 17470 : TransamVariables->latestCompletedXid =
1012 17470 : FullXidRelativeTo(rel, latestXid);
1013 : }
1014 :
1015 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1016 23315 : }
1017 :
1018 : /*
1019 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1020 : *
1021 : * Remember up to where the startup process initialized the CLOG and subtrans
1022 : * so we can ensure it's initialized gaplessly up to the point where necessary
1023 : * while in recovery.
1024 : */
1025 : void
1026 114 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1027 : {
1028 : Assert(standbyState == STANDBY_INITIALIZED);
1029 : Assert(TransactionIdIsNormal(initializedUptoXID));
1030 :
1031 : /*
1032 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1033 : * to, so we can extend it from that point onwards in
1034 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1035 : * ProcArrayApplyRecoveryInfo().
1036 : */
1037 114 : latestObservedXid = initializedUptoXID;
1038 114 : TransactionIdRetreat(latestObservedXid);
1039 114 : }
1040 :
1041 : /*
1042 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1043 : *
1044 : * Takes us through 3 states: Initialized, Pending and Ready.
1045 : * Normal case is to go all the way to Ready straight away, though there
1046 : * are atypical cases where we need to take it in steps.
1047 : *
1048 : * Use the data about running transactions on the primary to create the initial
1049 : * state of KnownAssignedXids. We also use these records to regularly prune
1050 : * KnownAssignedXids because we know it is possible that some transactions
1051 : * with FATAL errors fail to write abort records, which could cause eventual
1052 : * overflow.
1053 : *
1054 : * See comments for LogStandbySnapshot().
1055 : */
1056 : void
1057 822 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1058 : {
1059 : TransactionId *xids;
1060 : TransactionId advanceNextXid;
1061 : int nxids;
1062 : int i;
1063 :
1064 : Assert(standbyState >= STANDBY_INITIALIZED);
1065 : Assert(TransactionIdIsValid(running->nextXid));
1066 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1067 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1068 :
1069 : /*
1070 : * Remove stale transactions, if any.
1071 : */
1072 822 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1073 :
1074 : /*
1075 : * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1076 : * because we will need it up to date for accessing two-phase transactions
1077 : * in StandbyReleaseOldLocks().
1078 : */
1079 822 : advanceNextXid = running->nextXid;
1080 822 : TransactionIdRetreat(advanceNextXid);
1081 822 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1082 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1083 :
1084 : /*
1085 : * Remove stale locks, if any.
1086 : */
1087 822 : StandbyReleaseOldLocks(running->oldestRunningXid);
1088 :
1089 : /*
1090 : * If our snapshot is already valid, nothing else to do...
1091 : */
1092 822 : if (standbyState == STANDBY_SNAPSHOT_READY)
1093 708 : return;
1094 :
1095 : /*
1096 : * If our initial RunningTransactionsData had an overflowed snapshot then
1097 : * we knew we were missing some subxids from our snapshot. If we continue
1098 : * to see overflowed snapshots then we might never be able to start up, so
1099 : * we make another test to see if our snapshot is now valid. We know that
1100 : * the missing subxids are equal to or earlier than nextXid. After we
1101 : * initialise we continue to apply changes during recovery, so once the
1102 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1103 : * know that we no longer have missing information and can mark the
1104 : * snapshot as valid.
1105 : */
1106 114 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1107 : {
1108 : /*
1109 : * If the snapshot isn't overflowed or if its empty we can reset our
1110 : * pending state and use this snapshot instead.
1111 : */
1112 0 : if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1113 : {
1114 : /*
1115 : * If we have already collected known assigned xids, we need to
1116 : * throw them away before we apply the recovery snapshot.
1117 : */
1118 0 : KnownAssignedXidsReset();
1119 0 : standbyState = STANDBY_INITIALIZED;
1120 : }
1121 : else
1122 : {
1123 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1124 : running->oldestRunningXid))
1125 : {
1126 0 : standbyState = STANDBY_SNAPSHOT_READY;
1127 0 : elog(DEBUG1,
1128 : "recovery snapshots are now enabled");
1129 : }
1130 : else
1131 0 : elog(DEBUG1,
1132 : "recovery snapshot waiting for non-overflowed snapshot or "
1133 : "until oldest active xid on standby is at least %u (now %u)",
1134 : standbySnapshotPendingXmin,
1135 : running->oldestRunningXid);
1136 0 : return;
1137 : }
1138 : }
1139 :
1140 : Assert(standbyState == STANDBY_INITIALIZED);
1141 :
1142 : /*
1143 : * NB: this can be reached at least twice, so make sure new code can deal
1144 : * with that.
1145 : */
1146 :
1147 : /*
1148 : * Nobody else is running yet, but take locks anyhow
1149 : */
1150 114 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1151 :
1152 : /*
1153 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1154 : * sort them first.
1155 : *
1156 : * Some of the new xids are top-level xids and some are subtransactions.
1157 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1158 : * aren't overflowed then all xids will fit in snapshot and so we don't
1159 : * need subtrans. If we later overflow, an xid assignment record will add
1160 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1161 : * don't have enough information to correctly update subtrans anyway.
1162 : */
1163 :
1164 : /*
1165 : * Allocate a temporary array to avoid modifying the array passed as
1166 : * argument.
1167 : */
1168 114 : xids = palloc_array(TransactionId, running->xcnt + running->subxcnt);
1169 :
1170 : /*
1171 : * Add to the temp array any xids which have not already completed.
1172 : */
1173 114 : nxids = 0;
1174 118 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1175 : {
1176 4 : TransactionId xid = running->xids[i];
1177 :
1178 : /*
1179 : * The running-xacts snapshot can contain xids that were still visible
1180 : * in the procarray when the snapshot was taken, but were already
1181 : * WAL-logged as completed. They're not running anymore, so ignore
1182 : * them.
1183 : */
1184 4 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1185 0 : continue;
1186 :
1187 4 : xids[nxids++] = xid;
1188 : }
1189 :
1190 114 : if (nxids > 0)
1191 : {
1192 4 : if (procArray->numKnownAssignedXids != 0)
1193 : {
1194 0 : LWLockRelease(ProcArrayLock);
1195 0 : elog(ERROR, "KnownAssignedXids is not empty");
1196 : }
1197 :
1198 : /*
1199 : * Sort the array so that we can add them safely into
1200 : * KnownAssignedXids.
1201 : *
1202 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1203 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1204 : * come from RUNNING_XACTS, which means there are only normal XIDs
1205 : * from the same epoch, so this is safe.
1206 : */
1207 4 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1208 :
1209 : /*
1210 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1211 : * snapshot may include duplicated xids because of prepared
1212 : * transactions, so ignore them.
1213 : */
1214 8 : for (i = 0; i < nxids; i++)
1215 : {
1216 4 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1217 : {
1218 0 : elog(DEBUG1,
1219 : "found duplicated transaction %u for KnownAssignedXids insertion",
1220 : xids[i]);
1221 0 : continue;
1222 : }
1223 4 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1224 : }
1225 :
1226 4 : KnownAssignedXidsDisplay(DEBUG3);
1227 : }
1228 :
1229 114 : pfree(xids);
1230 :
1231 : /*
1232 : * latestObservedXid is at least set to the point where SUBTRANS was
1233 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1234 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1235 : * subtrans from thereon, up to nextXid - 1.
1236 : *
1237 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1238 : * because we've just added xids to the known assigned xids machinery that
1239 : * haven't gone through RecordKnownAssignedTransactionId().
1240 : */
1241 : Assert(TransactionIdIsNormal(latestObservedXid));
1242 114 : TransactionIdAdvance(latestObservedXid);
1243 228 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1244 : {
1245 0 : ExtendSUBTRANS(latestObservedXid);
1246 0 : TransactionIdAdvance(latestObservedXid);
1247 : }
1248 114 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1249 :
1250 : /* ----------
1251 : * Now we've got the running xids we need to set the global values that
1252 : * are used to track snapshots as they evolve further.
1253 : *
1254 : * - latestCompletedXid which will be the xmax for snapshots
1255 : * - lastOverflowedXid which shows whether snapshots overflow
1256 : * - nextXid
1257 : *
1258 : * If the snapshot overflowed, then we still initialise with what we know,
1259 : * but the recovery snapshot isn't fully valid yet because we know there
1260 : * are some subxids missing. We don't know the specific subxids that are
1261 : * missing, so conservatively assume the last one is latestObservedXid.
1262 : * ----------
1263 : */
1264 114 : if (running->subxid_status == SUBXIDS_MISSING)
1265 : {
1266 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1267 :
1268 0 : standbySnapshotPendingXmin = latestObservedXid;
1269 0 : procArray->lastOverflowedXid = latestObservedXid;
1270 : }
1271 : else
1272 : {
1273 114 : standbyState = STANDBY_SNAPSHOT_READY;
1274 :
1275 114 : standbySnapshotPendingXmin = InvalidTransactionId;
1276 :
1277 : /*
1278 : * If the 'xids' array didn't include all subtransactions, we have to
1279 : * mark any snapshots taken as overflowed.
1280 : */
1281 114 : if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1282 26 : procArray->lastOverflowedXid = latestObservedXid;
1283 : else
1284 : {
1285 : Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
1286 88 : procArray->lastOverflowedXid = InvalidTransactionId;
1287 : }
1288 : }
1289 :
1290 : /*
1291 : * If a transaction wrote a commit record in the gap between taking and
1292 : * logging the snapshot then latestCompletedXid may already be higher than
1293 : * the value from the snapshot, so check before we use the incoming value.
1294 : * It also might not yet be set at all.
1295 : */
1296 114 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1297 :
1298 : /*
1299 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1300 : * nobody can see it yet.
1301 : */
1302 :
1303 114 : LWLockRelease(ProcArrayLock);
1304 :
1305 114 : KnownAssignedXidsDisplay(DEBUG3);
1306 114 : if (standbyState == STANDBY_SNAPSHOT_READY)
1307 114 : elog(DEBUG1, "recovery snapshots are now enabled");
1308 : else
1309 0 : elog(DEBUG1,
1310 : "recovery snapshot waiting for non-overflowed snapshot or "
1311 : "until oldest active xid on standby is at least %u (now %u)",
1312 : standbySnapshotPendingXmin,
1313 : running->oldestRunningXid);
1314 : }
1315 :
1316 : /*
1317 : * ProcArrayApplyXidAssignment
1318 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1319 : */
1320 : void
1321 21 : ProcArrayApplyXidAssignment(TransactionId topxid,
1322 : int nsubxids, TransactionId *subxids)
1323 : {
1324 : TransactionId max_xid;
1325 : int i;
1326 :
1327 : Assert(standbyState >= STANDBY_INITIALIZED);
1328 :
1329 21 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1330 :
1331 : /*
1332 : * Mark all the subtransactions as observed.
1333 : *
1334 : * NOTE: This will fail if the subxid contains too many previously
1335 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1336 : * as the code stands, because xid-assignment records should never contain
1337 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1338 : */
1339 21 : RecordKnownAssignedTransactionIds(max_xid);
1340 :
1341 : /*
1342 : * Notice that we update pg_subtrans with the top-level xid, rather than
1343 : * the parent xid. This is a difference between normal processing and
1344 : * recovery, yet is still correct in all cases. The reason is that
1345 : * subtransaction commit is not marked in clog until commit processing, so
1346 : * all aborted subtransactions have already been clearly marked in clog.
1347 : * As a result we are able to refer directly to the top-level
1348 : * transaction's state rather than skipping through all the intermediate
1349 : * states in the subtransaction tree. This should be the first time we
1350 : * have attempted to SubTransSetParent().
1351 : */
1352 1365 : for (i = 0; i < nsubxids; i++)
1353 1344 : SubTransSetParent(subxids[i], topxid);
1354 :
1355 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1356 21 : if (standbyState == STANDBY_INITIALIZED)
1357 0 : return;
1358 :
1359 : /*
1360 : * Uses same locking as transaction commit
1361 : */
1362 21 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1363 :
1364 : /*
1365 : * Remove subxids from known-assigned-xacts.
1366 : */
1367 21 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1368 :
1369 : /*
1370 : * Advance lastOverflowedXid to be at least the last of these subxids.
1371 : */
1372 21 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1373 21 : procArray->lastOverflowedXid = max_xid;
1374 :
1375 21 : LWLockRelease(ProcArrayLock);
1376 : }
1377 :
1378 : /*
1379 : * TransactionIdIsInProgress -- is given transaction running in some backend
1380 : *
1381 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1382 : * there are four possibilities for finding a running transaction:
1383 : *
1384 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1385 : * by looking at ProcGlobal->xids.
1386 : *
1387 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1388 : * We can find this out cheaply too.
1389 : *
1390 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1391 : * if the Xid is running on the primary.
1392 : *
1393 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1394 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1395 : * This is the slowest way, but sadly it has to be done always if the others
1396 : * failed, unless we see that the cached subxact sets are complete (none have
1397 : * overflowed).
1398 : *
1399 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1400 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1401 : * This buys back some concurrency (and we can't retrieve the main Xids from
1402 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1403 : */
1404 : bool
1405 13225353 : TransactionIdIsInProgress(TransactionId xid)
1406 : {
1407 : static TransactionId *xids = NULL;
1408 : static TransactionId *other_xids;
1409 : XidCacheStatus *other_subxidstates;
1410 13225353 : int nxids = 0;
1411 13225353 : ProcArrayStruct *arrayP = procArray;
1412 : TransactionId topxid;
1413 : TransactionId latestCompletedXid;
1414 : int mypgxactoff;
1415 : int numProcs;
1416 : int j;
1417 :
1418 : /*
1419 : * Don't bother checking a transaction older than RecentXmin; it could not
1420 : * possibly still be running. (Note: in particular, this guarantees that
1421 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1422 : * running.)
1423 : */
1424 13225353 : if (TransactionIdPrecedes(xid, RecentXmin))
1425 : {
1426 : xc_by_recent_xmin_inc();
1427 4650137 : return false;
1428 : }
1429 :
1430 : /*
1431 : * We may have just checked the status of this transaction, so if it is
1432 : * already known to be completed, we can fall out without any access to
1433 : * shared memory.
1434 : */
1435 8575216 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1436 : {
1437 : xc_by_known_xact_inc();
1438 971346 : return false;
1439 : }
1440 :
1441 : /*
1442 : * Also, we can handle our own transaction (and subtransactions) without
1443 : * any access to shared memory.
1444 : */
1445 7603870 : if (TransactionIdIsCurrentTransactionId(xid))
1446 : {
1447 : xc_by_my_xact_inc();
1448 200487 : return true;
1449 : }
1450 :
1451 : /*
1452 : * If first time through, get workspace to remember main XIDs in. We
1453 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1454 : */
1455 7403383 : if (xids == NULL)
1456 : {
1457 : /*
1458 : * In hot standby mode, reserve enough space to hold all xids in the
1459 : * known-assigned list. If we later finish recovery, we no longer need
1460 : * the bigger array, but we don't bother to shrink it.
1461 : */
1462 1560 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1463 :
1464 1560 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1465 1560 : if (xids == NULL)
1466 0 : ereport(ERROR,
1467 : (errcode(ERRCODE_OUT_OF_MEMORY),
1468 : errmsg("out of memory")));
1469 : }
1470 :
1471 7403383 : other_xids = ProcGlobal->xids;
1472 7403383 : other_subxidstates = ProcGlobal->subxidStates;
1473 :
1474 7403383 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1475 :
1476 : /*
1477 : * Now that we have the lock, we can check latestCompletedXid; if the
1478 : * target Xid is after that, it's surely still running.
1479 : */
1480 7403383 : latestCompletedXid =
1481 7403383 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1482 7403383 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1483 : {
1484 1538835 : LWLockRelease(ProcArrayLock);
1485 : xc_by_latest_xid_inc();
1486 1538835 : return true;
1487 : }
1488 :
1489 : /* No shortcuts, gotta grovel through the array */
1490 5864548 : mypgxactoff = MyProc->pgxactoff;
1491 5864548 : numProcs = arrayP->numProcs;
1492 6011579 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1493 : {
1494 : int pgprocno;
1495 : PGPROC *proc;
1496 : TransactionId pxid;
1497 : int pxids;
1498 :
1499 : /* Ignore ourselves --- dealt with it above */
1500 5999467 : if (pgxactoff == mypgxactoff)
1501 13687 : continue;
1502 :
1503 : /* Fetch xid just once - see GetNewTransactionId */
1504 5985780 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1505 :
1506 5985780 : if (!TransactionIdIsValid(pxid))
1507 88661 : continue;
1508 :
1509 : /*
1510 : * Step 1: check the main Xid
1511 : */
1512 5897119 : if (TransactionIdEquals(pxid, xid))
1513 : {
1514 5852302 : LWLockRelease(ProcArrayLock);
1515 : xc_by_main_xid_inc();
1516 5852302 : return true;
1517 : }
1518 :
1519 : /*
1520 : * We can ignore main Xids that are younger than the target Xid, since
1521 : * the target could not possibly be their child.
1522 : */
1523 44817 : if (TransactionIdPrecedes(xid, pxid))
1524 23131 : continue;
1525 :
1526 : /*
1527 : * Step 2: check the cached child-Xids arrays
1528 : */
1529 21686 : pxids = other_subxidstates[pgxactoff].count;
1530 21686 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1531 21686 : pgprocno = arrayP->pgprocnos[pgxactoff];
1532 21686 : proc = &allProcs[pgprocno];
1533 39591 : for (j = pxids - 1; j >= 0; j--)
1534 : {
1535 : /* Fetch xid just once - see GetNewTransactionId */
1536 18039 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1537 :
1538 18039 : if (TransactionIdEquals(cxid, xid))
1539 : {
1540 134 : LWLockRelease(ProcArrayLock);
1541 : xc_by_child_xid_inc();
1542 134 : return true;
1543 : }
1544 : }
1545 :
1546 : /*
1547 : * Save the main Xid for step 4. We only need to remember main Xids
1548 : * that have uncached children. (Note: there is no race condition
1549 : * here because the overflowed flag cannot be cleared, only set, while
1550 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1551 : * worry about.)
1552 : */
1553 21552 : if (other_subxidstates[pgxactoff].overflowed)
1554 213 : xids[nxids++] = pxid;
1555 : }
1556 :
1557 : /*
1558 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1559 : * in the list must be treated as running.
