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 "utils/acl.h"
64 : #include "utils/builtins.h"
65 : #include "utils/injection_point.h"
66 : #include "utils/lsyscache.h"
67 : #include "utils/rel.h"
68 : #include "utils/snapmgr.h"
69 :
70 : #define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var))))
71 :
72 : /* Our shared memory area */
73 : typedef struct ProcArrayStruct
74 : {
75 : int numProcs; /* number of valid procs entries */
76 : int maxProcs; /* allocated size of procs array */
77 :
78 : /*
79 : * Known assigned XIDs handling
80 : */
81 : int maxKnownAssignedXids; /* allocated size of array */
82 : int numKnownAssignedXids; /* current # of valid entries */
83 : int tailKnownAssignedXids; /* index of oldest valid element */
84 : int headKnownAssignedXids; /* index of newest element, + 1 */
85 :
86 : /*
87 : * Highest subxid that has been removed from KnownAssignedXids array to
88 : * prevent overflow; or InvalidTransactionId if none. We track this for
89 : * similar reasons to tracking overflowing cached subxids in PGPROC
90 : * entries. Must hold exclusive ProcArrayLock to change this, and shared
91 : * lock to read it.
92 : */
93 : TransactionId lastOverflowedXid;
94 :
95 : /* oldest xmin of any replication slot */
96 : TransactionId replication_slot_xmin;
97 : /* oldest catalog xmin of any replication slot */
98 : TransactionId replication_slot_catalog_xmin;
99 :
100 : /* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */
101 : int pgprocnos[FLEXIBLE_ARRAY_MEMBER];
102 : } ProcArrayStruct;
103 :
104 : /*
105 : * State for the GlobalVisTest* family of functions. Those functions can
106 : * e.g. be used to decide if a deleted row can be removed without violating
107 : * MVCC semantics: If the deleted row's xmax is not considered to be running
108 : * by anyone, the row can be removed.
109 : *
110 : * To avoid slowing down GetSnapshotData(), we don't calculate a precise
111 : * cutoff XID while building a snapshot (looking at the frequently changing
112 : * xmins scales badly). Instead we compute two boundaries while building the
113 : * snapshot:
114 : *
115 : * 1) definitely_needed, indicating that rows deleted by XIDs >=
116 : * definitely_needed are definitely still visible.
117 : *
118 : * 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can
119 : * definitely be removed
120 : *
121 : * When testing an XID that falls in between the two (i.e. XID >= maybe_needed
122 : * && XID < definitely_needed), the boundaries can be recomputed (using
123 : * ComputeXidHorizons()) to get a more accurate answer. This is cheaper than
124 : * maintaining an accurate value all the time.
125 : *
126 : * As it is not cheap to compute accurate boundaries, we limit the number of
127 : * times that happens in short succession. See GlobalVisTestShouldUpdate().
128 : *
129 : *
130 : * There are three backend lifetime instances of this struct, optimized for
131 : * different types of relations. As e.g. a normal user defined table in one
132 : * database is inaccessible to backends connected to another database, a test
133 : * specific to a relation can be more aggressive than a test for a shared
134 : * relation. Currently we track four different states:
135 : *
136 : * 1) GlobalVisSharedRels, which only considers an XID's
137 : * effects visible-to-everyone if neither snapshots in any database, nor a
138 : * replication slot's xmin, nor a replication slot's catalog_xmin might
139 : * still consider XID as running.
140 : *
141 : * 2) GlobalVisCatalogRels, which only considers an XID's
142 : * effects visible-to-everyone if neither snapshots in the current
143 : * database, nor a replication slot's xmin, nor a replication slot's
144 : * catalog_xmin might still consider XID as running.
145 : *
146 : * I.e. the difference to GlobalVisSharedRels is that
147 : * snapshot in other databases are ignored.
148 : *
149 : * 3) GlobalVisDataRels, which only considers an XID's
150 : * effects visible-to-everyone if neither snapshots in the current
151 : * database, nor a replication slot's xmin consider XID as running.
152 : *
153 : * I.e. the difference to GlobalVisCatalogRels is that
154 : * replication slot's catalog_xmin is not taken into account.
155 : *
156 : * 4) GlobalVisTempRels, which only considers the current session, as temp
157 : * tables are not visible to other sessions.
158 : *
159 : * GlobalVisTestFor(relation) returns the appropriate state
160 : * for the relation.
161 : *
162 : * The boundaries are FullTransactionIds instead of TransactionIds to avoid
163 : * wraparound dangers. There e.g. would otherwise exist no procarray state to
164 : * prevent maybe_needed to become old enough after the GetSnapshotData()
165 : * call.
166 : *
167 : * The typedef is in the header.
168 : */
169 : struct GlobalVisState
170 : {
171 : /* XIDs >= are considered running by some backend */
172 : FullTransactionId definitely_needed;
173 :
174 : /* XIDs < are not considered to be running by any backend */
175 : FullTransactionId maybe_needed;
176 : };
177 :
178 : /*
179 : * Result of ComputeXidHorizons().
180 : */
181 : typedef struct ComputeXidHorizonsResult
182 : {
183 : /*
184 : * The value of TransamVariables->latestCompletedXid when
185 : * ComputeXidHorizons() held ProcArrayLock.
186 : */
187 : FullTransactionId latest_completed;
188 :
189 : /*
190 : * The same for procArray->replication_slot_xmin and
191 : * procArray->replication_slot_catalog_xmin.
192 : */
193 : TransactionId slot_xmin;
194 : TransactionId slot_catalog_xmin;
195 :
196 : /*
197 : * Oldest xid that any backend might still consider running. This needs to
198 : * include processes running VACUUM, in contrast to the normal visibility
199 : * cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when
200 : * determining visibility, but doesn't care about rows above its xmin to
201 : * be removed.
202 : *
203 : * This likely should only be needed to determine whether pg_subtrans can
204 : * be truncated. It currently includes the effects of replication slots,
205 : * for historical reasons. But that could likely be changed.
206 : */
207 : TransactionId oldest_considered_running;
208 :
209 : /*
210 : * Oldest xid for which deleted tuples need to be retained in shared
211 : * tables.