1560 : */
1561 12112 : if (RecoveryInProgress())
1562 : {
1563 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1564 : Assert(nxids == 0);
1565 :
1566 1 : if (KnownAssignedXidExists(xid))
1567 : {
1568 0 : LWLockRelease(ProcArrayLock);
1569 : xc_by_known_assigned_inc();
1570 0 : return true;
1571 : }
1572 :
1573 : /*
1574 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1575 : * too. Fetch all xids from KnownAssignedXids that are lower than
1576 : * xid, since if xid is a subtransaction its parent will always have a
1577 : * lower value. Note we will collect both main and subXIDs here, but
1578 : * there's no help for it.
1579 : */
1580 1 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1581 0 : nxids = KnownAssignedXidsGet(xids, xid);
1582 : }
1583 :
1584 12112 : LWLockRelease(ProcArrayLock);
1585 :
1586 : /*
1587 : * If none of the relevant caches overflowed, we know the Xid is not
1588 : * running without even looking at pg_subtrans.
1589 : */
1590 12112 : if (nxids == 0)
1591 : {
1592 : xc_no_overflow_inc();
1593 11899 : cachedXidIsNotInProgress = xid;
1594 11899 : return false;
1595 : }
1596 :
1597 : /*
1598 : * Step 4: have to check pg_subtrans.
1599 : *
1600 : * At this point, we know it's either a subtransaction of one of the Xids
1601 : * in xids[], or it's not running. If it's an already-failed
1602 : * subtransaction, we want to say "not running" even though its parent may
1603 : * still be running. So first, check pg_xact to see if it's been aborted.
1604 : */
1605 : xc_slow_answer_inc();
1606 :
1607 213 : if (TransactionIdDidAbort(xid))
1608 : {
1609 0 : cachedXidIsNotInProgress = xid;
1610 0 : return false;
1611 : }
1612 :
1613 : /*
1614 : * It isn't aborted, so check whether the transaction tree it belongs to
1615 : * is still running (or, more precisely, whether it was running when we
1616 : * held ProcArrayLock).
1617 : */
1618 213 : topxid = SubTransGetTopmostTransaction(xid);
1619 : Assert(TransactionIdIsValid(topxid));
1620 370 : if (!TransactionIdEquals(topxid, xid) &&
1621 157 : pg_lfind32(topxid, xids, nxids))
1622 157 : return true;
1623 :
1624 56 : cachedXidIsNotInProgress = xid;
1625 56 : return false;
1626 : }
1627 :
1628 :
1629 : /*
1630 : * Determine XID horizons.
1631 : *
1632 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1633 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1634 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1635 : * GlobalVisState).
1636 : *
1637 : * See the definition of ComputeXidHorizonsResult for the various computed
1638 : * horizons.
1639 : *
1640 : * For VACUUM separate horizons (used to decide which deleted tuples must
1641 : * be preserved), for shared and non-shared tables are computed. For shared
1642 : * relations backends in all databases must be considered, but for non-shared
1643 : * relations that's not required, since only backends in my own database could
1644 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1645 : * VACUUMs because (a) they must be working on other tables, and (b) they
1646 : * don't need to do snapshot-based lookups.
1647 : *
1648 : * This also computes a horizon used to truncate pg_subtrans. For that
1649 : * backends in all databases have to be considered, and concurrently running
1650 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1651 : * accesses.
1652 : *
1653 : * Note: we include all currently running xids in the set of considered xids.
1654 : * This ensures that if a just-started xact has not yet set its snapshot,
1655 : * when it does set the snapshot it cannot set xmin less than what we compute.
1656 : * See notes in src/backend/access/transam/README.
1657 : *
1658 : * Note: despite the above, it's possible for the calculated values to move
1659 : * backwards on repeated calls. The calculated values are conservative, so
1660 : * that anything older is definitely not considered as running by anyone
1661 : * anymore, but the exact values calculated depend on a number of things. For
1662 : * example, if there are no transactions running in the current database, the
1663 : * horizon for normal tables will be latestCompletedXid. If a transaction
1664 : * begins after that, its xmin will include in-progress transactions in other
1665 : * databases that started earlier, so another call will return a lower value.
1666 : * Nonetheless it is safe to vacuum a table in the current database with the
1667 : * first result. There are also replication-related effects: a walsender
1668 : * process can set its xmin based on transactions that are no longer running
1669 : * on the primary but are still being replayed on the standby, thus possibly
1670 : * making the values go backwards. In this case there is a possibility that
1671 : * we lose data that the standby would like to have, but unless the standby
1672 : * uses a replication slot to make its xmin persistent there is little we can
1673 : * do about that --- data is only protected if the walsender runs continuously
1674 : * while queries are executed on the standby. (The Hot Standby code deals
1675 : * with such cases by failing standby queries that needed to access
1676 : * already-removed data, so there's no integrity bug.)
1677 : *
1678 : * Note: the approximate horizons (see definition of GlobalVisState) are
1679 : * updated by the computations done here. That's currently required for
1680 : * correctness and a small optimization. Without doing so it's possible that
1681 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1682 : * horizon than later when deciding which tuples can be removed - which the
1683 : * code doesn't expect (breaking HOT).
1684 : */
1685 : static void
1686 206317 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1687 : {
1688 206317 : ProcArrayStruct *arrayP = procArray;
1689 : TransactionId kaxmin;
1690 206317 : bool in_recovery = RecoveryInProgress();
1691 206317 : TransactionId *other_xids = ProcGlobal->xids;
1692 :
1693 : /* inferred after ProcArrayLock is released */
1694 206317 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1695 :
1696 206317 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1697 :
1698 206317 : h->latest_completed = TransamVariables->latestCompletedXid;
1699 :
1700 : /*
1701 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1702 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1703 : * and so protects us against overestimating the result due to future
1704 : * additions.
1705 : */
1706 : {
1707 : TransactionId initial;
1708 :
1709 206317 : initial = XidFromFullTransactionId(h->latest_completed);
1710 : Assert(TransactionIdIsValid(initial));
1711 206317 : TransactionIdAdvance(initial);
1712 :
1713 206317 : h->oldest_considered_running = initial;
1714 206317 : h->shared_oldest_nonremovable = initial;
1715 206317 : h->data_oldest_nonremovable = initial;
1716 :
1717 : /*
1718 : * Only modifications made by this backend affect the horizon for
1719 : * temporary relations. Instead of a check in each iteration of the
1720 : * loop over all PGPROCs it is cheaper to just initialize to the
1721 : * current top-level xid any.
1722 : *
1723 : * Without an assigned xid we could use a horizon as aggressive as
1724 : * GetNewTransactionId(), but we can get away with the much cheaper
1725 : * latestCompletedXid + 1: If this backend has no xid there, by
1726 : * definition, can't be any newer changes in the temp table than
1727 : * latestCompletedXid.
1728 : */
1729 206317 : if (TransactionIdIsValid(MyProc->xid))
1730 34004 : h->temp_oldest_nonremovable = MyProc->xid;
1731 : else
1732 172313 : h->temp_oldest_nonremovable = initial;
1733 : }
1734 :
1735 : /*
1736 : * Fetch slot horizons while ProcArrayLock is held - the
1737 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1738 : * the lock.
1739 : */
1740 206317 : h->slot_xmin = procArray->replication_slot_xmin;
1741 206317 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1742 :
1743 1302964 : for (int index = 0; index < arrayP->numProcs; index++)
1744 : {
1745 1096647 : int pgprocno = arrayP->pgprocnos[index];
1746 1096647 : PGPROC *proc = &allProcs[pgprocno];
1747 1096647 : int8 statusFlags = ProcGlobal->statusFlags[index];
1748 : TransactionId xid;
1749 : TransactionId xmin;
1750 :
1751 : /* Fetch xid just once - see GetNewTransactionId */
1752 1096647 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1753 1096647 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1754 :
1755 : /*
1756 : * Consider both the transaction's Xmin, and its Xid.
1757 : *
1758 : * We must check both because a transaction might have an Xmin but not
1759 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1760 : * some not-yet-set Xmin.
1761 : */
1762 1096647 : xmin = TransactionIdOlder(xmin, xid);
1763 :
1764 : /* if neither is set, this proc doesn't influence the horizon */
1765 1096647 : if (!TransactionIdIsValid(xmin))
1766 462530 : continue;
1767 :
1768 : /*
1769 : * Don't ignore any procs when determining which transactions might be
1770 : * considered running. While slots should ensure logical decoding
1771 : * backends are protected even without this check, it can't hurt to
1772 : * include them here as well..
1773 : */
1774 634117 : h->oldest_considered_running =
1775 634117 : TransactionIdOlder(h->oldest_considered_running, xmin);
1776 :
1777 : /*
1778 : * Skip over backends either vacuuming (which is ok with rows being
1779 : * removed, as long as pg_subtrans is not truncated) or doing logical
1780 : * decoding (which manages xmin separately, check below).
1781 : */
1782 634117 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1783 222561 : continue;
1784 :
1785 : /* shared tables need to take backends in all databases into account */
1786 411556 : h->shared_oldest_nonremovable =
1787 411556 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1788 :
1789 : /*
1790 : * Normally sessions in other databases are ignored for anything but
1791 : * the shared horizon.
1792 : *
1793 : * However, include them when MyDatabaseId is not (yet) set. A
1794 : * backend in the process of starting up must not compute a "too
1795 : * aggressive" horizon, otherwise we could end up using it to prune
1796 : * still-needed data away. If the current backend never connects to a
1797 : * database this is harmless, because data_oldest_nonremovable will
1798 : * never be utilized.
1799 : *
1800 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1801 : * be included. (This flag is used for hot standby feedback, which
1802 : * can't be tied to a specific database.)
1803 : *
1804 : * Also, while in recovery we cannot compute an accurate per-database
1805 : * horizon, as all xids are managed via the KnownAssignedXids
1806 : * machinery.
1807 : */
1808 411556 : if (proc->databaseId == MyDatabaseId ||
1809 19698 : MyDatabaseId == InvalidOid ||
1810 12070 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1811 : in_recovery)
1812 : {
1813 399488 : h->data_oldest_nonremovable =
1814 399488 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1815 : }
1816 : }
1817 :
1818 : /*
1819 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1820 : * after lock is released.
1821 : */
1822 206317 : if (in_recovery)
1823 351 : kaxmin = KnownAssignedXidsGetOldestXmin();
1824 :
1825 : /*
1826 : * No other information from shared state is needed, release the lock
1827 : * immediately. The rest of the computations can be done without a lock.
1828 : */
1829 206317 : LWLockRelease(ProcArrayLock);
1830 :
1831 206317 : if (in_recovery)
1832 : {
1833 351 : h->oldest_considered_running =
1834 351 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1835 351 : h->shared_oldest_nonremovable =
1836 351 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1837 351 : h->data_oldest_nonremovable =
1838 351 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1839 : /* temp relations cannot be accessed in recovery */
1840 : }
1841 :
1842 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1843 : h->shared_oldest_nonremovable));
1844 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1845 : h->data_oldest_nonremovable));
1846 :
1847 : /*
1848 : * Check whether there are replication slots requiring an older xmin.
1849 : */
1850 206317 : h->shared_oldest_nonremovable =
1851 206317 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1852 206317 : h->data_oldest_nonremovable =
1853 206317 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1854 :
1855 : /*
1856 : * The only difference between catalog / data horizons is that the slot's
1857 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1858 : * for logical decoding). Initialize with data horizon, and then back up
1859 : * further if necessary. Have to back up the shared horizon as well, since
1860 : * that also can contain catalogs.
1861 : */
1862 206317 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1863 206317 : h->shared_oldest_nonremovable =
1864 206317 : TransactionIdOlder(h->shared_oldest_nonremovable,
1865 : h->slot_catalog_xmin);
1866 206317 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1867 206317 : h->catalog_oldest_nonremovable =
1868 206317 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1869 : h->slot_catalog_xmin);
1870 :
1871 : /*
1872 : * It's possible that slots backed up the horizons further than
1873 : * oldest_considered_running. Fix.
1874 : */
1875 206317 : h->oldest_considered_running =
1876 206317 : TransactionIdOlder(h->oldest_considered_running,
1877 : h->shared_oldest_nonremovable);
1878 206317 : h->oldest_considered_running =
1879 206317 : TransactionIdOlder(h->oldest_considered_running,
1880 : h->catalog_oldest_nonremovable);
1881 206317 : h->oldest_considered_running =
1882 206317 : TransactionIdOlder(h->oldest_considered_running,
1883 : h->data_oldest_nonremovable);
1884 :
1885 : /*
1886 : * shared horizons have to be at least as old as the oldest visible in
1887 : * current db
1888 : */
1889 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1890 : h->data_oldest_nonremovable));
1891 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1892 : h->catalog_oldest_nonremovable));
1893 :
1894 : /*
1895 : * Horizons need to ensure that pg_subtrans access is still possible for
1896 : * the relevant backends.
1897 : */
1898 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1899 : h->shared_oldest_nonremovable));
1900 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1901 : h->catalog_oldest_nonremovable));
1902 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1903 : h->data_oldest_nonremovable));
1904 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1905 : h->temp_oldest_nonremovable));
1906 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1907 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1908 : h->slot_xmin));
1909 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1910 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1911 : h->slot_catalog_xmin));
1912 :
1913 : /* update approximate horizons with the computed horizons */
1914 206317 : GlobalVisUpdateApply(h);
1915 206317 : }
1916 :
1917 : /*
1918 : * Determine what kind of visibility horizon needs to be used for a
1919 : * relation. If rel is NULL, the most conservative horizon is used.
1920 : */
1921 : static inline GlobalVisHorizonKind
1922 14603155 : GlobalVisHorizonKindForRel(Relation rel)
1923 : {
1924 : /*
1925 : * Other relkinds currently don't contain xids, nor always the necessary
1926 : * logical decoding markers.
1927 : */
1928 : Assert(!rel ||
1929 : rel->rd_rel->relkind == RELKIND_RELATION ||
1930 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1931 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1932 :
1933 14603155 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1934 111834 : return VISHORIZON_SHARED;
1935 14491321 : else if (IsCatalogRelation(rel) ||
1936 11380923 : RelationIsAccessibleInLogicalDecoding(rel))
1937 3110402 : return VISHORIZON_CATALOG;
1938 11380919 : else if (!RELATION_IS_LOCAL(rel))
1939 11326302 : return VISHORIZON_DATA;
1940 : else
1941 54617 : return VISHORIZON_TEMP;
1942 : }
1943 :
1944 : /*
1945 : * Return the oldest XID for which deleted tuples must be preserved in the
1946 : * passed table.
1947 : *
1948 : * If rel is not NULL the horizon may be considerably more recent than
1949 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1950 : * that is correct (but not optimal) for all relations will be returned.
1951 : *
1952 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1953 : * the passed in table.
1954 : */
1955 : TransactionId
1956 149089 : GetOldestNonRemovableTransactionId(Relation rel)
1957 : {
1958 : ComputeXidHorizonsResult horizons;
1959 :
1960 149089 : ComputeXidHorizons(&horizons);
1961 :
1962 149089 : switch (GlobalVisHorizonKindForRel(rel))
1963 : {
1964 22980 : case VISHORIZON_SHARED:
1965 22980 : return horizons.shared_oldest_nonremovable;
1966 91316 : case VISHORIZON_CATALOG:
1967 91316 : return horizons.catalog_oldest_nonremovable;
1968 22136 : case VISHORIZON_DATA:
1969 22136 : return horizons.data_oldest_nonremovable;
1970 12657 : case VISHORIZON_TEMP:
1971 12657 : return horizons.temp_oldest_nonremovable;
1972 : }
1973 :
1974 : /* just to prevent compiler warnings */
1975 0 : return InvalidTransactionId;
1976 : }
1977 :
1978 : /*
1979 : * Return the oldest transaction id any currently running backend might still
1980 : * consider running. This should not be used for visibility / pruning
1981 : * determinations (see GetOldestNonRemovableTransactionId()), but for
1982 : * decisions like up to where pg_subtrans can be truncated.
1983 : */
1984 : TransactionId
1985 1754 : GetOldestTransactionIdConsideredRunning(void)
1986 : {
1987 : ComputeXidHorizonsResult horizons;
1988 :
1989 1754 : ComputeXidHorizons(&horizons);
1990 :
1991 1754 : return horizons.oldest_considered_running;
1992 : }
1993 :
1994 : /*
1995 : * Return the visibility horizons for a hot standby feedback message.
1996 : */
1997 : void
1998 51 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
1999 : {
2000 : ComputeXidHorizonsResult horizons;
2001 :
2002 51 : ComputeXidHorizons(&horizons);
2003 :
2004 : /*
2005 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2006 : * effect of replication slot's catalog_xmin. We want to send a separate
2007 : * feedback for the catalog horizon, so the primary can remove data table
2008 : * contents more aggressively.
2009 : */
2010 51 : *xmin = horizons.shared_oldest_nonremovable_raw;
2011 51 : *catalog_xmin = horizons.slot_catalog_xmin;
2012 51 : }
2013 :
2014 : /*
2015 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2016 : *
2017 : * We have to export this for use by snapmgr.c.
2018 : */
2019 : int
2020 34857 : GetMaxSnapshotXidCount(void)
2021 : {
2022 34857 : return procArray->maxProcs;
2023 : }
2024 :
2025 : /*
2026 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2027 : *
2028 : * We have to export this for use by snapmgr.c.
2029 : */
2030 : int
2031 34666 : GetMaxSnapshotSubxidCount(void)
2032 : {
2033 34666 : return TOTAL_MAX_CACHED_SUBXIDS;
2034 : }
2035 :
2036 : /*
2037 : * Helper function for GetSnapshotData() that checks if the bulk of the
2038 : * visibility information in the snapshot is still valid. If so, it updates
2039 : * the fields that need to change and returns true. Otherwise it returns
2040 : * false.
2041 : *
2042 : * This very likely can be evolved to not need ProcArrayLock held (at very
2043 : * least in the case we already hold a snapshot), but that's for another day.
2044 : */
2045 : static bool
2046 2213418 : GetSnapshotDataReuse(Snapshot snapshot)
2047 : {
2048 : uint64 curXactCompletionCount;
2049 :
2050 : Assert(LWLockHeldByMe(ProcArrayLock));
2051 :
2052 2213418 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2053 34647 : return false;
2054 :
2055 2178771 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2056 2178771 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2057 366467 : return false;
2058 :
2059 : /*
2060 : * If the current xactCompletionCount is still the same as it was at the
2061 : * time the snapshot was built, we can be sure that rebuilding the
2062 : * contents of the snapshot the hard way would result in the same snapshot
2063 : * contents:
2064 : *
2065 : * As explained in transam/README, the set of xids considered running by
2066 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2067 : * contents only depend on transactions with xids and xactCompletionCount
2068 : * is incremented whenever a transaction with an xid finishes (while
2069 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2070 : * ensures we would detect if the snapshot would have changed.
2071 : *
2072 : * As the snapshot contents are the same as it was before, it is safe to
2073 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2074 : * visible under the snapshot could already have been removed (that'd
2075 : * require the set of running transactions to change) and it fulfills the
2076 : * requirement that concurrent GetSnapshotData() calls yield the same
2077 : * xmin.
2078 : */
2079 1812304 : if (!TransactionIdIsValid(MyProc->xmin))
2080 703169 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2081 :
2082 1812304 : RecentXmin = snapshot->xmin;
2083 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2084 :
2085 1812304 : snapshot->curcid = GetCurrentCommandId(false);
2086 1812304 : snapshot->active_count = 0;
2087 1812304 : snapshot->regd_count = 0;
2088 1812304 : snapshot->copied = false;
2089 :
2090 1812304 : return true;
2091 : }
2092 :
2093 : /*
2094 : * GetSnapshotData -- returns information about running transactions.
2095 : *
2096 : * The returned snapshot includes xmin (lowest still-running xact ID),
2097 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2098 : * in the range xmin <= xid < xmax. It is used as follows:
2099 : * All xact IDs < xmin are considered finished.
2100 : * All xact IDs >= xmax are considered still running.
2101 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2102 : * it is considered running or not.
2103 : * This ensures that the set of transactions seen as "running" by the
2104 : * current xact will not change after it takes the snapshot.
2105 : *
2106 : * All running top-level XIDs are included in the snapshot, except for lazy
2107 : * VACUUM processes. We also try to include running subtransaction XIDs,
2108 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2109 : * information may not be available. If we find any overflowed subxid arrays,
2110 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2111 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2112 : *
2113 : * We also update the following backend-global variables:
2114 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2115 : * current transaction (this is the same as MyProc->xmin).
2116 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2117 : * older than this are known not running any more.
2118 : *
2119 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2120 : * for the benefit of the GlobalVisTest* family of functions.
2121 : *
2122 : * Note: this function should probably not be called with an argument that's
2123 : * not statically allocated (see xip allocation below).
2124 : */
2125 : Snapshot
2126 2213418 : GetSnapshotData(Snapshot snapshot)
2127 : {
2128 2213418 : ProcArrayStruct *arrayP = procArray;
2129 2213418 : TransactionId *other_xids = ProcGlobal->xids;
2130 : TransactionId xmin;
2131 : TransactionId xmax;
2132 2213418 : int count = 0;
2133 2213418 : int subcount = 0;
2134 2213418 : bool suboverflowed = false;
2135 : FullTransactionId latest_completed;
2136 : TransactionId oldestxid;
2137 : int mypgxactoff;
2138 : TransactionId myxid;
2139 : uint64 curXactCompletionCount;
2140 :
2141 2213418 : TransactionId replication_slot_xmin = InvalidTransactionId;
2142 2213418 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2143 :
2144 : Assert(snapshot != NULL);
2145 :
2146 : /*
2147 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2148 : * be sufficient. But it seems better to do the malloc while not holding
2149 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2150 : * more subxip storage than is probably needed.
2151 : *
2152 : * This does open a possibility for avoiding repeated malloc/free: since
2153 : * maxProcs does not change at runtime, we can simply reuse the previous
2154 : * xip arrays if any. (This relies on the fact that all callers pass
2155 : * static SnapshotData structs.)
2156 : */
2157 2213418 : if (snapshot->xip == NULL)
2158 : {
2159 : /*
2160 : * First call for this snapshot. Snapshot is same size whether or not
2161 : * we are in recovery, see later comments.
2162 : */
2163 34641 : snapshot->xip = (TransactionId *)
2164 34641 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2165 34641 : if (snapshot->xip == NULL)
2166 0 : ereport(ERROR,
2167 : (errcode(ERRCODE_OUT_OF_MEMORY),
2168 : errmsg("out of memory")));
2169 : Assert(snapshot->subxip == NULL);
2170 34641 : snapshot->subxip = (TransactionId *)
2171 34641 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2172 34641 : if (snapshot->subxip == NULL)
2173 0 : ereport(ERROR,
2174 : (errcode(ERRCODE_OUT_OF_MEMORY),
2175 : errmsg("out of memory")));
2176 : }
2177 :
2178 : /*
2179 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2180 : * going to set MyProc->xmin.
2181 : */
2182 2213418 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2183 :
2184 2213418 : if (GetSnapshotDataReuse(snapshot))
2185 : {
2186 1812304 : LWLockRelease(ProcArrayLock);
2187 1812304 : return snapshot;
2188 : }
2189 :
2190 401114 : latest_completed = TransamVariables->latestCompletedXid;
2191 401114 : mypgxactoff = MyProc->pgxactoff;
2192 401114 : myxid = other_xids[mypgxactoff];
2193 : Assert(myxid == MyProc->xid);
2194 :
2195 401114 : oldestxid = TransamVariables->oldestXid;
2196 401114 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2197 :
2198 : /* xmax is always latestCompletedXid + 1 */
2199 401114 : xmax = XidFromFullTransactionId(latest_completed);
2200 401114 : TransactionIdAdvance(xmax);
2201 : Assert(TransactionIdIsNormal(xmax));
2202 :
2203 : /* initialize xmin calculation with xmax */
2204 401114 : xmin = xmax;
2205 :
2206 : /* take own xid into account, saves a check inside the loop */
2207 401114 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2208 27246 : xmin = myxid;
2209 :
2210 401114 : snapshot->takenDuringRecovery = RecoveryInProgress();
2211 :
2212 401114 : if (!snapshot->takenDuringRecovery)
2213 : {
2214 399834 : int numProcs = arrayP->numProcs;
2215 399834 : TransactionId *xip = snapshot->xip;
2216 399834 : int *pgprocnos = arrayP->pgprocnos;
2217 399834 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2218 399834 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2219 :
2220 : /*
2221 : * First collect set of pgxactoff/xids that need to be included in the
2222 : * snapshot.
2223 : */
2224 3485277 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2225 : {
2226 : /* Fetch xid just once - see GetNewTransactionId */
2227 3085443 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2228 : uint8 statusFlags;
2229 :
2230 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2231 :
2232 : /*
2233 : * If the transaction has no XID assigned, we can skip it; it
2234 : * won't have sub-XIDs either.
2235 : */
2236 3085443 : if (likely(xid == InvalidTransactionId))
2237 2307889 : continue;
2238 :
2239 : /*
2240 : * We don't include our own XIDs (if any) in the snapshot. It
2241 : * needs to be included in the xmin computation, but we did so
2242 : * outside the loop.
2243 : */
2244 777554 : if (pgxactoff == mypgxactoff)
2245 53211 : continue;
2246 :
2247 : /*
2248 : * The only way we are able to get here with a non-normal xid is
2249 : * during bootstrap - with this backend using
2250 : * BootstrapTransactionId. But the above test should filter that
2251 : * out.
2252 : */
2253 : Assert(TransactionIdIsNormal(xid));
2254 :
2255 : /*
2256 : * If the XID is >= xmax, we can skip it; such transactions will
2257 : * be treated as running anyway (and any sub-XIDs will also be >=
2258 : * xmax).
2259 : */
2260 724343 : if (!NormalTransactionIdPrecedes(xid, xmax))
2261 203161 : continue;
2262 :
2263 : /*
2264 : * Skip over backends doing logical decoding which manages xmin
2265 : * separately (check below) and ones running LAZY VACUUM.
2266 : */
2267 521182 : statusFlags = allStatusFlags[pgxactoff];
2268 521182 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2269 101 : continue;
2270 :
2271 521081 : if (NormalTransactionIdPrecedes(xid, xmin))
2272 273260 : xmin = xid;
2273 :
2274 : /* Add XID to snapshot. */
2275 521081 : xip[count++] = xid;
2276 :
2277 : /*
2278 : * Save subtransaction XIDs if possible (if we've already
2279 : * overflowed, there's no point). Note that the subxact XIDs must
2280 : * be later than their parent, so no need to check them against
2281 : * xmin. We could filter against xmax, but it seems better not to
2282 : * do that much work while holding the ProcArrayLock.
2283 : *
2284 : * The other backend can add more subxids concurrently, but cannot
2285 : * remove any. Hence it's important to fetch nxids just once.
2286 : * Should be safe to use memcpy, though. (We needn't worry about
2287 : * missing any xids added concurrently, because they must postdate
2288 : * xmax.)
2289 : *
2290 : * Again, our own XIDs are not included in the snapshot.
2291 : */
2292 521081 : if (!suboverflowed)
2293 : {
2294 :
2295 521077 : if (subxidStates[pgxactoff].overflowed)
2296 907 : suboverflowed = true;
2297 : else
2298 : {
2299 520170 : int nsubxids = subxidStates[pgxactoff].count;
2300 :
2301 520170 : if (nsubxids > 0)
2302 : {
2303 5560 : int pgprocno = pgprocnos[pgxactoff];
2304 5560 : PGPROC *proc = &allProcs[pgprocno];
2305 :
2306 5560 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2307 :
2308 5560 : memcpy(snapshot->subxip + subcount,
2309 5560 : proc->subxids.xids,
2310 : nsubxids * sizeof(TransactionId));
2311 5560 : subcount += nsubxids;
2312 : }
2313 : }
2314 : }
2315 : }
2316 : }
2317 : else
2318 : {
2319 : /*
2320 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2321 : *
2322 : * We store all xids directly into subxip[]. Here's why:
2323 : *
2324 : * In recovery we don't know which xids are top-level and which are
2325 : * subxacts, a design choice that greatly simplifies xid processing.
2326 : *
2327 : * It seems like we would want to try to put xids into xip[] only, but
2328 : * that is fairly small. We would either need to make that bigger or
2329 : * to increase the rate at which we WAL-log xid assignment; neither is
2330 : * an appealing choice.
2331 : *
2332 : * We could try to store xids into xip[] first and then into subxip[]
2333 : * if there are too many xids. That only works if the snapshot doesn't
2334 : * overflow because we do not search subxip[] in that case. A simpler
2335 : * way is to just store all xids in the subxip array because this is
2336 : * by far the bigger array. We just leave the xip array empty.
2337 : *
2338 : * Either way we need to change the way XidInMVCCSnapshot() works
2339 : * depending upon when the snapshot was taken, or change normal
2340 : * snapshot processing so it matches.
2341 : *
2342 : * Note: It is possible for recovery to end before we finish taking
2343 : * the snapshot, and for newly assigned transaction ids to be added to
2344 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2345 : * those newly added transaction ids would be filtered away, so we
2346 : * need not be concerned about them.
2347 : */
2348 1280 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2349 : xmax);
2350 :
2351 1280 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2352 6 : suboverflowed = true;
2353 : }
2354 :
2355 :
2356 : /*
2357 : * Fetch into local variable while ProcArrayLock is held - the
2358 : * LWLockRelease below is a barrier, ensuring this happens inside the
2359 : * lock.
2360 : */
2361 401114 : replication_slot_xmin = procArray->replication_slot_xmin;
2362 401114 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2363 :
2364 401114 : if (!TransactionIdIsValid(MyProc->xmin))
2365 221197 : MyProc->xmin = TransactionXmin = xmin;
2366 :
2367 401114 : LWLockRelease(ProcArrayLock);
2368 :
2369 : /* maintain state for GlobalVis* */
2370 : {
2371 : TransactionId def_vis_xid;
2372 : TransactionId def_vis_xid_data;
2373 : FullTransactionId def_vis_fxid;
2374 : FullTransactionId def_vis_fxid_data;
2375 : FullTransactionId oldestfxid;
2376 :
2377 : /*
2378 : * Converting oldestXid is only safe when xid horizon cannot advance,
2379 : * i.e. holding locks. While we don't hold the lock anymore, all the
2380 : * necessary data has been gathered with lock held.
2381 : */
2382 401114 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2383 :
2384 : /* Check whether there's a replication slot requiring an older xmin. */
2385 : def_vis_xid_data =
2386 401114 : TransactionIdOlder(xmin, replication_slot_xmin);
2387 :
2388 : /*
2389 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2390 : * if older than this xid.
2391 : */
2392 401114 : def_vis_xid = def_vis_xid_data;
2393 :
2394 : /*
2395 : * Check whether there's a replication slot requiring an older catalog
2396 : * xmin.
2397 : */
2398 : def_vis_xid =
2399 401114 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2400 :
2401 401114 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2402 401114 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2403 :
2404 : /*
2405 : * Check if we can increase upper bound. As a previous
2406 : * GlobalVisUpdate() might have computed more aggressive values, don't
2407 : * overwrite them if so.
2408 : */
2409 : GlobalVisSharedRels.definitely_needed =
2410 401114 : FullTransactionIdNewer(def_vis_fxid,
2411 : GlobalVisSharedRels.definitely_needed);
2412 : GlobalVisCatalogRels.definitely_needed =
2413 401114 : FullTransactionIdNewer(def_vis_fxid,
2414 : GlobalVisCatalogRels.definitely_needed);
2415 : GlobalVisDataRels.definitely_needed =
2416 401114 : FullTransactionIdNewer(def_vis_fxid_data,
2417 : GlobalVisDataRels.definitely_needed);
2418 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2419 401114 : if (TransactionIdIsNormal(myxid))
2420 : GlobalVisTempRels.definitely_needed =
2421 53109 : FullXidRelativeTo(latest_completed, myxid);
2422 : else
2423 : {
2424 348005 : GlobalVisTempRels.definitely_needed = latest_completed;
2425 348005 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2426 : }
2427 :
2428 : /*
2429 : * Check if we know that we can initialize or increase the lower
2430 : * bound. Currently the only cheap way to do so is to use
2431 : * TransamVariables->oldestXid as input.
2432 : *
2433 : * We should definitely be able to do better. We could e.g. put a
2434 : * global lower bound value into TransamVariables.
2435 : */
2436 : GlobalVisSharedRels.maybe_needed =
2437 401114 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2438 : oldestfxid);
2439 : GlobalVisCatalogRels.maybe_needed =
2440 401114 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2441 : oldestfxid);
2442 : GlobalVisDataRels.maybe_needed =
2443 401114 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2444 : oldestfxid);
2445 : /* accurate value known */
2446 401114 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2447 : }
2448 :
2449 401114 : RecentXmin = xmin;
2450 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2451 :
2452 401114 : snapshot->xmin = xmin;
2453 401114 : snapshot->xmax = xmax;
2454 401114 : snapshot->xcnt = count;
2455 401114 : snapshot->subxcnt = subcount;
2456 401114 : snapshot->suboverflowed = suboverflowed;
2457 401114 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2458 :
2459 401114 : snapshot->curcid = GetCurrentCommandId(false);
2460 :
2461 : /*
2462 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2463 : * not copied in persistent memory.
2464 : */
2465 401114 : snapshot->active_count = 0;
2466 401114 : snapshot->regd_count = 0;
2467 401114 : snapshot->copied = false;
2468 :
2469 401114 : return snapshot;
2470 : }
2471 :
2472 : /*
2473 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2474 : *
2475 : * This is called when installing a snapshot imported from another
2476 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2477 : * check that the source transaction is still running, and we'd better do
2478 : * that atomically with installing the new xmin.
2479 : *
2480 : * Returns true if successful, false if source xact is no longer running.
2481 : */
2482 : bool
2483 16 : ProcArrayInstallImportedXmin(TransactionId xmin,
2484 : VirtualTransactionId *sourcevxid)
2485 : {
2486 16 : bool result = false;
2487 16 : ProcArrayStruct *arrayP = procArray;
2488 : int index;
2489 :
2490 : Assert(TransactionIdIsNormal(xmin));
2491 16 : if (!sourcevxid)
2492 0 : return false;
2493 :
2494 : /* Get lock so source xact can't end while we're doing this */
2495 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2496 :
2497 : /*
2498 : * Find the PGPROC entry of the source transaction. (This could use
2499 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2500 : * performance critical.)
2501 : */
2502 16 : for (index = 0; index < arrayP->numProcs; index++)
2503 : {
2504 16 : int pgprocno = arrayP->pgprocnos[index];
2505 16 : PGPROC *proc = &allProcs[pgprocno];
2506 16 : int statusFlags = ProcGlobal->statusFlags[index];
2507 : TransactionId xid;
2508 :
2509 : /* Ignore procs running LAZY VACUUM */
2510 16 : if (statusFlags & PROC_IN_VACUUM)
2511 0 : continue;
2512 :
2513 : /* We are only interested in the specific virtual transaction. */
2514 16 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2515 0 : continue;
2516 16 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2517 0 : continue;
2518 :
2519 : /*
2520 : * We check the transaction's database ID for paranoia's sake: if it's
2521 : * in another DB then its xmin does not cover us. Caller should have
2522 : * detected this already, so we just treat any funny cases as
2523 : * "transaction not found".
2524 : */
2525 16 : if (proc->databaseId != MyDatabaseId)
2526 0 : continue;
2527 :
2528 : /*
2529 : * Likewise, let's just make real sure its xmin does cover us.
2530 : */
2531 16 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2532 16 : if (!TransactionIdIsNormal(xid) ||
2533 16 : !TransactionIdPrecedesOrEquals(xid, xmin))
2534 0 : continue;
2535 :
2536 : /*
2537 : * We're good. Install the new xmin. As in GetSnapshotData, set
2538 : * TransactionXmin too. (Note that because snapmgr.c called
2539 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2540 : * we don't check that.)
2541 : */
2542 16 : MyProc->xmin = TransactionXmin = xmin;
2543 :
2544 16 : result = true;
2545 16 : break;
2546 : }
2547 :
2548 16 : LWLockRelease(ProcArrayLock);
2549 :
2550 16 : return result;
2551 : }
2552 :
2553 : /*
2554 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2555 : *
2556 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2557 : * PGPROC of the transaction from which we imported the snapshot, rather than
2558 : * an XID.