212 : *
213 : * This includes the effects of replication slots. If that's not desired,
214 : * look at shared_oldest_nonremovable_raw;
215 : */
216 : TransactionId shared_oldest_nonremovable;
217 :
218 : /*
219 : * Oldest xid that may be necessary to retain in shared tables. This is
220 : * the same as shared_oldest_nonremovable, except that is not affected by
221 : * replication slot's catalog_xmin.
222 : *
223 : * This is mainly useful to be able to send the catalog_xmin to upstream
224 : * streaming replication servers via hot_standby_feedback, so they can
225 : * apply the limit only when accessing catalog tables.
226 : */
227 : TransactionId shared_oldest_nonremovable_raw;
228 :
229 : /*
230 : * Oldest xid for which deleted tuples need to be retained in non-shared
231 : * catalog tables.
232 : */
233 : TransactionId catalog_oldest_nonremovable;
234 :
235 : /*
236 : * Oldest xid for which deleted tuples need to be retained in normal user
237 : * defined tables.
238 : */
239 : TransactionId data_oldest_nonremovable;
240 :
241 : /*
242 : * Oldest xid for which deleted tuples need to be retained in this
243 : * session's temporary tables.
244 : */
245 : TransactionId temp_oldest_nonremovable;
246 : } ComputeXidHorizonsResult;
247 :
248 : /*
249 : * Return value for GlobalVisHorizonKindForRel().
250 : */
251 : typedef enum GlobalVisHorizonKind
252 : {
253 : VISHORIZON_SHARED,
254 : VISHORIZON_CATALOG,
255 : VISHORIZON_DATA,
256 : VISHORIZON_TEMP,
257 : } GlobalVisHorizonKind;
258 :
259 : /*
260 : * Reason codes for KnownAssignedXidsCompress().
261 : */
262 : typedef enum KAXCompressReason
263 : {
264 : KAX_NO_SPACE, /* need to free up space at array end */
265 : KAX_PRUNE, /* we just pruned old entries */
266 : KAX_TRANSACTION_END, /* we just committed/removed some XIDs */
267 : KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */
268 : } KAXCompressReason;
269 :
270 :
271 : static ProcArrayStruct *procArray;
272 :
273 : static PGPROC *allProcs;
274 :
275 : /*
276 : * Cache to reduce overhead of repeated calls to TransactionIdIsInProgress()
277 : */
278 : static TransactionId cachedXidIsNotInProgress = InvalidTransactionId;
279 :
280 : /*
281 : * Bookkeeping for tracking emulated transactions in recovery
282 : */
283 : static TransactionId *KnownAssignedXids;
284 : static bool *KnownAssignedXidsValid;
285 : static TransactionId latestObservedXid = InvalidTransactionId;
286 :
287 : /*
288 : * If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is
289 : * the highest xid that might still be running that we don't have in
290 : * KnownAssignedXids.
291 : */
292 : static TransactionId standbySnapshotPendingXmin;
293 :
294 : /*
295 : * State for visibility checks on different types of relations. See struct
296 : * GlobalVisState for details. As shared, catalog, normal and temporary
297 : * relations can have different horizons, one such state exists for each.
298 : */
299 : static GlobalVisState GlobalVisSharedRels;
300 : static GlobalVisState GlobalVisCatalogRels;
301 : static GlobalVisState GlobalVisDataRels;
302 : static GlobalVisState GlobalVisTempRels;
303 :
304 : /*
305 : * This backend's RecentXmin at the last time the accurate xmin horizon was
306 : * recomputed, or InvalidTransactionId if it has not. Used to limit how many
307 : * times accurate horizons are recomputed. See GlobalVisTestShouldUpdate().
308 : */
309 : static TransactionId ComputeXidHorizonsResultLastXmin;
310 :
311 : #ifdef XIDCACHE_DEBUG
312 :
313 : /* counters for XidCache measurement */
314 : static long xc_by_recent_xmin = 0;
315 : static long xc_by_known_xact = 0;
316 : static long xc_by_my_xact = 0;
317 : static long xc_by_latest_xid = 0;
318 : static long xc_by_main_xid = 0;
319 : static long xc_by_child_xid = 0;
320 : static long xc_by_known_assigned = 0;
321 : static long xc_no_overflow = 0;
322 : static long xc_slow_answer = 0;
323 :
324 : #define xc_by_recent_xmin_inc() (xc_by_recent_xmin++)
325 : #define xc_by_known_xact_inc() (xc_by_known_xact++)
326 : #define xc_by_my_xact_inc() (xc_by_my_xact++)
327 : #define xc_by_latest_xid_inc() (xc_by_latest_xid++)
328 : #define xc_by_main_xid_inc() (xc_by_main_xid++)
329 : #define xc_by_child_xid_inc() (xc_by_child_xid++)
330 : #define xc_by_known_assigned_inc() (xc_by_known_assigned++)
331 : #define xc_no_overflow_inc() (xc_no_overflow++)
332 : #define xc_slow_answer_inc() (xc_slow_answer++)
333 :
334 : static void DisplayXidCache(void);
335 : #else /* !XIDCACHE_DEBUG */
336 :
337 : #define xc_by_recent_xmin_inc() ((void) 0)
338 : #define xc_by_known_xact_inc() ((void) 0)
339 : #define xc_by_my_xact_inc() ((void) 0)
340 : #define xc_by_latest_xid_inc() ((void) 0)
341 : #define xc_by_main_xid_inc() ((void) 0)
342 : #define xc_by_child_xid_inc() ((void) 0)
343 : #define xc_by_known_assigned_inc() ((void) 0)
344 : #define xc_no_overflow_inc() ((void) 0)
345 : #define xc_slow_answer_inc() ((void) 0)
346 : #endif /* XIDCACHE_DEBUG */
347 :
348 : /* Primitives for KnownAssignedXids array handling for standby */
349 : static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock);
350 : static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
351 : bool exclusive_lock);
352 : static bool KnownAssignedXidsSearch(TransactionId xid, bool remove);
353 : static bool KnownAssignedXidExists(TransactionId xid);
354 : static void KnownAssignedXidsRemove(TransactionId xid);
355 : static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
356 : TransactionId *subxids);
357 : static void KnownAssignedXidsRemovePreceding(TransactionId removeXid);
358 : static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax);
359 : static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray,
360 : TransactionId *xmin,
361 : TransactionId xmax);
362 : static TransactionId KnownAssignedXidsGetOldestXmin(void);
363 : static void KnownAssignedXidsDisplay(int trace_level);
364 : static void KnownAssignedXidsReset(void);
365 : static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid);
366 : static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid);
367 : static void MaintainLatestCompletedXid(TransactionId latestXid);
368 : static void MaintainLatestCompletedXidRecovery(TransactionId latestXid);
369 :
370 : static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel,
371 : TransactionId xid);
372 : static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons);
373 :
374 : /*
375 : * Report shared-memory space needed by ProcArrayShmemInit
376 : */
377 : Size
378 4238 : ProcArrayShmemSize(void)
379 : {
380 : Size size;
381 :
382 : /* Size of the ProcArray structure itself */
383 : #define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts)
384 :
385 4238 : size = offsetof(ProcArrayStruct, pgprocnos);
386 4238 : size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS));
387 :
388 : /*
389 : * During Hot Standby processing we have a data structure called
390 : * KnownAssignedXids, created in shared memory. Local data structures are
391 : * also created in various backends during GetSnapshotData(),
392 : * TransactionIdIsInProgress() and GetRunningTransactionData(). All of the
393 : * main structures created in those functions must be identically sized,
394 : * since we may at times copy the whole of the data structures around. We
395 : * refer to this size as TOTAL_MAX_CACHED_SUBXIDS.