2559 : *
2560 : * Note that this function also copies statusFlags from the source `proc` in
2561 : * order to avoid the case where MyProc's xmin needs to be skipped for
2562 : * computing xid horizon.
2563 : *
2564 : * Returns true if successful, false if source xact is no longer running.
2565 : */
2566 : bool
2567 1683 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2568 : {
2569 1683 : bool result = false;
2570 : TransactionId xid;
2571 :
2572 : Assert(TransactionIdIsNormal(xmin));
2573 : Assert(proc != NULL);
2574 :
2575 : /*
2576 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2577 : */
2578 1683 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2579 :
2580 : /*
2581 : * Be certain that the referenced PGPROC has an advertised xmin which is
2582 : * no later than the one we're installing, so that the system-wide xmin
2583 : * can't go backwards. Also, make sure it's running in the same database,
2584 : * so that the per-database xmin cannot go backwards.
2585 : */
2586 1683 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2587 1683 : if (proc->databaseId == MyDatabaseId &&
2588 1683 : TransactionIdIsNormal(xid) &&
2589 1683 : TransactionIdPrecedesOrEquals(xid, xmin))
2590 : {
2591 : /*
2592 : * Install xmin and propagate the statusFlags that affect how the
2593 : * value is interpreted by vacuum.
2594 : */
2595 1683 : MyProc->xmin = TransactionXmin = xmin;
2596 1683 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2597 1683 : (proc->statusFlags & PROC_XMIN_FLAGS);
2598 1683 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2599 :
2600 1683 : result = true;
2601 : }
2602 :
2603 1683 : LWLockRelease(ProcArrayLock);
2604 :
2605 1683 : return result;
2606 : }
2607 :
2608 : /*
2609 : * GetRunningTransactionData -- returns information about running transactions.
2610 : *
2611 : * Similar to GetSnapshotData but returns more information. We include
2612 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2613 : * prepared transactions.
2614 : *
2615 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2616 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2617 : * array until the caller has WAL-logged this snapshot, and releases the
2618 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2619 : * lock is released.
2620 : *
2621 : * The returned data structure is statically allocated; caller should not
2622 : * modify it, and must not assume it is valid past the next call.
2623 : *
2624 : * This is never executed during recovery so there is no need to look at
2625 : * KnownAssignedXids.
2626 : *
2627 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2628 : * may return entries with duplicated TransactionId values coming from
2629 : * transaction finishing to prepare. Nothing is done about duplicated
2630 : * entries here to not hold on ProcArrayLock more than necessary.
2631 : *
2632 : * We don't worry about updating other counters, we want to keep this as
2633 : * simple as possible and leave GetSnapshotData() as the primary code for
2634 : * that bookkeeping.
2635 : *
2636 : * Note that if any transaction has overflowed its cached subtransactions
2637 : * then there is no real need include any subtransactions.
2638 : */
2639 : RunningTransactions
2640 1448 : GetRunningTransactionData(void)
2641 : {
2642 : /* result workspace */
2643 : static RunningTransactionsData CurrentRunningXactsData;
2644 :
2645 1448 : ProcArrayStruct *arrayP = procArray;
2646 1448 : TransactionId *other_xids = ProcGlobal->xids;
2647 1448 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2648 : TransactionId latestCompletedXid;
2649 : TransactionId oldestRunningXid;
2650 : TransactionId oldestDatabaseRunningXid;
2651 : TransactionId *xids;
2652 : int index;
2653 : int count;
2654 : int subcount;
2655 : bool suboverflowed;
2656 :
2657 : Assert(!RecoveryInProgress());
2658 :
2659 : /*
2660 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2661 : * be sufficient. But it seems better to do the malloc while not holding
2662 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2663 : * more subxip storage than is probably needed.
2664 : *
2665 : * Should only be allocated in bgwriter, since only ever executed during
2666 : * checkpoints.
2667 : */
2668 1448 : if (CurrentRunningXacts->xids == NULL)
2669 : {
2670 : /*
2671 : * First call
2672 : */
2673 562 : CurrentRunningXacts->xids = (TransactionId *)
2674 562 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2675 562 : if (CurrentRunningXacts->xids == NULL)
2676 0 : ereport(ERROR,
2677 : (errcode(ERRCODE_OUT_OF_MEMORY),
2678 : errmsg("out of memory")));
2679 : }
2680 :
2681 1448 : xids = CurrentRunningXacts->xids;
2682 :
2683 1448 : count = subcount = 0;
2684 1448 : suboverflowed = false;
2685 :
2686 : /*
2687 : * Ensure that no xids enter or leave the procarray while we obtain
2688 : * snapshot.
2689 : */
2690 1448 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2691 1448 : LWLockAcquire(XidGenLock, LW_SHARED);
2692 :
2693 1448 : latestCompletedXid =
2694 1448 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2695 1448 : oldestDatabaseRunningXid = oldestRunningXid =
2696 1448 : XidFromFullTransactionId(TransamVariables->nextXid);
2697 :
2698 : /*
2699 : * Spin over procArray collecting all xids
2700 : */
2701 7101 : for (index = 0; index < arrayP->numProcs; index++)
2702 : {
2703 : TransactionId xid;
2704 :
2705 : /* Fetch xid just once - see GetNewTransactionId */
2706 5653 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2707 :
2708 : /*
2709 : * We don't need to store transactions that don't have a TransactionId
2710 : * yet because they will not show as running on a standby server.
2711 : */
2712 5653 : if (!TransactionIdIsValid(xid))
2713 4718 : continue;
2714 :
2715 : /*
2716 : * Be careful not to exclude any xids before calculating the values of
2717 : * oldestRunningXid and suboverflowed, since these are used to clean
2718 : * up transaction information held on standbys.
2719 : */
2720 935 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2721 612 : oldestRunningXid = xid;
2722 :
2723 : /*
2724 : * Also, update the oldest running xid within the current database. As
2725 : * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2726 : * TransactionId comparison first.
2727 : */
2728 935 : if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2729 : {
2730 935 : int pgprocno = arrayP->pgprocnos[index];
2731 935 : PGPROC *proc = &allProcs[pgprocno];
2732 :
2733 935 : if (proc->databaseId == MyDatabaseId)
2734 214 : oldestDatabaseRunningXid = xid;
2735 : }
2736 :
2737 935 : if (ProcGlobal->subxidStates[index].overflowed)
2738 2 : suboverflowed = true;
2739 :
2740 : /*
2741 : * If we wished to exclude xids this would be the right place for it.
2742 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2743 : * but they do during truncation at the end when they get the lock and
2744 : * truncate, so it is not much of a problem to include them if they
2745 : * are seen and it is cleaner to include them.
2746 : */
2747 :
2748 935 : xids[count++] = xid;
2749 : }
2750 :
2751 : /*
2752 : * Spin over procArray collecting all subxids, but only if there hasn't
2753 : * been a suboverflow.
2754 : */
2755 1448 : if (!suboverflowed)
2756 : {
2757 1446 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2758 :
2759 7093 : for (index = 0; index < arrayP->numProcs; index++)
2760 : {
2761 5647 : int pgprocno = arrayP->pgprocnos[index];
2762 5647 : PGPROC *proc = &allProcs[pgprocno];
2763 : int nsubxids;
2764 :
2765 : /*
2766 : * Save subtransaction XIDs. Other backends can't add or remove
2767 : * entries while we're holding XidGenLock.
2768 : */
2769 5647 : nsubxids = other_subxidstates[index].count;
2770 5647 : if (nsubxids > 0)
2771 : {
2772 : /* barrier not really required, as XidGenLock is held, but ... */
2773 19 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2774 :
2775 19 : memcpy(&xids[count], proc->subxids.xids,
2776 : nsubxids * sizeof(TransactionId));
2777 19 : count += nsubxids;
2778 19 : subcount += nsubxids;
2779 :
2780 : /*
2781 : * Top-level XID of a transaction is always less than any of
2782 : * its subxids, so we don't need to check if any of the
2783 : * subxids are smaller than oldestRunningXid
2784 : */
2785 : }
2786 : }
2787 : }
2788 :
2789 : /*
2790 : * It's important *not* to include the limits set by slots here because
2791 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2792 : * were to be included here the initial value could never increase because
2793 : * of a circular dependency where slots only increase their limits when
2794 : * running xacts increases oldestRunningXid and running xacts only
2795 : * increases if slots do.
2796 : */
2797 :
2798 1448 : CurrentRunningXacts->xcnt = count - subcount;
2799 1448 : CurrentRunningXacts->subxcnt = subcount;
2800 1448 : CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2801 1448 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2802 1448 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2803 1448 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2804 1448 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2805 :
2806 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2807 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2808 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2809 :
2810 : /* We don't release the locks here, the caller is responsible for that */
2811 :
2812 1448 : return CurrentRunningXacts;
2813 : }
2814 :
2815 : /*
2816 : * GetOldestActiveTransactionId()
2817 : *
2818 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2819 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2820 : *
2821 : * If allDbs is true, we look at all databases, though there is no need to
2822 : * include WALSender since this has no effect on hot standby conflicts. If
2823 : * allDbs is false, skip processes attached to other databases.
2824 : *
2825 : * This is never executed during recovery so there is no need to look at
2826 : * KnownAssignedXids.
2827 : *
2828 : * We don't worry about updating other counters, we want to keep this as
2829 : * simple as possible and leave GetSnapshotData() as the primary code for
2830 : * that bookkeeping.
2831 : *
2832 : * inCommitOnly indicates getting the oldestActiveXid among the transactions
2833 : * in the commit critical section.
2834 : */
2835 : TransactionId
2836 3267 : GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
2837 : {
2838 3267 : ProcArrayStruct *arrayP = procArray;
2839 3267 : TransactionId *other_xids = ProcGlobal->xids;
2840 : TransactionId oldestRunningXid;
2841 : int index;
2842 :
2843 : Assert(!RecoveryInProgress());
2844 :
2845 : /*
2846 : * Read nextXid, as the upper bound of what's still active.
2847 : *
2848 : * Reading a TransactionId is atomic, but we must grab the lock to make
2849 : * sure that all XIDs < nextXid are already present in the proc array (or
2850 : * have already completed), when we spin over it.
2851 : */
2852 3267 : LWLockAcquire(XidGenLock, LW_SHARED);
2853 3267 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2854 3267 : LWLockRelease(XidGenLock);
2855 :
2856 : /*
2857 : * Spin over procArray collecting all xids and subxids.
2858 : */
2859 3267 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2860 18478 : for (index = 0; index < arrayP->numProcs; index++)
2861 : {
2862 : TransactionId xid;
2863 15211 : int pgprocno = arrayP->pgprocnos[index];
2864 15211 : PGPROC *proc = &allProcs[pgprocno];
2865 :
2866 : /* Fetch xid just once - see GetNewTransactionId */
2867 15211 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2868 :
2869 15211 : if (!TransactionIdIsNormal(xid))
2870 12181 : continue;
2871 :
2872 3030 : if (inCommitOnly &&
2873 2130 : (proc->delayChkptFlags & DELAY_CHKPT_IN_COMMIT) == 0)
2874 1999 : continue;
2875 :
2876 1031 : if (!allDbs && proc->databaseId != MyDatabaseId)
2877 0 : continue;
2878 :
2879 1031 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2880 718 : oldestRunningXid = xid;
2881 :
2882 : /*
2883 : * Top-level XID of a transaction is always less than any of its
2884 : * subxids, so we don't need to check if any of the subxids are
2885 : * smaller than oldestRunningXid
2886 : */
2887 : }
2888 3267 : LWLockRelease(ProcArrayLock);
2889 :
2890 3267 : return oldestRunningXid;
2891 : }
2892 :
2893 : /*
2894 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2895 : *
2896 : * Returns the oldest xid that we can guarantee not to have been affected by
2897 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2898 : * transaction aborted. Note that the value can (and most of the time will) be
2899 : * much more conservative than what really has been affected by vacuum, but we
2900 : * currently don't have better data available.
2901 : *
2902 : * This is useful to initialize the cutoff xid after which a new changeset
2903 : * extraction replication slot can start decoding changes.
2904 : *
2905 : * Must be called with ProcArrayLock held either shared or exclusively,
2906 : * although most callers will want to use exclusive mode since it is expected
2907 : * that the caller will immediately use the xid to peg the xmin horizon.
2908 : */
2909 : TransactionId
2910 687 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2911 : {
2912 687 : ProcArrayStruct *arrayP = procArray;
2913 : TransactionId oldestSafeXid;
2914 : int index;
2915 687 : bool recovery_in_progress = RecoveryInProgress();
2916 :
2917 : Assert(LWLockHeldByMe(ProcArrayLock));
2918 :
2919 : /*
2920 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2921 : * running. If no transaction with xid were running concurrently a new xid
2922 : * could influence the RecentXmin et al.
2923 : *
2924 : * We initialize the computation to nextXid since that's guaranteed to be
2925 : * a safe, albeit pessimal, value.
2926 : */
2927 687 : LWLockAcquire(XidGenLock, LW_SHARED);
2928 687 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2929 :
2930 : /*
2931 : * If there's already a slot pegging the xmin horizon, we can start with
2932 : * that value, it's guaranteed to be safe since it's computed by this
2933 : * routine initially and has been enforced since. We can always use the
2934 : * slot's general xmin horizon, but the catalog horizon is only usable
2935 : * when only catalog data is going to be looked at.
2936 : */
2937 910 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2938 223 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2939 : oldestSafeXid))
2940 14 : oldestSafeXid = procArray->replication_slot_xmin;
2941 :
2942 687 : if (catalogOnly &&
2943 357 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2944 74 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2945 : oldestSafeXid))
2946 29 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2947 :
2948 : /*
2949 : * If we're not in recovery, we walk over the procarray and collect the
2950 : * lowest xid. Since we're called with ProcArrayLock held and have
2951 : * acquired XidGenLock, no entries can vanish concurrently, since
2952 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2953 : * with ProcArrayLock held.
2954 : *
2955 : * In recovery we can't lower the safe value besides what we've computed
2956 : * above, so we'll have to wait a bit longer there. We unfortunately can
2957 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2958 : * machinery can miss values and return an older value than is safe.
2959 : */
2960 687 : if (!recovery_in_progress)
2961 : {
2962 654 : TransactionId *other_xids = ProcGlobal->xids;
2963 :
2964 : /*
2965 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2966 : */
2967 3367 : for (index = 0; index < arrayP->numProcs; index++)
2968 : {
2969 : TransactionId xid;
2970 :
2971 : /* Fetch xid just once - see GetNewTransactionId */
2972 2713 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2973 :
2974 2713 : if (!TransactionIdIsNormal(xid))
2975 2703 : continue;
2976 :
2977 10 : if (TransactionIdPrecedes(xid, oldestSafeXid))
2978 9 : oldestSafeXid = xid;
2979 : }
2980 : }
2981 :
2982 687 : LWLockRelease(XidGenLock);
2983 :
2984 687 : return oldestSafeXid;
2985 : }
2986 :
2987 : /*
2988 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2989 : * delaying checkpoint because they have critical actions in progress.
2990 : *
2991 : * Constructs an array of VXIDs of transactions that are currently in commit
2992 : * critical sections, as shown by having specified delayChkptFlags bits set
2993 : * in their PGPROC.
2994 : *
2995 : * Returns a palloc'd array that should be freed by the caller.
2996 : * *nvxids is the number of valid entries.
2997 : *
2998 : * Note that because backends set or clear delayChkptFlags without holding any
2999 : * lock, the result is somewhat indeterminate, but we don't really care. Even
3000 : * in a multiprocessor with delayed writes to shared memory, it should be
3001 : * certain that setting of delayChkptFlags will propagate to shared memory
3002 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
3003 : * delayChkptFlags if it's already inserted its commit record. Whether it
3004 : * takes a little while for clearing of delayChkptFlags to propagate is
3005 : * unimportant for correctness.
3006 : */
3007 : VirtualTransactionId *
3008 3172 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3009 : {
3010 : VirtualTransactionId *vxids;
3011 3172 : ProcArrayStruct *arrayP = procArray;
3012 3172 : int count = 0;
3013 : int index;
3014 :
3015 : Assert(type != 0);
3016 :
3017 : /* allocate what's certainly enough result space */
3018 3172 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3019 :
3020 3172 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3021 :
3022 10426 : for (index = 0; index < arrayP->numProcs; index++)
3023 : {
3024 7254 : int pgprocno = arrayP->pgprocnos[index];
3025 7254 : PGPROC *proc = &allProcs[pgprocno];
3026 :
3027 7254 : if ((proc->delayChkptFlags & type) != 0)
3028 : {
3029 : VirtualTransactionId vxid;
3030 :
3031 40 : GET_VXID_FROM_PGPROC(vxid, *proc);
3032 40 : if (VirtualTransactionIdIsValid(vxid))
3033 40 : vxids[count++] = vxid;
3034 : }
3035 : }
3036 :
3037 3172 : LWLockRelease(ProcArrayLock);
3038 :
3039 3172 : *nvxids = count;
3040 3172 : return vxids;
3041 : }
3042 :
3043 : /*
3044 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3045 : *
3046 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3047 : * of the specified VXIDs are still in critical sections of code.
3048 : *
3049 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3050 : * those numbers should be small enough for it not to be a problem.
3051 : */
3052 : bool
3053 42 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3054 : {
3055 42 : bool result = false;
3056 42 : ProcArrayStruct *arrayP = procArray;
3057 : int index;
3058 :
3059 : Assert(type != 0);
3060 :
3061 42 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3062 :
3063 431 : for (index = 0; index < arrayP->numProcs; index++)
3064 : {
3065 395 : int pgprocno = arrayP->pgprocnos[index];
3066 395 : PGPROC *proc = &allProcs[pgprocno];
3067 : VirtualTransactionId vxid;
3068 :
3069 395 : GET_VXID_FROM_PGPROC(vxid, *proc);
3070 :
3071 395 : if ((proc->delayChkptFlags & type) != 0 &&
3072 23 : VirtualTransactionIdIsValid(vxid))
3073 : {
3074 : int i;
3075 :
3076 43 : for (i = 0; i < nvxids; i++)
3077 : {
3078 26 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3079 : {
3080 6 : result = true;
3081 6 : break;
3082 : }
3083 : }
3084 23 : if (result)
3085 6 : break;
3086 : }
3087 : }
3088 :
3089 42 : LWLockRelease(ProcArrayLock);
3090 :
3091 42 : return result;
3092 : }
3093 :
3094 : /*
3095 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3096 : *
3097 : * The result may be out of date arbitrarily quickly, so the caller
3098 : * must be careful about how this information is used. NULL is
3099 : * returned if the backend is not active.
3100 : */
3101 : PGPROC *
3102 588 : ProcNumberGetProc(ProcNumber procNumber)
3103 : {
3104 : PGPROC *result;
3105 :
3106 588 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3107 1 : return NULL;
3108 587 : result = GetPGProcByNumber(procNumber);
3109 :
3110 587 : if (result->pid == 0)
3111 3 : return NULL;
3112 :
3113 584 : return result;
3114 : }
3115 :
3116 : /*
3117 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3118 : *
3119 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3120 : * result may be out of date arbitrarily quickly, so the caller must be
3121 : * careful about how this information is used.
3122 : */
3123 : void
3124 11043 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3125 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3126 : {
3127 : PGPROC *proc;
3128 :
3129 11043 : *xid = InvalidTransactionId;
3130 11043 : *xmin = InvalidTransactionId;
3131 11043 : *nsubxid = 0;
3132 11043 : *overflowed = false;
3133 :
3134 11043 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3135 0 : return;
3136 11043 : proc = GetPGProcByNumber(procNumber);
3137 :
3138 : /* Need to lock out additions/removals of backends */
3139 11043 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3140 :
3141 11043 : if (proc->pid != 0)
3142 : {
3143 11043 : *xid = proc->xid;
3144 11043 : *xmin = proc->xmin;
3145 11043 : *nsubxid = proc->subxidStatus.count;
3146 11043 : *overflowed = proc->subxidStatus.overflowed;
3147 : }
3148 :
3149 11043 : LWLockRelease(ProcArrayLock);
3150 : }
3151 :
3152 : /*
3153 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3154 : *
3155 : * Returns NULL if not found. Note that it is up to the caller to be
3156 : * sure that the question remains meaningful for long enough for the
3157 : * answer to be used ...
3158 : */
3159 : PGPROC *
3160 12510 : BackendPidGetProc(int pid)
3161 : {
3162 : PGPROC *result;
3163 :
3164 12510 : if (pid == 0) /* never match dummy PGPROCs */
3165 3 : return NULL;
3166 :
3167 12507 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3168 :
3169 12507 : result = BackendPidGetProcWithLock(pid);
3170 :
3171 12507 : LWLockRelease(ProcArrayLock);
3172 :
3173 12507 : return result;
3174 : }
3175 :
3176 : /*
3177 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3178 : *
3179 : * Same as above, except caller must be holding ProcArrayLock. The found
3180 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3181 : */
3182 : PGPROC *
3183 14732 : BackendPidGetProcWithLock(int pid)
3184 : {
3185 14732 : PGPROC *result = NULL;
3186 14732 : ProcArrayStruct *arrayP = procArray;
3187 : int index;
3188 :
3189 14732 : if (pid == 0) /* never match dummy PGPROCs */
3190 0 : return NULL;
3191 :
3192 56643 : for (index = 0; index < arrayP->numProcs; index++)
3193 : {
3194 50589 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3195 :
3196 50589 : if (proc->pid == pid)
3197 : {
3198 8678 : result = proc;
3199 8678 : break;
3200 : }
3201 : }
3202 :
3203 14732 : return result;
3204 : }
3205 :
3206 : /*
3207 : * BackendXidGetPid -- get a backend's pid given its XID
3208 : *
3209 : * Returns 0 if not found or it's a prepared transaction. Note that
3210 : * it is up to the caller to be sure that the question remains
3211 : * meaningful for long enough for the answer to be used ...
3212 : *
3213 : * Only main transaction Ids are considered. This function is mainly
3214 : * useful for determining what backend owns a lock.
3215 : *
3216 : * Beware that not every xact has an XID assigned. However, as long as you
3217 : * only call this using an XID found on disk, you're safe.
3218 : */
3219 : int
3220 30 : BackendXidGetPid(TransactionId xid)
3221 : {
3222 30 : int result = 0;
3223 30 : ProcArrayStruct *arrayP = procArray;
3224 30 : TransactionId *other_xids = ProcGlobal->xids;
3225 : int index;
3226 :
3227 30 : if (xid == InvalidTransactionId) /* never match invalid xid */
3228 0 : return 0;
3229 :
3230 30 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3231 :
3232 92 : for (index = 0; index < arrayP->numProcs; index++)
3233 : {
3234 84 : if (other_xids[index] == xid)
3235 : {
3236 22 : int pgprocno = arrayP->pgprocnos[index];
3237 22 : PGPROC *proc = &allProcs[pgprocno];
3238 :
3239 22 : result = proc->pid;
3240 22 : break;
3241 : }
3242 : }
3243 :
3244 30 : LWLockRelease(ProcArrayLock);
3245 :
3246 30 : return result;
3247 : }
3248 :
3249 : /*
3250 : * IsBackendPid -- is a given pid a running backend
3251 : *
3252 : * This is not called by the backend, but is called by external modules.
3253 : */
3254 : bool
3255 2 : IsBackendPid(int pid)
3256 : {
3257 2 : return (BackendPidGetProc(pid) != NULL);
3258 : }
3259 :
3260 :
3261 : /*
3262 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3263 : *
3264 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3265 : *
3266 : * The arguments allow filtering the set of VXIDs returned. Our own process
3267 : * is always skipped. In addition:
3268 : * If limitXmin is not InvalidTransactionId, skip processes with
3269 : * xmin > limitXmin.
3270 : * If excludeXmin0 is true, skip processes with xmin = 0.
3271 : * If allDbs is false, skip processes attached to other databases.
3272 : * If excludeVacuum isn't zero, skip processes for which
3273 : * (statusFlags & excludeVacuum) is not zero.
3274 : *
3275 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3276 : * allow skipping backends whose oldest live snapshot is no older than
3277 : * some snapshot we have. Since we examine the procarray with only shared
3278 : * lock, there are race conditions: a backend could set its xmin just after
3279 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3280 : * other backend could have set its xmin *before* we look. We know however
3281 : * that such a backend must have held shared ProcArrayLock overlapping our
3282 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3283 : * any snapshot the other backend is taking concurrently with our scan cannot
3284 : * consider any transactions as still running that we think are committed
3285 : * (since backends must hold ProcArrayLock exclusive to commit).
3286 : */
3287 : VirtualTransactionId *
3288 404 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3289 : bool allDbs, int excludeVacuum,
3290 : int *nvxids)
3291 : {
3292 : VirtualTransactionId *vxids;
3293 404 : ProcArrayStruct *arrayP = procArray;
3294 404 : int count = 0;
3295 : int index;
3296 :
3297 : /* allocate what's certainly enough result space */
3298 404 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3299 :
3300 404 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3301 :
3302 2439 : for (index = 0; index < arrayP->numProcs; index++)
3303 : {
3304 2035 : int pgprocno = arrayP->pgprocnos[index];
3305 2035 : PGPROC *proc = &allProcs[pgprocno];
3306 2035 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3307 :
3308 2035 : if (proc == MyProc)
3309 404 : continue;
3310 :
3311 1631 : if (excludeVacuum & statusFlags)
3312 6 : continue;
3313 :
3314 1625 : if (allDbs || proc->databaseId == MyDatabaseId)
3315 : {
3316 : /* Fetch xmin just once - might change on us */
3317 744 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3318 :
3319 744 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3320 433 : continue;
3321 :
3322 : /*
3323 : * InvalidTransactionId precedes all other XIDs, so a proc that
3324 : * hasn't set xmin yet will not be rejected by this test.
3325 : */
3326 622 : if (!TransactionIdIsValid(limitXmin) ||
3327 311 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3328 : {
3329 : VirtualTransactionId vxid;
3330 :
3331 290 : GET_VXID_FROM_PGPROC(vxid, *proc);
3332 290 : if (VirtualTransactionIdIsValid(vxid))
3333 290 : vxids[count++] = vxid;
3334 : }
3335 : }
3336 : }
3337 :
3338 404 : LWLockRelease(ProcArrayLock);
3339 :
3340 404 : *nvxids = count;
3341 404 : return vxids;
3342 : }
3343 :
3344 : /*
3345 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3346 : *
3347 : * Usage is limited to conflict resolution during recovery on standby servers.
3348 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3349 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3350 : * determine a safe snapshotConflictHorizon value.
3351 : *
3352 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3353 : * so we're not worried if they have a snapshot or not, nor does it really
3354 : * matter what type of lock we hold. Caller must avoid calling here with
3355 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3356 : * during original execution, since that actually indicates that there is
3357 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3358 : * convention for InvalidTransactionId values is the opposite of our own!).
3359 : *
3360 : * All callers that are checking xmins always now supply a valid and useful
3361 : * value for limitXmin. The limitXmin is always lower than the lowest
3362 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3363 : * because we only care about cleanup records that are cleaning up tuple
3364 : * versions from committed transactions. In that case they will only occur
3365 : * at the point where the record is less than the lowest running xid. That
3366 : * allows us to say that if any backend takes a snapshot concurrently with
3367 : * us then the conflict assessment made here would never include the snapshot
3368 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3369 : * concurrent snapshots when limitXmin is valid. We might think about adding
3370 : * Assert(limitXmin < lowest(KnownAssignedXids))
3371 : * but that would not be true in the case of FATAL errors lagging in array,
3372 : * but we already know those are bogus anyway, so we skip that test.
3373 : *
3374 : * If dbOid is valid we skip backends attached to other databases.
3375 : *
3376 : * Be careful to *not* pfree the result from this function. We reuse
3377 : * this array sufficiently often that we use malloc for the result.
3378 : */
3379 : VirtualTransactionId *
3380 14043 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3381 : {
3382 : static VirtualTransactionId *vxids;
3383 14043 : ProcArrayStruct *arrayP = procArray;
3384 14043 : int count = 0;
3385 : int index;
3386 :
3387 : /*
3388 : * If first time through, get workspace to remember main XIDs in. We
3389 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3390 : * result space, remembering room for a terminator.
3391 : */
3392 14043 : if (vxids == NULL)
3393 : {
3394 22 : vxids = (VirtualTransactionId *)
3395 22 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3396 22 : if (vxids == NULL)
3397 0 : ereport(ERROR,
3398 : (errcode(ERRCODE_OUT_OF_MEMORY),
3399 : errmsg("out of memory")));
3400 : }
3401 :
3402 14043 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3403 :
3404 14191 : for (index = 0; index < arrayP->numProcs; index++)
3405 : {
3406 148 : int pgprocno = arrayP->pgprocnos[index];
3407 148 : PGPROC *proc = &allProcs[pgprocno];
3408 :
3409 : /* Exclude prepared transactions */
3410 148 : if (proc->pid == 0)
3411 0 : continue;
3412 :
3413 148 : if (!OidIsValid(dbOid) ||
3414 142 : proc->databaseId == dbOid)
3415 : {
3416 : /* Fetch xmin just once - can't change on us, but good coding */
3417 19 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3418 :
3419 : /*
3420 : * We ignore an invalid pxmin because this means that backend has
3421 : * no snapshot currently. We hold a Share lock to avoid contention
3422 : * with users taking snapshots. That is not a problem because the
3423 : * current xmin is always at least one higher than the latest
3424 : * removed xid, so any new snapshot would never conflict with the
3425 : * test here.
3426 : */
3427 19 : if (!TransactionIdIsValid(limitXmin) ||
3428 3 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3429 : {
3430 : VirtualTransactionId vxid;
3431 :
3432 2 : GET_VXID_FROM_PGPROC(vxid, *proc);
3433 2 : if (VirtualTransactionIdIsValid(vxid))
3434 2 : vxids[count++] = vxid;
3435 : }
3436 : }
3437 : }
3438 :
3439 14043 : LWLockRelease(ProcArrayLock);
3440 :
3441 : /* add the terminator */
3442 14043 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3443 14043 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3444 :
3445 14043 : return vxids;
3446 : }
3447 :
3448 : /*
3449 : * SignalRecoveryConflict -- signal that a process is blocking recovery
3450 : *
3451 : * The 'pid' is redundant with 'proc', but it acts as a cross-check to
3452 : * detect process had exited and the PGPROC entry was reused for a different
3453 : * process.
3454 : *
3455 : * Returns true if the process was signaled, or false if not found.
3456 : */
3457 : bool
3458 5 : SignalRecoveryConflict(PGPROC *proc, pid_t pid, RecoveryConflictReason reason)
3459 : {
3460 5 : bool found = false;
3461 :
3462 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3463 :
3464 : /*
3465 : * Kill the pid if it's still here. If not, that's what we wanted so
3466 : * ignore any errors.
3467 : */
3468 5 : if (proc->pid == pid)
3469 : {
3470 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3471 :
3472 : /* wake up the process */
3473 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, GetNumberFromPGProc(proc));
3474 5 : found = true;
3475 : }
3476 :
3477 5 : LWLockRelease(ProcArrayLock);
3478 :
3479 5 : return found;
3480 : }
3481 :
3482 : /*
3483 : * SignalRecoveryConflictWithVirtualXID -- signal that a VXID is blocking recovery
3484 : *
3485 : * Like SignalRecoveryConflict, but the target is identified by VXID
3486 : */
3487 : bool
3488 5 : SignalRecoveryConflictWithVirtualXID(VirtualTransactionId vxid, RecoveryConflictReason reason)
3489 : {
3490 5 : ProcArrayStruct *arrayP = procArray;
3491 : int index;
3492 5 : pid_t pid = 0;
3493 :
3494 5 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3495 :
3496 5 : for (index = 0; index < arrayP->numProcs; index++)
3497 : {
3498 5 : int pgprocno = arrayP->pgprocnos[index];
3499 5 : PGPROC *proc = &allProcs[pgprocno];
3500 : VirtualTransactionId procvxid;
3501 :
3502 5 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3503 :
3504 5 : if (procvxid.procNumber == vxid.procNumber &&
3505 5 : procvxid.localTransactionId == vxid.localTransactionId)
3506 : {
3507 5 : pid = proc->pid;
3508 5 : if (pid != 0)
3509 : {
3510 5 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3511 :
3512 : /*
3513 : * Kill the pid if it's still here. If not, that's what we
3514 : * wanted so ignore any errors.
3515 : */
3516 5 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, vxid.procNumber);
3517 : }
3518 5 : break;
3519 : }
3520 : }
3521 :
3522 5 : LWLockRelease(ProcArrayLock);
3523 :
3524 5 : return pid != 0;
3525 : }
3526 :
3527 : /*
3528 : * SignalRecoveryConflictWithDatabase --- signal all backends specified database
3529 : *
3530 : * Like SignalRecoveryConflict, but signals all backends using the database.
3531 : */
3532 : void
3533 9 : SignalRecoveryConflictWithDatabase(Oid databaseid, RecoveryConflictReason reason)
3534 : {
3535 9 : ProcArrayStruct *arrayP = procArray;
3536 : int index;
3537 :
3538 : /* tell all backends to die */
3539 9 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3540 :
3541 18 : for (index = 0; index < arrayP->numProcs; index++)
3542 : {
3543 9 : int pgprocno = arrayP->pgprocnos[index];
3544 9 : PGPROC *proc = &allProcs[pgprocno];
3545 :
3546 9 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3547 : {
3548 : VirtualTransactionId procvxid;
3549 : pid_t pid;
3550 :
3551 9 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3552 :
3553 9 : pid = proc->pid;
3554 9 : if (pid != 0)
3555 : {
3556 9 : (void) pg_atomic_fetch_or_u32(&proc->pendingRecoveryConflicts, (1 << reason));
3557 :
3558 : /*
3559 : * Kill the pid if it's still here. If not, that's what we
3560 : * wanted so ignore any errors.
3561 : */
3562 9 : (void) SendProcSignal(pid, PROCSIG_RECOVERY_CONFLICT, procvxid.procNumber);
3563 : }
3564 : }
3565 : }
3566 :
3567 9 : LWLockRelease(ProcArrayLock);
3568 9 : }
3569 :
3570 : /*
3571 : * MinimumActiveBackends --- count backends (other than myself) that are
3572 : * in active transactions. Return true if the count exceeds the
3573 : * minimum threshold passed. This is used as a heuristic to decide if
3574 : * a pre-XLOG-flush delay is worthwhile during commit.
3575 : *
3576 : * Do not count backends that are blocked waiting for locks, since they are
3577 : * not going to get to run until someone else commits.
3578 : */
3579 : bool
3580 0 : MinimumActiveBackends(int min)
3581 : {
3582 0 : ProcArrayStruct *arrayP = procArray;
3583 0 : int count = 0;
3584 : int index;
3585 :
3586 : /* Quick short-circuit if no minimum is specified */
3587 0 : if (min == 0)
3588 0 : return true;
3589 :
3590 : /*
3591 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3592 : * bogus, but since we are only testing fields for zero or nonzero, it
3593 : * should be OK. The result is only used for heuristic purposes anyway...
3594 : */
3595 0 : for (index = 0; index < arrayP->numProcs; index++)
3596 : {
3597 0 : int pgprocno = arrayP->pgprocnos[index];
3598 0 : PGPROC *proc = &allProcs[pgprocno];
3599 :
3600 : /*
3601 : * Since we're not holding a lock, need to be prepared to deal with
3602 : * garbage, as someone could have incremented numProcs but not yet
3603 : * filled the structure.
3604 : *
3605 : * If someone just decremented numProcs, 'proc' could also point to a
3606 : * PGPROC entry that's no longer in the array. It still points to a
3607 : * PGPROC struct, though, because freed PGPROC entries just go to the
3608 : * free list and are recycled. Its contents are nonsense in that case,
3609 : * but that's acceptable for this function.