396 : *
397 : * Ideally we'd only create this structure if we were actually doing hot
398 : * standby in the current run, but we don't know that yet at the time
399 : * shared memory is being set up.
400 : */
401 : #define TOTAL_MAX_CACHED_SUBXIDS \
402 : ((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS)
403 :
404 4238 : if (EnableHotStandby)
405 : {
406 4214 : size = add_size(size,
407 : mul_size(sizeof(TransactionId),
408 4214 : TOTAL_MAX_CACHED_SUBXIDS));
409 4214 : size = add_size(size,
410 4214 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS));
411 : }
412 :
413 4238 : return size;
414 : }
415 :
416 : /*
417 : * Initialize the shared PGPROC array during postmaster startup.
418 : */
419 : void
420 2272 : ProcArrayShmemInit(void)
421 : {
422 : bool found;
423 :
424 : /* Create or attach to the ProcArray shared structure */
425 2272 : procArray = (ProcArrayStruct *)
426 2272 : ShmemInitStruct("Proc Array",
427 : add_size(offsetof(ProcArrayStruct, pgprocnos),
428 : mul_size(sizeof(int),
429 2272 : PROCARRAY_MAXPROCS)),
430 : &found);
431 :
432 2272 : if (!found)
433 : {
434 : /*
435 : * We're the first - initialize.
436 : */
437 2272 : procArray->numProcs = 0;
438 2272 : procArray->maxProcs = PROCARRAY_MAXPROCS;
439 2272 : procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS;
440 2272 : procArray->numKnownAssignedXids = 0;
441 2272 : procArray->tailKnownAssignedXids = 0;
442 2272 : procArray->headKnownAssignedXids = 0;
443 2272 : procArray->lastOverflowedXid = InvalidTransactionId;
444 2272 : procArray->replication_slot_xmin = InvalidTransactionId;
445 2272 : procArray->replication_slot_catalog_xmin = InvalidTransactionId;
446 2272 : TransamVariables->xactCompletionCount = 1;
447 : }
448 :
449 2272 : allProcs = ProcGlobal->allProcs;
450 :
451 : /* Create or attach to the KnownAssignedXids arrays too, if needed */
452 2272 : if (EnableHotStandby)
453 : {
454 2260 : KnownAssignedXids = (TransactionId *)
455 2260 : ShmemInitStruct("KnownAssignedXids",
456 : mul_size(sizeof(TransactionId),
457 2260 : TOTAL_MAX_CACHED_SUBXIDS),
458 : &found);
459 2260 : KnownAssignedXidsValid = (bool *)
460 2260 : ShmemInitStruct("KnownAssignedXidsValid",
461 2260 : mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS),
462 : &found);
463 : }
464 2272 : }
465 :
466 : /*
467 : * Add the specified PGPROC to the shared array.
468 : */
469 : void
470 37008 : ProcArrayAdd(PGPROC *proc)
471 : {
472 37008 : int pgprocno = GetNumberFromPGProc(proc);
473 37008 : ProcArrayStruct *arrayP = procArray;
474 : int index;
475 : int movecount;
476 :
477 : /* See ProcGlobal comment explaining why both locks are held */
478 37008 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
479 37008 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
480 :
481 37008 : if (arrayP->numProcs >= arrayP->maxProcs)
482 : {
483 : /*
484 : * Oops, no room. (This really shouldn't happen, since there is a
485 : * fixed supply of PGPROC structs too, and so we should have failed
486 : * earlier.)
487 : */
488 0 : ereport(FATAL,
489 : (errcode(ERRCODE_TOO_MANY_CONNECTIONS),
490 : errmsg("sorry, too many clients already")));
491 : }
492 :
493 : /*
494 : * Keep the procs array sorted by (PGPROC *) so that we can utilize
495 : * locality of references much better. This is useful while traversing the
496 : * ProcArray because there is an increased likelihood of finding the next
497 : * PGPROC structure in the cache.