3610 : */
3611 0 : if (pgprocno == -1)
3612 0 : continue; /* do not count deleted entries */
3613 0 : if (proc == MyProc)
3614 0 : continue; /* do not count myself */
3615 0 : if (proc->xid == InvalidTransactionId)
3616 0 : continue; /* do not count if no XID assigned */
3617 0 : if (proc->pid == 0)
3618 0 : continue; /* do not count prepared xacts */
3619 0 : if (proc->waitLock != NULL)
3620 0 : continue; /* do not count if blocked on a lock */
3621 0 : count++;
3622 0 : if (count >= min)
3623 0 : break;
3624 : }
3625 :
3626 0 : return count >= min;
3627 : }
3628 :
3629 : /*
3630 : * CountDBBackends --- count backends that are using specified database
3631 : */
3632 : int
3633 16 : CountDBBackends(Oid databaseid)
3634 : {
3635 16 : ProcArrayStruct *arrayP = procArray;
3636 16 : int count = 0;
3637 : int index;
3638 :
3639 16 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3640 :
3641 23 : for (index = 0; index < arrayP->numProcs; index++)
3642 : {
3643 7 : int pgprocno = arrayP->pgprocnos[index];
3644 7 : PGPROC *proc = &allProcs[pgprocno];
3645 :
3646 7 : if (proc->pid == 0)
3647 0 : continue; /* do not count prepared xacts */
3648 7 : if (!OidIsValid(databaseid) ||
3649 7 : proc->databaseId == databaseid)
3650 2 : count++;
3651 : }
3652 :
3653 16 : LWLockRelease(ProcArrayLock);
3654 :
3655 16 : return count;
3656 : }
3657 :
3658 : /*
3659 : * CountDBConnections --- counts database backends (only regular backends)
3660 : */
3661 : int
3662 0 : CountDBConnections(Oid databaseid)
3663 : {
3664 0 : ProcArrayStruct *arrayP = procArray;
3665 0 : int count = 0;
3666 : int index;
3667 :
3668 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3669 :
3670 0 : for (index = 0; index < arrayP->numProcs; index++)
3671 : {
3672 0 : int pgprocno = arrayP->pgprocnos[index];
3673 0 : PGPROC *proc = &allProcs[pgprocno];
3674 :
3675 0 : if (proc->pid == 0)
3676 0 : continue; /* do not count prepared xacts */
3677 0 : if (proc->backendType != B_BACKEND)
3678 0 : continue; /* count only regular backend processes */
3679 0 : if (!OidIsValid(databaseid) ||
3680 0 : proc->databaseId == databaseid)
3681 0 : count++;
3682 : }
3683 :
3684 0 : LWLockRelease(ProcArrayLock);
3685 :
3686 0 : return count;
3687 : }
3688 :
3689 : /*
3690 : * CountUserBackends --- count backends that are used by specified user
3691 : * (only regular backends, not any type of background worker)
3692 : */
3693 : int
3694 0 : CountUserBackends(Oid roleid)
3695 : {
3696 0 : ProcArrayStruct *arrayP = procArray;
3697 0 : int count = 0;
3698 : int index;
3699 :
3700 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3701 :
3702 0 : for (index = 0; index < arrayP->numProcs; index++)
3703 : {
3704 0 : int pgprocno = arrayP->pgprocnos[index];
3705 0 : PGPROC *proc = &allProcs[pgprocno];
3706 :
3707 0 : if (proc->pid == 0)
3708 0 : continue; /* do not count prepared xacts */
3709 0 : if (proc->backendType != B_BACKEND)
3710 0 : continue; /* count only regular backend processes */
3711 0 : if (proc->roleId == roleid)
3712 0 : count++;
3713 : }
3714 :
3715 0 : LWLockRelease(ProcArrayLock);
3716 :
3717 0 : return count;
3718 : }
3719 :
3720 : /*
3721 : * CountOtherDBBackends -- check for other backends running in the given DB
3722 : *
3723 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3724 : * for them to exit (or 0.3s for testing purposes). Autovacuum backends are
3725 : * encouraged to exit early by sending them SIGTERM, but normal user backends
3726 : * are just waited for. If background workers connected to this database are
3727 : * marked as interruptible, they are terminated.
3728 : *
3729 : * The current backend is always ignored; it is caller's responsibility to
3730 : * check whether the current backend uses the given DB, if it's important.
3731 : *
3732 : * Returns true if there are (still) other backends in the DB, false if not.
3733 : * Also, *nbackends and *nprepared are set to the number of other backends
3734 : * and prepared transactions in the DB, respectively.
3735 : *
3736 : * This function is used to interlock DROP DATABASE and related commands
3737 : * against there being any active backends in the target DB --- dropping the
3738 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3739 : * detect here the possibility of a newly-started backend that is trying to
3740 : * connect to the doomed database, so additional interlocking is needed during
3741 : * backend startup. The caller should normally hold an exclusive lock on the
3742 : * target DB before calling this, which is one reason we mustn't wait
3743 : * indefinitely.
3744 : */
3745 : bool
3746 460 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3747 : {
3748 460 : ProcArrayStruct *arrayP = procArray;
3749 :
3750 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3751 : int autovac_pids[MAXAUTOVACPIDS];
3752 :
3753 : /*
3754 : * Retry up to 50 times with 100ms between attempts (max 5s total). Can be
3755 : * reduced to 3 attempts (max 0.3s total) to speed up tests.
3756 : */
3757 460 : int ntries = 50;
3758 :
3759 : #ifdef USE_INJECTION_POINTS
3760 460 : if (IS_INJECTION_POINT_ATTACHED("procarray-reduce-count"))
3761 1 : ntries = 3;
3762 : #endif
3763 :
3764 467 : for (int tries = 0; tries < ntries; tries++)
3765 : {
3766 466 : int nautovacs = 0;
3767 466 : bool found = false;
3768 : int index;
3769 :
3770 466 : CHECK_FOR_INTERRUPTS();
3771 :
3772 466 : *nbackends = *nprepared = 0;
3773 :
3774 466 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3775 :
3776 1739 : for (index = 0; index < arrayP->numProcs; index++)
3777 : {
3778 1273 : int pgprocno = arrayP->pgprocnos[index];
3779 1273 : PGPROC *proc = &allProcs[pgprocno];
3780 1273 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3781 :
3782 1273 : if (proc->databaseId != databaseId)
3783 1163 : continue;
3784 110 : if (proc == MyProc)
3785 103 : continue;
3786 :
3787 7 : found = true;
3788 :
3789 7 : if (proc->pid == 0)
3790 0 : (*nprepared)++;
3791 : else
3792 : {
3793 7 : (*nbackends)++;
3794 7 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3795 : nautovacs < MAXAUTOVACPIDS)
3796 0 : autovac_pids[nautovacs++] = proc->pid;
3797 : }
3798 : }
3799 :
3800 466 : LWLockRelease(ProcArrayLock);
3801 :
3802 466 : if (!found)
3803 459 : return false; /* no conflicting backends, so done */
3804 :
3805 : /*
3806 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3807 : * postpone this step until after the loop because we don't want to
3808 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3809 : * block kill() inside the kernel...
3810 : */
3811 7 : for (index = 0; index < nautovacs; index++)
3812 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3813 :
3814 : /*
3815 : * Terminate all background workers for this database, if they have
3816 : * requested it (BGWORKER_INTERRUPTIBLE).
3817 : */
3818 7 : TerminateBackgroundWorkersForDatabase(databaseId);
3819 :
3820 : /* sleep, then try again */
3821 7 : pg_usleep(100 * 1000L); /* 100ms */
3822 : }
3823 :
3824 1 : return true; /* timed out, still conflicts */
3825 : }
3826 :
3827 : /*
3828 : * Terminate existing connections to the specified database. This routine
3829 : * is used by the DROP DATABASE command when user has asked to forcefully
3830 : * drop the database.
3831 : *
3832 : * The current backend is always ignored; it is caller's responsibility to
3833 : * check whether the current backend uses the given DB, if it's important.
3834 : *
3835 : * If the target database has a prepared transaction or permissions checks
3836 : * fail for a connection, this fails without terminating anything.
3837 : */
3838 : void
3839 1 : TerminateOtherDBBackends(Oid databaseId)
3840 : {
3841 1 : ProcArrayStruct *arrayP = procArray;
3842 1 : List *pids = NIL;
3843 1 : int nprepared = 0;
3844 : int i;
3845 :
3846 1 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3847 :
3848 4 : for (i = 0; i < procArray->numProcs; i++)
3849 : {
3850 3 : int pgprocno = arrayP->pgprocnos[i];
3851 3 : PGPROC *proc = &allProcs[pgprocno];
3852 :
3853 3 : if (proc->databaseId != databaseId)
3854 3 : continue;
3855 0 : if (proc == MyProc)
3856 0 : continue;
3857 :
3858 0 : if (proc->pid != 0)
3859 0 : pids = lappend_int(pids, proc->pid);
3860 : else
3861 0 : nprepared++;
3862 : }
3863 :
3864 1 : LWLockRelease(ProcArrayLock);
3865 :
3866 1 : if (nprepared > 0)
3867 0 : ereport(ERROR,
3868 : (errcode(ERRCODE_OBJECT_IN_USE),
3869 : errmsg("database \"%s\" is being used by prepared transactions",
3870 : get_database_name(databaseId)),
3871 : errdetail_plural("There is %d prepared transaction using the database.",
3872 : "There are %d prepared transactions using the database.",
3873 : nprepared,
3874 : nprepared)));
3875 :
3876 1 : if (pids)
3877 : {
3878 : ListCell *lc;
3879 :
3880 : /*
3881 : * Permissions checks relax the pg_terminate_backend checks in two
3882 : * ways, both by omitting the !OidIsValid(proc->roleId) check:
3883 : *
3884 : * - Accept terminating autovacuum workers, since DROP DATABASE
3885 : * without FORCE terminates them.
3886 : *
3887 : * - Accept terminating bgworkers. For bgworker authors, it's
3888 : * convenient to be able to recommend FORCE if a worker is blocking
3889 : * DROP DATABASE unexpectedly.
3890 : *
3891 : * Unlike pg_terminate_backend, we don't raise some warnings - like
3892 : * "PID %d is not a PostgreSQL server process", because for us already
3893 : * finished session is not a problem.
3894 : */
3895 0 : foreach(lc, pids)
3896 : {
3897 0 : int pid = lfirst_int(lc);
3898 0 : PGPROC *proc = BackendPidGetProc(pid);
3899 :
3900 0 : if (proc != NULL)
3901 : {
3902 0 : if (superuser_arg(proc->roleId) && !superuser())
3903 0 : ereport(ERROR,
3904 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3905 : errmsg("permission denied to terminate process"),
3906 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3907 : "SUPERUSER", "SUPERUSER")));
3908 :
3909 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3910 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3911 0 : ereport(ERROR,
3912 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3913 : errmsg("permission denied to terminate process"),
3914 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3915 : "pg_signal_backend")));
3916 : }
3917 : }
3918 :
3919 : /*
3920 : * There's a race condition here: once we release the ProcArrayLock,
3921 : * it's possible for the session to exit before we issue kill. That
3922 : * race condition possibility seems too unlikely to worry about. See
3923 : * pg_signal_backend.
3924 : */
3925 0 : foreach(lc, pids)
3926 : {
3927 0 : int pid = lfirst_int(lc);
3928 0 : PGPROC *proc = BackendPidGetProc(pid);
3929 :
3930 0 : if (proc != NULL)
3931 : {
3932 : /*
3933 : * If we have setsid(), signal the backend's whole process
3934 : * group
3935 : */
3936 : #ifdef HAVE_SETSID
3937 0 : (void) kill(-pid, SIGTERM);
3938 : #else
3939 : (void) kill(pid, SIGTERM);
3940 : #endif
3941 : }
3942 : }
3943 : }
3944 1 : }
3945 :
3946 : /*
3947 : * ProcArraySetReplicationSlotXmin
3948 : *
3949 : * Install limits to future computations of the xmin horizon to prevent vacuum
3950 : * and HOT pruning from removing affected rows still needed by clients with
3951 : * replication slots.
3952 : */
3953 : void
3954 2452 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3955 : bool already_locked)
3956 : {
3957 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3958 :
3959 2452 : if (!already_locked)
3960 1965 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3961 :
3962 2452 : procArray->replication_slot_xmin = xmin;
3963 2452 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3964 :
3965 2452 : if (!already_locked)
3966 1965 : LWLockRelease(ProcArrayLock);
3967 :
3968 2452 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3969 : xmin, catalog_xmin);
3970 2452 : }
3971 :
3972 : /*
3973 : * ProcArrayGetReplicationSlotXmin
3974 : *
3975 : * Return the current slot xmin limits. That's useful to be able to remove
3976 : * data that's older than those limits.
3977 : */
3978 : void
3979 22 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3980 : TransactionId *catalog_xmin)
3981 : {
3982 22 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3983 :
3984 22 : if (xmin != NULL)
3985 0 : *xmin = procArray->replication_slot_xmin;
3986 :
3987 22 : if (catalog_xmin != NULL)
3988 22 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3989 :
3990 22 : LWLockRelease(ProcArrayLock);
3991 22 : }
3992 :
3993 : /*
3994 : * XidCacheRemoveRunningXids
3995 : *
3996 : * Remove a bunch of TransactionIds from the list of known-running
3997 : * subtransactions for my backend. Both the specified xid and those in
3998 : * the xids[] array (of length nxids) are removed from the subxids cache.
3999 : * latestXid must be the latest XID among the group.
4000 : */
4001 : void
4002 667 : XidCacheRemoveRunningXids(TransactionId xid,
4003 : int nxids, const TransactionId *xids,
4004 : TransactionId latestXid)
4005 : {
4006 : int i,
4007 : j;
4008 : XidCacheStatus *mysubxidstat;
4009 :
4010 : Assert(TransactionIdIsValid(xid));
4011 :
4012 : /*
4013 : * We must hold ProcArrayLock exclusively in order to remove transactions
4014 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
4015 : * possible this could be relaxed since we know this routine is only used
4016 : * to abort subtransactions, but pending closer analysis we'd best be
4017 : * conservative.
4018 : *
4019 : * Note that we do not have to be careful about memory ordering of our own
4020 : * reads wrt. GetNewTransactionId() here - only this process can modify
4021 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
4022 : * careful about our own writes being well ordered.
4023 : */
4024 667 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4025 :
4026 667 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
4027 :
4028 : /*
4029 : * Under normal circumstances xid and xids[] will be in increasing order,
4030 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
4031 : * behavior when removing a lot of xids.
4032 : */
4033 697 : for (i = nxids - 1; i >= 0; i--)
4034 : {
4035 30 : TransactionId anxid = xids[i];
4036 :
4037 30 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4038 : {
4039 30 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4040 : {
4041 30 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4042 30 : pg_write_barrier();
4043 30 : mysubxidstat->count--;
4044 30 : MyProc->subxidStatus.count--;
4045 30 : break;
4046 : }
4047 : }
4048 :
4049 : /*
4050 : * Ordinarily we should have found it, unless the cache has
4051 : * overflowed. However it's also possible for this routine to be
4052 : * invoked multiple times for the same subtransaction, in case of an
4053 : * error during AbortSubTransaction. So instead of Assert, emit a
4054 : * debug warning.
4055 : */
4056 30 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4057 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4058 : }
4059 :
4060 731 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4061 : {
4062 730 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4063 : {
4064 666 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4065 666 : pg_write_barrier();
4066 666 : mysubxidstat->count--;
4067 666 : MyProc->subxidStatus.count--;
4068 666 : break;
4069 : }
4070 : }
4071 : /* Ordinarily we should have found it, unless the cache has overflowed */
4072 667 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4073 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4074 :
4075 : /* Also advance global latestCompletedXid while holding the lock */
4076 667 : MaintainLatestCompletedXid(latestXid);
4077 :
4078 : /* ... and xactCompletionCount */
4079 667 : TransamVariables->xactCompletionCount++;
4080 :
4081 667 : LWLockRelease(ProcArrayLock);
4082 667 : }
4083 :
4084 : #ifdef XIDCACHE_DEBUG
4085 :
4086 : /*
4087 : * Print stats about effectiveness of XID cache
4088 : */
4089 : static void
4090 : DisplayXidCache(void)
4091 : {
4092 : fprintf(stderr,
4093 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4094 : xc_by_recent_xmin,
4095 : xc_by_known_xact,
4096 : xc_by_my_xact,
4097 : xc_by_latest_xid,
4098 : xc_by_main_xid,
4099 : xc_by_child_xid,
4100 : xc_by_known_assigned,
4101 : xc_no_overflow,
4102 : xc_slow_answer);
4103 : }
4104 : #endif /* XIDCACHE_DEBUG */
4105 :
4106 : /*
4107 : * If rel != NULL, return test state appropriate for relation, otherwise
4108 : * return state usable for all relations. The latter may consider XIDs as
4109 : * not-yet-visible-to-everyone that a state for a specific relation would
4110 : * already consider visible-to-everyone.
4111 : *
4112 : * This needs to be called while a snapshot is active or registered, otherwise
4113 : * there are wraparound and other dangers.
4114 : *
4115 : * See comment for GlobalVisState for details.
4116 : */
4117 : GlobalVisState *
4118 14454066 : GlobalVisTestFor(Relation rel)
4119 : {
4120 14454066 : GlobalVisState *state = NULL;
4121 :
4122 : /* XXX: we should assert that a snapshot is pushed or registered */
4123 : Assert(RecentXmin);
4124 :
4125 14454066 : switch (GlobalVisHorizonKindForRel(rel))
4126 : {
4127 88854 : case VISHORIZON_SHARED:
4128 88854 : state = &GlobalVisSharedRels;
4129 88854 : break;
4130 3019086 : case VISHORIZON_CATALOG:
4131 3019086 : state = &GlobalVisCatalogRels;
4132 3019086 : break;
4133 11304166 : case VISHORIZON_DATA:
4134 11304166 : state = &GlobalVisDataRels;
4135 11304166 : break;
4136 41960 : case VISHORIZON_TEMP:
4137 41960 : state = &GlobalVisTempRels;
4138 41960 : break;
4139 : }
4140 :
4141 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4142 : FullTransactionIdIsValid(state->maybe_needed));
4143 :
4144 14454066 : return state;
4145 : }
4146 :
4147 : /*
4148 : * Return true if it's worth updating the accurate maybe_needed boundary.
4149 : *
4150 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4151 : * repeatedly do so when there is a low likelihood of it being beneficial.
4152 : *
4153 : * The current heuristic is that we update only if RecentXmin has changed
4154 : * since the last update. If the oldest currently running transaction has not
4155 : * finished, it is unlikely that recomputing the horizon would be useful.
4156 : */
4157 : static bool
4158 450070 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4159 : {
4160 : /* hasn't been updated yet */
4161 450070 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4162 8998 : return true;
4163 :
4164 : /*
4165 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4166 : * unlikely to be beneficial to refresh boundaries.