498 : *
499 : * Since the occurrence of adding/removing a proc is much lower than the
500 : * access to the ProcArray itself, the overhead should be marginal
501 : */
502 88430 : for (index = 0; index < arrayP->numProcs; index++)
503 : {
504 79388 : int this_procno = arrayP->pgprocnos[index];
505 :
506 : Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
507 : Assert(allProcs[this_procno].pgxactoff == index);
508 :
509 : /* If we have found our right position in the array, break */
510 79388 : if (this_procno > pgprocno)
511 27966 : break;
512 : }
513 :
514 37008 : movecount = arrayP->numProcs - index;
515 37008 : memmove(&arrayP->pgprocnos[index + 1],
516 37008 : &arrayP->pgprocnos[index],
517 : movecount * sizeof(*arrayP->pgprocnos));
518 37008 : memmove(&ProcGlobal->xids[index + 1],
519 37008 : &ProcGlobal->xids[index],
520 : movecount * sizeof(*ProcGlobal->xids));
521 37008 : memmove(&ProcGlobal->subxidStates[index + 1],
522 37008 : &ProcGlobal->subxidStates[index],
523 : movecount * sizeof(*ProcGlobal->subxidStates));
524 37008 : memmove(&ProcGlobal->statusFlags[index + 1],
525 37008 : &ProcGlobal->statusFlags[index],
526 : movecount * sizeof(*ProcGlobal->statusFlags));
527 :
528 37008 : arrayP->pgprocnos[index] = GetNumberFromPGProc(proc);
529 37008 : proc->pgxactoff = index;
530 37008 : ProcGlobal->xids[index] = proc->xid;
531 37008 : ProcGlobal->subxidStates[index] = proc->subxidStatus;
532 37008 : ProcGlobal->statusFlags[index] = proc->statusFlags;
533 :
534 37008 : arrayP->numProcs++;
535 :
536 : /* adjust pgxactoff for all following PGPROCs */
537 37008 : index++;
538 102032 : for (; index < arrayP->numProcs; index++)
539 : {
540 65024 : int procno = arrayP->pgprocnos[index];
541 :
542 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
543 : Assert(allProcs[procno].pgxactoff == index - 1);
544 :
545 65024 : allProcs[procno].pgxactoff = index;
546 : }
547 :
548 : /*
549 : * Release in reversed acquisition order, to reduce frequency of having to
550 : * wait for XidGenLock while holding ProcArrayLock.
551 : */
552 37008 : LWLockRelease(XidGenLock);
553 37008 : LWLockRelease(ProcArrayLock);
554 37008 : }
555 :
556 : /*
557 : * Remove the specified PGPROC from the shared array.
558 : *
559 : * When latestXid is a valid XID, we are removing a live 2PC gxact from the
560 : * array, and thus causing it to appear as "not running" anymore. In this
561 : * case we must advance latestCompletedXid. (This is essentially the same
562 : * as ProcArrayEndTransaction followed by removal of the PGPROC, but we take
563 : * the ProcArrayLock only once, and don't damage the content of the PGPROC;
564 : * twophase.c depends on the latter.)
565 : */
566 : void
567 36954 : ProcArrayRemove(PGPROC *proc, TransactionId latestXid)
568 : {
569 36954 : ProcArrayStruct *arrayP = procArray;
570 : int myoff;
571 : int movecount;
572 :
573 : #ifdef XIDCACHE_DEBUG
574 : /* dump stats at backend shutdown, but not prepared-xact end */
575 : if (proc->pid != 0)
576 : DisplayXidCache();
577 : #endif
578 :
579 : /* See ProcGlobal comment explaining why both locks are held */
580 36954 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
581 36954 : LWLockAcquire(XidGenLock, LW_EXCLUSIVE);
582 :
583 36954 : myoff = proc->pgxactoff;
584 :
585 : Assert(myoff >= 0 && myoff < arrayP->numProcs);
586 : Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff);
587 :
588 36954 : if (TransactionIdIsValid(latestXid))
589 : {
590 : Assert(TransactionIdIsValid(ProcGlobal->xids[myoff]));
591 :
592 : /* Advance global latestCompletedXid while holding the lock */
593 620 : MaintainLatestCompletedXid(latestXid);
594 :
595 : /* Same with xactCompletionCount */
596 620 : TransamVariables->xactCompletionCount++;
597 :
598 620 : ProcGlobal->xids[myoff] = InvalidTransactionId;
599 620 : ProcGlobal->subxidStates[myoff].overflowed = false;
600 620 : ProcGlobal->subxidStates[myoff].count = 0;
601 : }
602 : else
603 : {
604 : /* Shouldn't be trying to remove a live transaction here */
605 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
606 : }
607 :
608 : Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff]));
609 : Assert(ProcGlobal->subxidStates[myoff].count == 0);
610 : Assert(ProcGlobal->subxidStates[myoff].overflowed == false);
611 :
612 36954 : ProcGlobal->statusFlags[myoff] = 0;
613 :
614 : /* Keep the PGPROC array sorted. See notes above */
615 36954 : movecount = arrayP->numProcs - myoff - 1;
616 36954 : memmove(&arrayP->pgprocnos[myoff],
617 36954 : &arrayP->pgprocnos[myoff + 1],
618 : movecount * sizeof(*arrayP->pgprocnos));
619 36954 : memmove(&ProcGlobal->xids[myoff],
620 36954 : &ProcGlobal->xids[myoff + 1],
621 : movecount * sizeof(*ProcGlobal->xids));
622 36954 : memmove(&ProcGlobal->subxidStates[myoff],
623 36954 : &ProcGlobal->subxidStates[myoff + 1],
624 : movecount * sizeof(*ProcGlobal->subxidStates));
625 36954 : memmove(&ProcGlobal->statusFlags[myoff],
626 36954 : &ProcGlobal->statusFlags[myoff + 1],
627 : movecount * sizeof(*ProcGlobal->statusFlags));
628 :
629 36954 : arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */
630 36954 : arrayP->numProcs--;
631 :
632 : /*
633 : * Adjust pgxactoff of following procs for removed PGPROC (note that
634 : * numProcs already has been decremented).
635 : */
636 108510 : for (int index = myoff; index < arrayP->numProcs; index++)
637 : {
638 71556 : int procno = arrayP->pgprocnos[index];
639 :
640 : Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS));
641 : Assert(allProcs[procno].pgxactoff - 1 == index);
642 :
643 71556 : allProcs[procno].pgxactoff = index;
644 : }
645 :
646 : /*
647 : * Release in reversed acquisition order, to reduce frequency of having to
648 : * wait for XidGenLock while holding ProcArrayLock.