4167 : */
4168 441072 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4169 : state->definitely_needed))
4170 0 : return false;
4171 :
4172 : /* does the last snapshot built have a different xmin? */
4173 441072 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4174 : }
4175 :
4176 : static void
4177 206317 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4178 : {
4179 : GlobalVisSharedRels.maybe_needed =
4180 206317 : FullXidRelativeTo(horizons->latest_completed,
4181 : horizons->shared_oldest_nonremovable);
4182 : GlobalVisCatalogRels.maybe_needed =
4183 206317 : FullXidRelativeTo(horizons->latest_completed,
4184 : horizons->catalog_oldest_nonremovable);
4185 : GlobalVisDataRels.maybe_needed =
4186 206317 : FullXidRelativeTo(horizons->latest_completed,
4187 : horizons->data_oldest_nonremovable);
4188 : GlobalVisTempRels.maybe_needed =
4189 206317 : FullXidRelativeTo(horizons->latest_completed,
4190 : horizons->temp_oldest_nonremovable);
4191 :
4192 : /*
4193 : * In longer running transactions it's possible that transactions we
4194 : * previously needed to treat as running aren't around anymore. So update
4195 : * definitely_needed to not be earlier than maybe_needed.
4196 : */
4197 : GlobalVisSharedRels.definitely_needed =
4198 206317 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4199 : GlobalVisSharedRels.definitely_needed);
4200 : GlobalVisCatalogRels.definitely_needed =
4201 206317 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4202 : GlobalVisCatalogRels.definitely_needed);
4203 : GlobalVisDataRels.definitely_needed =
4204 206317 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4205 : GlobalVisDataRels.definitely_needed);
4206 206317 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4207 :
4208 206317 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4209 206317 : }
4210 :
4211 : /*
4212 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4213 : * using ComputeXidHorizons().
4214 : */
4215 : static void
4216 55423 : GlobalVisUpdate(void)
4217 : {
4218 : ComputeXidHorizonsResult horizons;
4219 :
4220 : /* updates the horizons as a side-effect */
4221 55423 : ComputeXidHorizons(&horizons);
4222 55423 : }
4223 :
4224 : /*
4225 : * Return true if no snapshot still considers fxid to be running.
4226 : *
4227 : * The state passed needs to have been initialized for the relation fxid is
4228 : * from (NULL is also OK), otherwise the result may not be correct.
4229 : *
4230 : * See comment for GlobalVisState for details.
4231 : */
4232 : bool
4233 9671260 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4234 : FullTransactionId fxid)
4235 : {
4236 : /*
4237 : * If fxid is older than maybe_needed bound, it definitely is visible to
4238 : * everyone.
4239 : */
4240 9671260 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4241 2473643 : return true;
4242 :
4243 : /*
4244 : * If fxid is >= definitely_needed bound, it is very likely to still be
4245 : * considered running.
4246 : */
4247 7197617 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4248 6747547 : return false;
4249 :
4250 : /*
4251 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4252 : * might not exist a snapshot considering fxid running. If it makes sense,
4253 : * update boundaries and recheck.
4254 : */
4255 450070 : if (GlobalVisTestShouldUpdate(state))
4256 : {
4257 55423 : GlobalVisUpdate();
4258 :
4259 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4260 :
4261 55423 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4262 : }
4263 : else
4264 394647 : return false;
4265 : }
4266 :
4267 : /*
4268 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4269 : *
4270 : * It is crucial that this only gets called for xids from a source that
4271 : * protects against xid wraparounds (e.g. from a table and thus protected by
4272 : * relfrozenxid).
4273 : */
4274 : bool
4275 9670766 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4276 : {
4277 : FullTransactionId fxid;
4278 :
4279 : /*
4280 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4281 : * because we know the xid has to, at the very least, be between
4282 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4283 : * lock to determine either, we can just compare with
4284 : * state->definitely_needed, which was based on those value at the time
4285 : * the current snapshot was built.
4286 : */
4287 9670766 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4288 :
4289 9670766 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4290 : }
4291 :
4292 : /*
4293 : * Convenience wrapper around GlobalVisTestFor() and
4294 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4295 : */
4296 : bool
4297 494 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4298 : {
4299 : GlobalVisState *state;
4300 :
4301 494 : state = GlobalVisTestFor(rel);
4302 :
4303 494 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4304 : }
4305 :
4306 : /*
4307 : * Convenience wrapper around GlobalVisTestFor() and
4308 : * GlobalVisTestIsRemovableXid(), see their comments.
4309 : */
4310 : bool
4311 6 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4312 : {
4313 : GlobalVisState *state;
4314 :
4315 6 : state = GlobalVisTestFor(rel);
4316 :
4317 6 : return GlobalVisTestIsRemovableXid(state, xid);
4318 : }
4319 :
4320 : /*
4321 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4322 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4323 : *
4324 : * Be very careful about when to use this function. It can only safely be used
4325 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4326 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4327 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4328 : * the xid has to be within that range), or if xid is from the procarray and
4329 : * prevents xid wraparound that way.
4330 : */
4331 : static inline FullTransactionId
4332 11896176 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4333 : {
4334 11896176 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4335 :
4336 : Assert(TransactionIdIsValid(xid));
4337 : Assert(TransactionIdIsValid(rel_xid));
4338 :
4339 : /* not guaranteed to find issues, but likely to catch mistakes */
4340 : AssertTransactionIdInAllowableRange(xid);
4341 :
4342 23792352 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4343 11896176 : + (int32) (xid - rel_xid));
4344 : }
4345 :
4346 :
4347 : /* ----------------------------------------------
4348 : * KnownAssignedTransactionIds sub-module
4349 : * ----------------------------------------------
4350 : */
4351 :
4352 : /*
4353 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4354 : * running on the primary at the current point in WAL. These XIDs must be
4355 : * treated as running by standby transactions, even though they are not in
4356 : * the standby server's PGPROC array.
4357 : *
4358 : * We record all XIDs that we know have been assigned. That includes all the
4359 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4360 : * been assigned. We can deduce the existence of unobserved XIDs because we
4361 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4362 : * list expands as new XIDs are observed or inferred, and contracts when
4363 : * transaction completion records arrive.
4364 : *
4365 : * During hot standby we do not fret too much about the distinction between
4366 : * top-level XIDs and subtransaction XIDs. We store both together in the
4367 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4368 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4369 : * doesn't care about the distinction either. Subtransaction XIDs are
4370 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4371 : * links are *not* maintained (which does not affect visibility).
4372 : *
4373 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4374 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4375 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4376 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4377 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4378 : * and then remove them from KnownAssignedXids. This prevents overflow of
4379 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4380 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4381 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4382 : * though it should be complete for top-level XIDs; this is the same situation
4383 : * that holds with respect to the PGPROC entries in normal running.
4384 : *
4385 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4386 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4387 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4388 : * As long as that is within the range of interesting XIDs, we have to assume
4389 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4390 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4391 : * subXID arrives - that is not an error.)
4392 : *
4393 : * Should a backend on primary somehow disappear before it can write an abort
4394 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4395 : * aborted but we think they were running; the distinction is irrelevant
4396 : * because either way any changes done by the transaction are not visible to
4397 : * backends in the standby. We prune KnownAssignedXids when
4398 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4399 : * array due to such dead XIDs.
4400 : */
4401 :
4402 : /*
4403 : * RecordKnownAssignedTransactionIds
4404 : * Record the given XID in KnownAssignedXids, as well as any preceding
4405 : * unobserved XIDs.
4406 : *
4407 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4408 : * associated with a transaction. Must be called for each record after we
4409 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4410 : *
4411 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4412 : */
4413 : void
4414 2518840 : RecordKnownAssignedTransactionIds(TransactionId xid)
4415 : {
4416 : Assert(standbyState >= STANDBY_INITIALIZED);
4417 : Assert(TransactionIdIsValid(xid));
4418 : Assert(TransactionIdIsValid(latestObservedXid));
4419 :
4420 2518840 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4421 : xid, latestObservedXid);
4422 :
4423 : /*
4424 : * When a newly observed xid arrives, it is frequently the case that it is
4425 : * *not* the next xid in sequence. When this occurs, we must treat the
4426 : * intervening xids as running also.
4427 : */
4428 2518840 : if (TransactionIdFollows(xid, latestObservedXid))
4429 : {
4430 : TransactionId next_expected_xid;
4431 :
4432 : /*
4433 : * Extend subtrans like we do in GetNewTransactionId() during normal
4434 : * operation using individual extend steps. Note that we do not need
4435 : * to extend clog since its extensions are WAL logged.
4436 : *
4437 : * This part has to be done regardless of standbyState since we
4438 : * immediately start assigning subtransactions to their toplevel
4439 : * transactions.
4440 : */
4441 22967 : next_expected_xid = latestObservedXid;
4442 46880 : while (TransactionIdPrecedes(next_expected_xid, xid))
4443 : {
4444 23913 : TransactionIdAdvance(next_expected_xid);
4445 23913 : ExtendSUBTRANS(next_expected_xid);
4446 : }
4447 : Assert(next_expected_xid == xid);
4448 :
4449 : /*
4450 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4451 : * more to do since we don't track assigned xids yet.
4452 : */
4453 22967 : if (standbyState <= STANDBY_INITIALIZED)
4454 : {
4455 0 : latestObservedXid = xid;
4456 0 : return;
4457 : }
4458 :
4459 : /*
4460 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4461 : */
4462 22967 : next_expected_xid = latestObservedXid;
4463 22967 : TransactionIdAdvance(next_expected_xid);
4464 22967 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4465 :
4466 : /*
4467 : * Now we can advance latestObservedXid
4468 : */
4469 22967 : latestObservedXid = xid;
4470 :
4471 : /* TransamVariables->nextXid must be beyond any observed xid */
4472 22967 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4473 : }
4474 : }
4475 :
4476 : /*
4477 : * ExpireTreeKnownAssignedTransactionIds
4478 : * Remove the given XIDs from KnownAssignedXids.
4479 : *
4480 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4481 : */
4482 : void
4483 22379 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4484 : TransactionId *subxids, TransactionId max_xid)
4485 : {
4486 : Assert(standbyState >= STANDBY_INITIALIZED);
4487 :
4488 : /*
4489 : * Uses same locking as transaction commit
4490 : */
4491 22379 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4492 :
4493 22379 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4494 :
4495 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4496 22379 : MaintainLatestCompletedXidRecovery(max_xid);
4497 :
4498 : /* ... and xactCompletionCount */
4499 22379 : TransamVariables->xactCompletionCount++;
4500 :
4501 22379 : LWLockRelease(ProcArrayLock);
4502 22379 : }
4503 :
4504 : /*
4505 : * ExpireAllKnownAssignedTransactionIds
4506 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4507 : */
4508 : void
4509 114 : ExpireAllKnownAssignedTransactionIds(void)
4510 : {
4511 : FullTransactionId latestXid;
4512 :
4513 114 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4514 114 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4515 :
4516 : /* Reset latestCompletedXid to nextXid - 1 */
4517 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
4518 114 : latestXid = TransamVariables->nextXid;
4519 114 : FullTransactionIdRetreat(&latestXid);
4520 114 : TransamVariables->latestCompletedXid = latestXid;
4521 :
4522 : /*
4523 : * Any transactions that were in-progress were effectively aborted, so
4524 : * advance xactCompletionCount.
4525 : */
4526 114 : TransamVariables->xactCompletionCount++;
4527 :
4528 : /*
4529 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4530 : * the call of this function. But do this for unification with what
4531 : * ExpireOldKnownAssignedTransactionIds() do.
4532 : */
4533 114 : procArray->lastOverflowedXid = InvalidTransactionId;
4534 114 : LWLockRelease(ProcArrayLock);
4535 114 : }
4536 :
4537 : /*
4538 : * ExpireOldKnownAssignedTransactionIds
4539 : * Remove KnownAssignedXids entries preceding the given XID and
4540 : * potentially reset lastOverflowedXid.
4541 : */
4542 : void
4543 822 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4544 : {
4545 : TransactionId latestXid;
4546 :
4547 822 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4548 :
4549 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4550 822 : latestXid = xid;
4551 822 : TransactionIdRetreat(latestXid);
4552 822 : MaintainLatestCompletedXidRecovery(latestXid);
4553 :
4554 : /* ... and xactCompletionCount */
4555 822 : TransamVariables->xactCompletionCount++;
4556 :
4557 : /*
4558 : * Reset lastOverflowedXid if we know all transactions that have been
4559 : * possibly running are being gone. Not doing so could cause an incorrect
4560 : * lastOverflowedXid value, which makes extra snapshots be marked as
4561 : * suboverflowed.
4562 : */
4563 822 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4564 814 : procArray->lastOverflowedXid = InvalidTransactionId;
4565 822 : KnownAssignedXidsRemovePreceding(xid);
4566 822 : LWLockRelease(ProcArrayLock);
4567 822 : }
4568 :
4569 : /*
4570 : * KnownAssignedTransactionIdsIdleMaintenance
4571 : * Opportunistically do maintenance work when the startup process
4572 : * is about to go idle.
4573 : */
4574 : void
4575 7515 : KnownAssignedTransactionIdsIdleMaintenance(void)
4576 : {
4577 7515 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4578 7515 : }
4579 :
4580 :
4581 : /*
4582 : * Private module functions to manipulate KnownAssignedXids
4583 : *
4584 : * There are 5 main uses of the KnownAssignedXids data structure:
4585 : *
4586 : * * backends taking snapshots - all valid XIDs need to be copied out
4587 : * * backends seeking to determine presence of a specific XID
4588 : * * startup process adding new known-assigned XIDs
4589 : * * startup process removing specific XIDs as transactions end
4590 : * * startup process pruning array when special WAL records arrive
4591 : *
4592 : * This data structure is known to be a hot spot during Hot Standby, so we
4593 : * go to some lengths to make these operations as efficient and as concurrent
4594 : * as possible.
4595 : *
4596 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4597 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4598 : * in general TransactionIdPrecedes would not provide a total order, but
4599 : * we know that the entries present at any instant should not extend across
4600 : * a large enough fraction of XID space to wrap around (the primary would
4601 : * shut down for fear of XID wrap long before that happens). So it's OK to
4602 : * use TransactionIdPrecedes as a binary-search comparator.
4603 : *
4604 : * It's cheap to maintain the sortedness during insertions, since new known
4605 : * XIDs are always reported in XID order; we just append them at the right.
4606 : *
4607 : * To keep individual deletions cheap, we need to allow gaps in the array.
4608 : * This is implemented by marking array elements as valid or invalid using
4609 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4610 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4611 : * XID entry itself. This preserves the property that the XID entries are
4612 : * sorted, so we can do binary searches easily. Periodically we compress
4613 : * out the unused entries; that's much cheaper than having to compress the
4614 : * array immediately on every deletion.
4615 : *
4616 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4617 : * are those with indexes tail <= i < head; items outside this subscript range
4618 : * have unspecified contents. When head reaches the end of the array, we
4619 : * force compression of unused entries rather than wrapping around, since
4620 : * allowing wraparound would greatly complicate the search logic. We maintain
4621 : * an explicit tail pointer so that pruning of old XIDs can be done without
4622 : * immediately moving the array contents. In most cases only a small fraction
4623 : * of the array contains valid entries at any instant.
4624 : *
4625 : * Although only the startup process can ever change the KnownAssignedXids
4626 : * data structure, we still need interlocking so that standby backends will
4627 : * not observe invalid intermediate states. The convention is that backends
4628 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4629 : * the array, the startup process must hold ProcArrayLock exclusively, for
4630 : * the usual transactional reasons (compare commit/abort of a transaction
4631 : * during normal running). Compressing unused entries out of the array
4632 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4633 : * them into slots to the right of the head pointer and then advance the head
4634 : * pointer. This doesn't require any lock at all, but on machines with weak
4635 : * memory ordering, we need to be careful that other processors see the array
4636 : * element changes before they see the head pointer change. We handle this by
4637 : * using memory barriers when reading or writing the head/tail pointers (unless
4638 : * the caller holds ProcArrayLock exclusively).
4639 : *
4640 : * Algorithmic analysis:
4641 : *
4642 : * If we have a maximum of M slots, with N XIDs currently spread across
4643 : * S elements then we have N <= S <= M always.
4644 : *
4645 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4646 : * happen)
4647 : * * Compressing the array is O(S) and requires exclusive lock
4648 : * * Removing an XID is O(logS) and requires exclusive lock
4649 : * * Taking a snapshot is O(S) and requires shared lock
4650 : * * Checking for an XID is O(logS) and requires shared lock
4651 : *
4652 : * In comparison, using a hash table for KnownAssignedXids would mean that
4653 : * taking snapshots would be O(M). If we can maintain S << M then the
4654 : * sorted array technique will deliver significantly faster snapshots.
4655 : * If we try to keep S too small then we will spend too much time compressing,
4656 : * so there is an optimal point for any workload mix. We use a heuristic to
4657 : * decide when to compress the array, though trimming also helps reduce
4658 : * frequency of compressing. The heuristic requires us to track the number of
4659 : * currently valid XIDs in the array (N). Except in special cases, we'll
4660 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4661 : * taking a snapshot to be O(N), which it would have to be anyway.
4662 : */
4663 :
4664 :
4665 : /*
4666 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4667 : * array, removing any gaps.
4668 : *
4669 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4670 : * we do it only if a heuristic indicates it's a good time to do it.
4671 : *
4672 : * Compression requires holding ProcArrayLock in exclusive mode.
4673 : * Caller must pass haveLock = true if it already holds the lock.
4674 : */
4675 : static void
4676 30737 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4677 : {
4678 30737 : ProcArrayStruct *pArray = procArray;
4679 : int head,
4680 : tail,
4681 : nelements;
4682 : int compress_index;
4683 : int i;
4684 :
4685 : /* Counters for compression heuristics */
4686 : static unsigned int transactionEndsCounter;
4687 : static TimestampTz lastCompressTs;
4688 :
4689 : /* Tuning constants */
4690 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4691 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4692 :
4693 : /*
4694 : * Since only the startup process modifies the head/tail pointers, we
4695 : * don't need a lock to read them here.
4696 : */
4697 30737 : head = pArray->headKnownAssignedXids;
4698 30737 : tail = pArray->tailKnownAssignedXids;
4699 30737 : nelements = head - tail;
4700 :
4701 : /*
4702 : * If we can choose whether to compress, use a heuristic to avoid
4703 : * compressing too often or not often enough. "Compress" here simply
4704 : * means moving the values to the beginning of the array, so it is not as
4705 : * complex or costly as typical data compression algorithms.
4706 : */
4707 30737 : if (nelements == pArray->numKnownAssignedXids)
4708 : {
4709 : /*
4710 : * When there are no gaps between head and tail, don't bother to
4711 : * compress, except in the KAX_NO_SPACE case where we must compress to
4712 : * create some space after the head.
4713 : */
4714 14366 : if (reason != KAX_NO_SPACE)
4715 14366 : return;
4716 : }
4717 16371 : else if (reason == KAX_TRANSACTION_END)
4718 : {
4719 : /*
4720 : * Consider compressing only once every so many commits. Frequency
4721 : * determined by benchmarks.
4722 : */
4723 13926 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4724 13808 : return;
4725 :
4726 : /*
4727 : * Furthermore, compress only if the used part of the array is less
4728 : * than 50% full (see comments above).
4729 : */
4730 118 : if (nelements < 2 * pArray->numKnownAssignedXids)
4731 9 : return;
4732 : }
4733 2445 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4734 : {
4735 : /*
4736 : * We're about to go idle for lack of new WAL, so we might as well
4737 : * compress. But not too often, to avoid ProcArray lock contention
4738 : * with readers.
4739 : */
4740 2330 : if (lastCompressTs != 0)
4741 : {
4742 : TimestampTz compress_after;
4743 :
4744 2329 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4745 : KAX_COMPRESS_IDLE_INTERVAL);
4746 2329 : if (GetCurrentTimestamp() < compress_after)
4747 2284 : return;
4748 : }
4749 : }
4750 :
4751 : /* Need to compress, so get the lock if we don't have it. */
4752 270 : if (!haveLock)
4753 46 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4754 :
4755 : /*
4756 : * We compress the array by reading the valid values from tail to head,
4757 : * re-aligning data to 0th element.
4758 : */
4759 270 : compress_index = 0;
4760 10072 : for (i = tail; i < head; i++)
4761 : {
4762 9802 : if (KnownAssignedXidsValid[i])
4763 : {
4764 1112 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4765 1112 : KnownAssignedXidsValid[compress_index] = true;
4766 1112 : compress_index++;
4767 : }
4768 : }
4769 : Assert(compress_index == pArray->numKnownAssignedXids);
4770 :
4771 270 : pArray->tailKnownAssignedXids = 0;
4772 270 : pArray->headKnownAssignedXids = compress_index;
4773 :
4774 270 : if (!haveLock)
4775 46 : LWLockRelease(ProcArrayLock);
4776 :
4777 : /* Update timestamp for maintenance. No need to hold lock for this. */
4778 270 : lastCompressTs = GetCurrentTimestamp();
4779 : }
4780 :
4781 : /*
4782 : * Add xids into KnownAssignedXids at the head of the array.
4783 : *
4784 : * xids from from_xid to to_xid, inclusive, are added to the array.
4785 : *
4786 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4787 : * exclusive mode, so we need no extra locking here. Else caller holds no
4788 : * lock, so we need to be sure we maintain sufficient interlocks against
4789 : * concurrent readers. (Only the startup process ever calls this, so no need
4790 : * to worry about concurrent writers.)
4791 : */
4792 : static void
4793 22971 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4794 : bool exclusive_lock)
4795 : {
4796 22971 : ProcArrayStruct *pArray = procArray;
4797 : TransactionId next_xid;
4798 : int head,
4799 : tail;
4800 : int nxids;
4801 : int i;
4802 :
4803 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4804 :
4805 : /*
4806 : * Calculate how many array slots we'll need. Normally this is cheap; in
4807 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4808 : * way.
4809 : */
4810 22971 : if (to_xid >= from_xid)
4811 22971 : nxids = to_xid - from_xid + 1;
4812 : else
4813 : {
4814 0 : nxids = 1;
4815 0 : next_xid = from_xid;
4816 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4817 : {
4818 0 : nxids++;
4819 0 : TransactionIdAdvance(next_xid);
4820 : }
4821 : }
4822 :
4823 : /*
4824 : * Since only the startup process modifies the head/tail pointers, we
4825 : * don't need a lock to read them here.
4826 : */
4827 22971 : head = pArray->headKnownAssignedXids;
4828 22971 : tail = pArray->tailKnownAssignedXids;
4829 :
4830 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4831 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4832 :
4833 : /*
4834 : * Verify that insertions occur in TransactionId sequence. Note that even
4835 : * if the last existing element is marked invalid, it must still have a
4836 : * correctly sequenced XID value.
4837 : */
4838 38003 : if (head > tail &&
4839 15032 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4840 : {
4841 0 : KnownAssignedXidsDisplay(LOG);
4842 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4843 : }
4844 :
4845 : /*
4846 : * If our xids won't fit in the remaining space, compress out free space
4847 : */
4848 22971 : if (head + nxids > pArray->maxKnownAssignedXids)
4849 : {
4850 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4851 :
4852 0 : head = pArray->headKnownAssignedXids;
4853 : /* note: we no longer care about the tail pointer */
4854 :
4855 : /*
4856 : * If it still won't fit then we're out of memory
4857 : */
4858 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4859 0 : elog(ERROR, "too many KnownAssignedXids");
4860 : }
4861 :
4862 : /* Now we can insert the xids into the space starting at head */
4863 22971 : next_xid = from_xid;
4864 46888 : for (i = 0; i < nxids; i++)
4865 : {
4866 23917 : KnownAssignedXids[head] = next_xid;
4867 23917 : KnownAssignedXidsValid[head] = true;
4868 23917 : TransactionIdAdvance(next_xid);
4869 23917 : head++;
4870 : }
4871 :
4872 : /* Adjust count of number of valid entries */
4873 22971 : pArray->numKnownAssignedXids += nxids;
4874 :
4875 : /*
4876 : * Now update the head pointer. We use a write barrier to ensure that
4877 : * other processors see the above array updates before they see the head
4878 : * pointer change. The barrier isn't required if we're holding
4879 : * ProcArrayLock exclusively.
4880 : */
4881 22971 : if (!exclusive_lock)
4882 22967 : pg_write_barrier();
4883 :
4884 22971 : pArray->headKnownAssignedXids = head;
4885 22971 : }
4886 :
4887 : /*
4888 : * KnownAssignedXidsSearch
4889 : *
4890 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4891 : * Returns true if it was found, false if not.
4892 : *
4893 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4894 : * Exclusive lock must be held for remove = true.
4895 : */
4896 : static bool
4897 25051 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4898 : {
4899 25051 : ProcArrayStruct *pArray = procArray;
4900 : int first,
4901 : last;
4902 : int head;
4903 : int tail;
4904 25051 : int result_index = -1;
4905 :
4906 25051 : tail = pArray->tailKnownAssignedXids;
4907 25051 : head = pArray->headKnownAssignedXids;
4908 :
4909 : /*
4910 : * Only the startup process removes entries, so we don't need the read
4911 : * barrier in that case.
4912 : */
4913 25051 : if (!remove)
4914 1 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4915 :
4916 : /*
4917 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4918 : * array here, since even invalid entries will contain sorted XIDs.
4919 : */
4920 25051 : first = tail;
4921 25051 : last = head - 1;
4922 86890 : while (first <= last)
4923 : {
4924 : int mid_index;
4925 : TransactionId mid_xid;
4926 :
4927 85700 : mid_index = (first + last) / 2;
4928 85700 : mid_xid = KnownAssignedXids[mid_index];
4929 :
4930 85700 : if (xid == mid_xid)
4931 : {
4932 23861 : result_index = mid_index;
4933 23861 : break;
4934 : }
4935 61839 : else if (TransactionIdPrecedes(xid, mid_xid))
4936 13385 : last = mid_index - 1;
4937 : else
4938 48454 : first = mid_index + 1;
4939 : }
4940 :
4941 25051 : if (result_index < 0)
4942 1190 : return false; /* not in array */
4943 :
4944 23861 : if (!KnownAssignedXidsValid[result_index])
4945 30 : return false; /* in array, but invalid */
4946 :
4947 23831 : if (remove)
4948 : {
4949 23831 : KnownAssignedXidsValid[result_index] = false;
4950 :
4951 23831 : pArray->numKnownAssignedXids--;
4952 : Assert(pArray->numKnownAssignedXids >= 0);
4953 :
4954 : /*
4955 : * If we're removing the tail element then advance tail pointer over
4956 : * any invalid elements. This will speed future searches.
4957 : */
4958 23831 : if (result_index == tail)
4959 : {
4960 8968 : tail++;
4961 15141 : while (tail < head && !KnownAssignedXidsValid[tail])
4962 6173 : tail++;
4963 8968 : if (tail >= head)
4964 : {
4965 : /* Array is empty, so we can reset both pointers */
4966 7929 : pArray->headKnownAssignedXids = 0;
4967 7929 : pArray->tailKnownAssignedXids = 0;
4968 : }
4969 : else
4970 : {
4971 1039 : pArray->tailKnownAssignedXids = tail;
4972 : }
4973 : }
4974 : }
4975 :
4976 23831 : return true;
4977 : }
4978 :
4979 : /*
4980 : * Is the specified XID present in KnownAssignedXids[]?
4981 : *
4982 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4983 : */
4984 : static bool
4985 1 : KnownAssignedXidExists(TransactionId xid)
4986 : {
4987 : Assert(TransactionIdIsValid(xid));
4988 :
4989 1 : return KnownAssignedXidsSearch(xid, false);
4990 : }
4991 :
4992 : /*
4993 : * Remove the specified XID from KnownAssignedXids[].
4994 : *
4995 : * Caller must hold ProcArrayLock in exclusive mode.
4996 : */
4997 : static void
4998 25050 : KnownAssignedXidsRemove(TransactionId xid)
4999 : {
5000 : Assert(TransactionIdIsValid(xid));
5001 :
5002 25050 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
5003 :
5004 : /*
5005 : * Note: we cannot consider it an error to remove an XID that's not
5006 : * present. We intentionally remove subxact IDs while processing
5007 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
5008 : * removed again when the top-level xact commits or aborts.
5009 : *
5010 : * It might be possible to track such XIDs to distinguish this case from
5011 : * actual errors, but it would be complicated and probably not worth it.
5012 : * So, just ignore the search result.
5013 : */
5014 25050 : (void) KnownAssignedXidsSearch(xid, true);
5015 25050 : }
5016 :
5017 : /*
5018 : * KnownAssignedXidsRemoveTree
5019 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
5020 : *
5021 : * Caller must hold ProcArrayLock in exclusive mode.
5022 : */
5023 : static void
5024 22400 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
5025 : TransactionId *subxids)
5026 : {
5027 : int i;
5028 :
5029 22400 : if (TransactionIdIsValid(xid))
5030 22379 : KnownAssignedXidsRemove(xid);
5031 :
5032 25071 : for (i = 0; i < nsubxids; i++)
5033 2671 : KnownAssignedXidsRemove(subxids[i]);
5034 :
5035 : /* Opportunistically compress the array */
5036 22400 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5037 22400 : }
5038 :
5039 : /*
5040 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5041 : * then clear the whole table.
5042 : *
5043 : * Caller must hold ProcArrayLock in exclusive mode.
5044 : */
5045 : static void
5046 936 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5047 : {
5048 936 : ProcArrayStruct *pArray = procArray;
5049 936 : int count = 0;
5050 : int head,
5051 : tail,
5052 : i;
5053 :
5054 936 : if (!TransactionIdIsValid(removeXid))
5055 : {
5056 114 : elog(DEBUG4, "removing all KnownAssignedXids");
5057 114 : pArray->numKnownAssignedXids = 0;
5058 114 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5059 114 : return;
5060 : }
5061 :
5062 822 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5063 :
5064 : /*
5065 : * Mark entries invalid starting at the tail. Since array is sorted, we
5066 : * can stop as soon as we reach an entry >= removeXid.
5067 : */
5068 822 : tail = pArray->tailKnownAssignedXids;
5069 822 : head = pArray->headKnownAssignedXids;
5070 :
5071 822 : for (i = tail; i < head; i++)
5072 : {
5073 207 : if (KnownAssignedXidsValid[i])
5074 : {
5075 207 : TransactionId knownXid = KnownAssignedXids[i];
5076 :
5077 207 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5078 207 : break;
5079 :
5080 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5081 : {
5082 0 : KnownAssignedXidsValid[i] = false;
5083 0 : count++;
5084 : }
5085 : }
5086 : }
5087 :
5088 822 : pArray->numKnownAssignedXids -= count;
5089 : Assert(pArray->numKnownAssignedXids >= 0);
5090 :
5091 : /*
5092 : * Advance the tail pointer if we've marked the tail item invalid.
5093 : */
5094 822 : for (i = tail; i < head; i++)
5095 : {
5096 207 : if (KnownAssignedXidsValid[i])
5097 207 : break;
5098 : }
5099 822 : if (i >= head)
5100 : {
5101 : /* Array is empty, so we can reset both pointers */
5102 615 : pArray->headKnownAssignedXids = 0;
5103 615 : pArray->tailKnownAssignedXids = 0;
5104 : }
5105 : else
5106 : {
5107 207 : pArray->tailKnownAssignedXids = i;
5108 : }
5109 :
5110 : /* Opportunistically compress the array */
5111 822 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5112 : }
5113 :
5114 : /*
5115 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5116 : * We filter out anything >= xmax.
5117 : *
5118 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5119 : * that array is large enough.
5120 : *
5121 : * Caller must hold ProcArrayLock in (at least) shared mode.
5122 : */
5123 : static int
5124 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5125 : {
5126 0 : TransactionId xtmp = InvalidTransactionId;
5127 :
5128 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5129 : }
5130 :
5131 : /*
5132 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5133 : * we reduce *xmin to the lowest xid value seen if not already lower.
5134 : *
5135 : * Caller must hold ProcArrayLock in (at least) shared mode.
5136 : */
5137 : static int
5138 1280 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5139 : TransactionId xmax)
5140 : {
5141 1280 : int count = 0;
5142 : int head,
5143 : tail;
5144 : int i;
5145 :
5146 : /*
5147 : * Fetch head just once, since it may change while we loop. We can stop
5148 : * once we reach the initially seen head, since we are certain that an xid
5149 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5150 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5151 : * anyway.
5152 : */
5153 1280 : tail = procArray->tailKnownAssignedXids;
5154 1280 : head = procArray->headKnownAssignedXids;
5155 :
5156 1280 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5157 :
5158 1318 : for (i = tail; i < head; i++)
5159 : {
5160 : /* Skip any gaps in the array */
5161 171 : if (KnownAssignedXidsValid[i])
5162 : {
5163 147 : TransactionId knownXid = KnownAssignedXids[i];
5164 :
5165 : /*
5166 : * Update xmin if required. Only the first XID need be checked,
5167 : * since the array is sorted.
5168 : */
5169 293 : if (count == 0 &&
5170 146 : TransactionIdPrecedes(knownXid, *xmin))
5171 14 : *xmin = knownXid;
5172 :
5173 : /*
5174 : * Filter out anything >= xmax, again relying on sorted property
5175 : * of array.
5176 : */
5177 294 : if (TransactionIdIsValid(xmax) &&
5178 147 : TransactionIdFollowsOrEquals(knownXid, xmax))
5179 133 : break;
5180 :
5181 : /* Add knownXid into output array */
5182 14 : xarray[count++] = knownXid;
5183 : }
5184 : }
5185 :
5186 1280 : return count;
5187 : }
5188 :
5189 : /*
5190 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5191 : * if nothing there.
5192 : */
5193 : static TransactionId
5194 351 : KnownAssignedXidsGetOldestXmin(void)
5195 : {
5196 : int head,
5197 : tail;
5198 : int i;
5199 :
5200 : /*
5201 : * Fetch head just once, since it may change while we loop.
5202 : */
5203 351 : tail = procArray->tailKnownAssignedXids;
5204 351 : head = procArray->headKnownAssignedXids;
5205 :
5206 351 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5207 :
5208 351 : for (i = tail; i < head; i++)
5209 : {
5210 : /* Skip any gaps in the array */
5211 146 : if (KnownAssignedXidsValid[i])
5212 146 : return KnownAssignedXids[i];
5213 : }
5214 :
5215 205 : return InvalidTransactionId;
5216 : }
5217 :
5218 : /*
5219 : * Display KnownAssignedXids to provide debug trail
5220 : *
5221 : * Currently this is only called within startup process, so we need no
5222 : * special locking.
5223 : *
5224 : * Note this is pretty expensive, and much of the expense will be incurred
5225 : * even if the elog message will get discarded. It's not currently called
5226 : * in any performance-critical places, however, so no need to be tenser.
5227 : */
5228 : static void
5229 118 : KnownAssignedXidsDisplay(int trace_level)
5230 : {
5231 118 : ProcArrayStruct *pArray = procArray;
5232 : StringInfoData buf;
5233 : int head,
5234 : tail,
5235 : i;
5236 118 : int nxids = 0;
5237 :
5238 118 : tail = pArray->tailKnownAssignedXids;
5239 118 : head = pArray->headKnownAssignedXids;
5240 :
5241 118 : initStringInfo(&buf);
5242 :
5243 126 : for (i = tail; i < head; i++)
5244 : {
5245 8 : if (KnownAssignedXidsValid[i])
5246 : {
5247 8 : nxids++;
5248 8 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5249 : }
5250 : }
5251 :
5252 118 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5253 : nxids,
5254 : pArray->numKnownAssignedXids,
5255 : pArray->tailKnownAssignedXids,
5256 : pArray->headKnownAssignedXids,
5257 : buf.data);
5258 :
5259 118 : pfree(buf.data);
5260 118 : }
5261 :
5262 : /*
5263 : * KnownAssignedXidsReset
5264 : * Resets KnownAssignedXids to be empty
5265 : */
5266 : static void
5267 0 : KnownAssignedXidsReset(void)
5268 : {
5269 0 : ProcArrayStruct *pArray = procArray;
5270 :
5271 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5272 :
5273 0 : pArray->numKnownAssignedXids = 0;
5274 0 : pArray->tailKnownAssignedXids = 0;
5275 0 : pArray->headKnownAssignedXids = 0;
5276 :
5277 0 : LWLockRelease(ProcArrayLock);
5278 0 : }
|