649 : */
650 36954 : LWLockRelease(XidGenLock);
651 36954 : LWLockRelease(ProcArrayLock);
652 36954 : }
653 :
654 :
655 : /*
656 : * ProcArrayEndTransaction -- mark a transaction as no longer running
657 : *
658 : * This is used interchangeably for commit and abort cases. The transaction
659 : * commit/abort must already be reported to WAL and pg_xact.
660 : *
661 : * proc is currently always MyProc, but we pass it explicitly for flexibility.
662 : * latestXid is the latest Xid among the transaction's main XID and
663 : * subtransactions, or InvalidTransactionId if it has no XID. (We must ask
664 : * the caller to pass latestXid, instead of computing it from the PGPROC's
665 : * contents, because the subxid information in the PGPROC might be
666 : * incomplete.)
667 : */
668 : void
669 964426 : ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid)
670 : {
671 964426 : if (TransactionIdIsValid(latestXid))
672 : {
673 : /*
674 : * We must lock ProcArrayLock while clearing our advertised XID, so
675 : * that we do not exit the set of "running" transactions while someone
676 : * else is taking a snapshot. See discussion in
677 : * src/backend/access/transam/README.
678 : */
679 : Assert(TransactionIdIsValid(proc->xid));
680 :
681 : /*
682 : * If we can immediately acquire ProcArrayLock, we clear our own XID
683 : * and release the lock. If not, use group XID clearing to improve
684 : * efficiency.
685 : */
686 274740 : if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE))
687 : {
688 274376 : ProcArrayEndTransactionInternal(proc, latestXid);
689 274376 : LWLockRelease(ProcArrayLock);
690 : }
691 : else
692 364 : ProcArrayGroupClearXid(proc, latestXid);
693 : }
694 : else
695 : {
696 : /*
697 : * If we have no XID, we don't need to lock, since we won't affect
698 : * anyone else's calculation of a snapshot. We might change their
699 : * estimate of global xmin, but that's OK.
700 : */
701 : Assert(!TransactionIdIsValid(proc->xid));
702 : Assert(proc->subxidStatus.count == 0);
703 : Assert(!proc->subxidStatus.overflowed);
704 :
705 689686 : proc->vxid.lxid = InvalidLocalTransactionId;
706 689686 : proc->xmin = InvalidTransactionId;
707 :
708 : /* be sure this is cleared in abort */
709 689686 : proc->delayChkptFlags = 0;
710 :
711 689686 : proc->recoveryConflictPending = false;
712 :
713 : /* must be cleared with xid/xmin: */
714 : /* avoid unnecessarily dirtying shared cachelines */
715 689686 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
716 : {
717 : Assert(!LWLockHeldByMe(ProcArrayLock));
718 167464 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
719 : Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]);
720 167464 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
721 167464 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
722 167464 : LWLockRelease(ProcArrayLock);
723 : }
724 : }
725 964426 : }
726 :
727 : /*
728 : * Mark a write transaction as no longer running.
729 : *
730 : * We don't do any locking here; caller must handle that.
731 : */
732 : static inline void
733 274740 : ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid)
734 : {
735 274740 : int pgxactoff = proc->pgxactoff;
736 :
737 : /*
738 : * Note: we need exclusive lock here because we're going to change other
739 : * processes' PGPROC entries.
740 : */
741 : Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE));
742 : Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff]));
743 : Assert(ProcGlobal->xids[pgxactoff] == proc->xid);
744 :
745 274740 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
746 274740 : proc->xid = InvalidTransactionId;
747 274740 : proc->vxid.lxid = InvalidLocalTransactionId;
748 274740 : proc->xmin = InvalidTransactionId;
749 :
750 : /* be sure this is cleared in abort */
751 274740 : proc->delayChkptFlags = 0;
752 :
753 274740 : proc->recoveryConflictPending = false;
754 :
755 : /* must be cleared with xid/xmin: */
756 : /* avoid unnecessarily dirtying shared cachelines */
757 274740 : if (proc->statusFlags & PROC_VACUUM_STATE_MASK)
758 : {
759 1516 : proc->statusFlags &= ~PROC_VACUUM_STATE_MASK;
760 1516 : ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags;
761 : }
762 :
763 : /* Clear the subtransaction-XID cache too while holding the lock */
764 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
765 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
766 274740 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
767 : {
768 1162 : ProcGlobal->subxidStates[pgxactoff].count = 0;
769 1162 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
770 1162 : proc->subxidStatus.count = 0;
771 1162 : proc->subxidStatus.overflowed = false;
772 : }
773 :
774 : /* Also advance global latestCompletedXid while holding the lock */
775 274740 : MaintainLatestCompletedXid(latestXid);
776 :
777 : /* Same with xactCompletionCount */
778 274740 : TransamVariables->xactCompletionCount++;
779 274740 : }
780 :
781 : /*
782 : * ProcArrayGroupClearXid -- group XID clearing
783 : *
784 : * When we cannot immediately acquire ProcArrayLock in exclusive mode at
785 : * commit time, add ourselves to a list of processes that need their XIDs
786 : * cleared. The first process to add itself to the list will acquire
787 : * ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal
788 : * on behalf of all group members. This avoids a great deal of contention
789 : * around ProcArrayLock when many processes are trying to commit at once,
790 : * since the lock need not be repeatedly handed off from one committing
791 : * process to the next.
792 : */
793 : static void
794 364 : ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid)
795 : {
796 364 : int pgprocno = GetNumberFromPGProc(proc);
797 364 : PROC_HDR *procglobal = ProcGlobal;
798 : uint32 nextidx;
799 : uint32 wakeidx;
800 :
801 : /* We should definitely have an XID to clear. */
802 : Assert(TransactionIdIsValid(proc->xid));
803 :
804 : /* Add ourselves to the list of processes needing a group XID clear. */
805 364 : proc->procArrayGroupMember = true;
806 364 : proc->procArrayGroupMemberXid = latestXid;
807 364 : nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst);
808 : while (true)
809 : {
810 364 : pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx);
811 :
812 364 : if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst,
813 : &nextidx,
814 : (uint32) pgprocno))
815 364 : break;
816 : }
817 :
818 : /*
819 : * If the list was not empty, the leader will clear our XID. It is
820 : * impossible to have followers without a leader because the first process
821 : * that has added itself to the list will always have nextidx as
822 : * INVALID_PROC_NUMBER.
823 : */
824 364 : if (nextidx != INVALID_PROC_NUMBER)
825 : {
826 36 : int extraWaits = 0;
827 :
828 : /* Sleep until the leader clears our XID. */
829 36 : pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE);
830 : for (;;)
831 : {
832 : /* acts as a read barrier */
833 36 : PGSemaphoreLock(proc->sem);
834 36 : if (!proc->procArrayGroupMember)
835 36 : break;
836 0 : extraWaits++;
837 : }
838 36 : pgstat_report_wait_end();
839 :
840 : Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER);
841 :
842 : /* Fix semaphore count for any absorbed wakeups */
843 36 : while (extraWaits-- > 0)
844 0 : PGSemaphoreUnlock(proc->sem);
845 36 : return;
846 : }
847 :
848 : /* We are the leader. Acquire the lock on behalf of everyone. */
849 328 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
850 :
851 : /*
852 : * Now that we've got the lock, clear the list of processes waiting for
853 : * group XID clearing, saving a pointer to the head of the list. Trying
854 : * to pop elements one at a time could lead to an ABA problem.
855 : */
856 328 : nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst,
857 : INVALID_PROC_NUMBER);
858 :
859 : /* Remember head of list so we can perform wakeups after dropping lock. */
860 328 : wakeidx = nextidx;
861 :
862 : /* Walk the list and clear all XIDs. */
863 692 : while (nextidx != INVALID_PROC_NUMBER)
864 : {
865 364 : PGPROC *nextproc = &allProcs[nextidx];
866 :
867 364 : ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid);
868 :
869 : /* Move to next proc in list. */
870 364 : nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
871 : }
872 :
873 : /* We're done with the lock now. */
874 328 : LWLockRelease(ProcArrayLock);
875 :
876 : /*
877 : * Now that we've released the lock, go back and wake everybody up. We
878 : * don't do this under the lock so as to keep lock hold times to a
879 : * minimum. The system calls we need to perform to wake other processes
880 : * up are probably much slower than the simple memory writes we did while
881 : * holding the lock.
882 : */
883 692 : while (wakeidx != INVALID_PROC_NUMBER)
884 : {
885 364 : PGPROC *nextproc = &allProcs[wakeidx];
886 :
887 364 : wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext);
888 364 : pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER);
889 :
890 : /* ensure all previous writes are visible before follower continues. */
891 364 : pg_write_barrier();
892 :
893 364 : nextproc->procArrayGroupMember = false;
894 :
895 364 : if (nextproc != MyProc)
896 36 : PGSemaphoreUnlock(nextproc->sem);
897 : }
898 : }
899 :
900 : /*
901 : * ProcArrayClearTransaction -- clear the transaction fields
902 : *
903 : * This is used after successfully preparing a 2-phase transaction. We are
904 : * not actually reporting the transaction's XID as no longer running --- it
905 : * will still appear as running because the 2PC's gxact is in the ProcArray
906 : * too. We just have to clear out our own PGPROC.
907 : */
908 : void
909 608 : ProcArrayClearTransaction(PGPROC *proc)
910 : {
911 : int pgxactoff;
912 :
913 : /*
914 : * Currently we need to lock ProcArrayLock exclusively here, as we
915 : * increment xactCompletionCount below. We also need it at least in shared
916 : * mode for pgproc->pgxactoff to stay the same below.
917 : *
918 : * We could however, as this action does not actually change anyone's view
919 : * of the set of running XIDs (our entry is duplicate with the gxact that
920 : * has already been inserted into the ProcArray), lower the lock level to
921 : * shared if we were to make xactCompletionCount an atomic variable. But
922 : * that doesn't seem worth it currently, as a 2PC commit is heavyweight
923 : * enough for this not to be the bottleneck. If it ever becomes a
924 : * bottleneck it may also be worth considering to combine this with the
925 : * subsequent ProcArrayRemove()
926 : */
927 608 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
928 :
929 608 : pgxactoff = proc->pgxactoff;
930 :
931 608 : ProcGlobal->xids[pgxactoff] = InvalidTransactionId;
932 608 : proc->xid = InvalidTransactionId;
933 :
934 608 : proc->vxid.lxid = InvalidLocalTransactionId;
935 608 : proc->xmin = InvalidTransactionId;
936 608 : proc->recoveryConflictPending = false;
937 :
938 : Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK));
939 : Assert(!proc->delayChkptFlags);
940 :
941 : /*
942 : * Need to increment completion count even though transaction hasn't
943 : * really committed yet. The reason for that is that GetSnapshotData()
944 : * omits the xid of the current transaction, thus without the increment we
945 : * otherwise could end up reusing the snapshot later. Which would be bad,
946 : * because it might not count the prepared transaction as running.
947 : */
948 608 : TransamVariables->xactCompletionCount++;
949 :
950 : /* Clear the subtransaction-XID cache too */
951 : Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count &&
952 : ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed);
953 608 : if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed)
954 : {
955 212 : ProcGlobal->subxidStates[pgxactoff].count = 0;
956 212 : ProcGlobal->subxidStates[pgxactoff].overflowed = false;
957 212 : proc->subxidStatus.count = 0;
958 212 : proc->subxidStatus.overflowed = false;
959 : }
960 :
961 608 : LWLockRelease(ProcArrayLock);
962 608 : }
963 :
964 : /*
965 : * Update TransamVariables->latestCompletedXid to point to latestXid if
966 : * currently older.
967 : */
968 : static void
969 276696 : MaintainLatestCompletedXid(TransactionId latestXid)
970 : {
971 276696 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
972 :
973 : Assert(FullTransactionIdIsValid(cur_latest));
974 : Assert(!RecoveryInProgress());
975 : Assert(LWLockHeldByMe(ProcArrayLock));
976 :
977 276696 : if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
978 : {
979 248278 : TransamVariables->latestCompletedXid =
980 248278 : FullXidRelativeTo(cur_latest, latestXid);
981 : }
982 :
983 : Assert(IsBootstrapProcessingMode() ||
984 : FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
985 276696 : }
986 :
987 : /*
988 : * Same as MaintainLatestCompletedXid, except for use during WAL replay.
989 : */
990 : static void
991 46336 : MaintainLatestCompletedXidRecovery(TransactionId latestXid)
992 : {
993 46336 : FullTransactionId cur_latest = TransamVariables->latestCompletedXid;
994 : FullTransactionId rel;
995 :
996 : Assert(AmStartupProcess() || !IsUnderPostmaster);
997 : Assert(LWLockHeldByMe(ProcArrayLock));
998 :
999 : /*
1000 : * Need a FullTransactionId to compare latestXid with. Can't rely on
1001 : * latestCompletedXid to be initialized in recovery. But in recovery it's
1002 : * safe to access nextXid without a lock for the startup process.
1003 : */
1004 46336 : rel = TransamVariables->nextXid;
1005 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1006 :
1007 92442 : if (!FullTransactionIdIsValid(cur_latest) ||
1008 46106 : TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid))
1009 : {
1010 34130 : TransamVariables->latestCompletedXid =
1011 34130 : FullXidRelativeTo(rel, latestXid);
1012 : }
1013 :
1014 : Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid));
1015 46336 : }
1016 :
1017 : /*
1018 : * ProcArrayInitRecovery -- initialize recovery xid mgmt environment
1019 : *
1020 : * Remember up to where the startup process initialized the CLOG and subtrans
1021 : * so we can ensure it's initialized gaplessly up to the point where necessary
1022 : * while in recovery.
1023 : */
1024 : void
1025 230 : ProcArrayInitRecovery(TransactionId initializedUptoXID)
1026 : {
1027 : Assert(standbyState == STANDBY_INITIALIZED);
1028 : Assert(TransactionIdIsNormal(initializedUptoXID));
1029 :
1030 : /*
1031 : * we set latestObservedXid to the xid SUBTRANS has been initialized up
1032 : * to, so we can extend it from that point onwards in
1033 : * RecordKnownAssignedTransactionIds, and when we get consistent in
1034 : * ProcArrayApplyRecoveryInfo().
1035 : */
1036 230 : latestObservedXid = initializedUptoXID;
1037 230 : TransactionIdRetreat(latestObservedXid);
1038 230 : }
1039 :
1040 : /*
1041 : * ProcArrayApplyRecoveryInfo -- apply recovery info about xids
1042 : *
1043 : * Takes us through 3 states: Initialized, Pending and Ready.
1044 : * Normal case is to go all the way to Ready straight away, though there
1045 : * are atypical cases where we need to take it in steps.
1046 : *
1047 : * Use the data about running transactions on the primary to create the initial
1048 : * state of KnownAssignedXids. We also use these records to regularly prune
1049 : * KnownAssignedXids because we know it is possible that some transactions
1050 : * with FATAL errors fail to write abort records, which could cause eventual
1051 : * overflow.
1052 : *
1053 : * See comments for LogStandbySnapshot().
1054 : */
1055 : void
1056 1684 : ProcArrayApplyRecoveryInfo(RunningTransactions running)
1057 : {
1058 : TransactionId *xids;
1059 : TransactionId advanceNextXid;
1060 : int nxids;
1061 : int i;
1062 :
1063 : Assert(standbyState >= STANDBY_INITIALIZED);
1064 : Assert(TransactionIdIsValid(running->nextXid));
1065 : Assert(TransactionIdIsValid(running->oldestRunningXid));
1066 : Assert(TransactionIdIsNormal(running->latestCompletedXid));
1067 :
1068 : /*
1069 : * Remove stale transactions, if any.
1070 : */
1071 1684 : ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid);
1072 :
1073 : /*
1074 : * Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(),
1075 : * because we will need it up to date for accessing two-phase transactions
1076 : * in StandbyReleaseOldLocks().
1077 : */
1078 1684 : advanceNextXid = running->nextXid;
1079 1684 : TransactionIdRetreat(advanceNextXid);
1080 1684 : AdvanceNextFullTransactionIdPastXid(advanceNextXid);
1081 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
1082 :
1083 : /*
1084 : * Remove stale locks, if any.
1085 : */
1086 1684 : StandbyReleaseOldLocks(running->oldestRunningXid);
1087 :
1088 : /*
1089 : * If our snapshot is already valid, nothing else to do...
1090 : */
1091 1684 : if (standbyState == STANDBY_SNAPSHOT_READY)
1092 1454 : return;
1093 :
1094 : /*
1095 : * If our initial RunningTransactionsData had an overflowed snapshot then
1096 : * we knew we were missing some subxids from our snapshot. If we continue
1097 : * to see overflowed snapshots then we might never be able to start up, so
1098 : * we make another test to see if our snapshot is now valid. We know that
1099 : * the missing subxids are equal to or earlier than nextXid. After we
1100 : * initialise we continue to apply changes during recovery, so once the
1101 : * oldestRunningXid is later than the nextXid from the initial snapshot we
1102 : * know that we no longer have missing information and can mark the
1103 : * snapshot as valid.
1104 : */
1105 230 : if (standbyState == STANDBY_SNAPSHOT_PENDING)
1106 : {
1107 : /*
1108 : * If the snapshot isn't overflowed or if its empty we can reset our
1109 : * pending state and use this snapshot instead.
1110 : */
1111 0 : if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0)
1112 : {
1113 : /*
1114 : * If we have already collected known assigned xids, we need to
1115 : * throw them away before we apply the recovery snapshot.
1116 : */
1117 0 : KnownAssignedXidsReset();
1118 0 : standbyState = STANDBY_INITIALIZED;
1119 : }
1120 : else
1121 : {
1122 0 : if (TransactionIdPrecedes(standbySnapshotPendingXmin,
1123 : running->oldestRunningXid))
1124 : {
1125 0 : standbyState = STANDBY_SNAPSHOT_READY;
1126 0 : elog(DEBUG1,
1127 : "recovery snapshots are now enabled");
1128 : }
1129 : else
1130 0 : elog(DEBUG1,
1131 : "recovery snapshot waiting for non-overflowed snapshot or "
1132 : "until oldest active xid on standby is at least %u (now %u)",
1133 : standbySnapshotPendingXmin,
1134 : running->oldestRunningXid);
1135 0 : return;
1136 : }
1137 : }
1138 :
1139 : Assert(standbyState == STANDBY_INITIALIZED);
1140 :
1141 : /*
1142 : * NB: this can be reached at least twice, so make sure new code can deal
1143 : * with that.
1144 : */
1145 :
1146 : /*
1147 : * Nobody else is running yet, but take locks anyhow
1148 : */
1149 230 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1150 :
1151 : /*
1152 : * KnownAssignedXids is sorted so we cannot just add the xids, we have to
1153 : * sort them first.
1154 : *
1155 : * Some of the new xids are top-level xids and some are subtransactions.
1156 : * We don't call SubTransSetParent because it doesn't matter yet. If we
1157 : * aren't overflowed then all xids will fit in snapshot and so we don't
1158 : * need subtrans. If we later overflow, an xid assignment record will add
1159 : * xids to subtrans. If RunningTransactionsData is overflowed then we
1160 : * don't have enough information to correctly update subtrans anyway.
1161 : */
1162 :
1163 : /*
1164 : * Allocate a temporary array to avoid modifying the array passed as
1165 : * argument.
1166 : */
1167 230 : xids = palloc_array(TransactionId, running->xcnt + running->subxcnt);
1168 :
1169 : /*
1170 : * Add to the temp array any xids which have not already completed.
1171 : */
1172 230 : nxids = 0;
1173 238 : for (i = 0; i < running->xcnt + running->subxcnt; i++)
1174 : {
1175 8 : TransactionId xid = running->xids[i];
1176 :
1177 : /*
1178 : * The running-xacts snapshot can contain xids that were still visible
1179 : * in the procarray when the snapshot was taken, but were already
1180 : * WAL-logged as completed. They're not running anymore, so ignore
1181 : * them.
1182 : */
1183 8 : if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid))
1184 0 : continue;
1185 :
1186 8 : xids[nxids++] = xid;
1187 : }
1188 :
1189 230 : if (nxids > 0)
1190 : {
1191 8 : if (procArray->numKnownAssignedXids != 0)
1192 : {
1193 0 : LWLockRelease(ProcArrayLock);
1194 0 : elog(ERROR, "KnownAssignedXids is not empty");
1195 : }
1196 :
1197 : /*
1198 : * Sort the array so that we can add them safely into
1199 : * KnownAssignedXids.
1200 : *
1201 : * We have to sort them logically, because in KnownAssignedXidsAdd we
1202 : * call TransactionIdFollowsOrEquals and so on. But we know these XIDs
1203 : * come from RUNNING_XACTS, which means there are only normal XIDs
1204 : * from the same epoch, so this is safe.
1205 : */
1206 8 : qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator);
1207 :
1208 : /*
1209 : * Add the sorted snapshot into KnownAssignedXids. The running-xacts
1210 : * snapshot may include duplicated xids because of prepared
1211 : * transactions, so ignore them.
1212 : */
1213 16 : for (i = 0; i < nxids; i++)
1214 : {
1215 8 : if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i]))
1216 : {
1217 0 : elog(DEBUG1,
1218 : "found duplicated transaction %u for KnownAssignedXids insertion",
1219 : xids[i]);
1220 0 : continue;
1221 : }
1222 8 : KnownAssignedXidsAdd(xids[i], xids[i], true);
1223 : }
1224 :
1225 8 : KnownAssignedXidsDisplay(DEBUG3);
1226 : }
1227 :
1228 230 : pfree(xids);
1229 :
1230 : /*
1231 : * latestObservedXid is at least set to the point where SUBTRANS was
1232 : * started up to (cf. ProcArrayInitRecovery()) or to the biggest xid
1233 : * RecordKnownAssignedTransactionIds() was called for. Initialize
1234 : * subtrans from thereon, up to nextXid - 1.
1235 : *
1236 : * We need to duplicate parts of RecordKnownAssignedTransactionId() here,
1237 : * because we've just added xids to the known assigned xids machinery that
1238 : * haven't gone through RecordKnownAssignedTransactionId().
1239 : */
1240 : Assert(TransactionIdIsNormal(latestObservedXid));
1241 230 : TransactionIdAdvance(latestObservedXid);
1242 460 : while (TransactionIdPrecedes(latestObservedXid, running->nextXid))
1243 : {
1244 0 : ExtendSUBTRANS(latestObservedXid);
1245 0 : TransactionIdAdvance(latestObservedXid);
1246 : }
1247 230 : TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */
1248 :
1249 : /* ----------
1250 : * Now we've got the running xids we need to set the global values that
1251 : * are used to track snapshots as they evolve further.
1252 : *
1253 : * - latestCompletedXid which will be the xmax for snapshots
1254 : * - lastOverflowedXid which shows whether snapshots overflow
1255 : * - nextXid
1256 : *
1257 : * If the snapshot overflowed, then we still initialise with what we know,
1258 : * but the recovery snapshot isn't fully valid yet because we know there
1259 : * are some subxids missing. We don't know the specific subxids that are
1260 : * missing, so conservatively assume the last one is latestObservedXid.
1261 : * ----------
1262 : */
1263 230 : if (running->subxid_status == SUBXIDS_MISSING)
1264 : {
1265 0 : standbyState = STANDBY_SNAPSHOT_PENDING;
1266 :
1267 0 : standbySnapshotPendingXmin = latestObservedXid;
1268 0 : procArray->lastOverflowedXid = latestObservedXid;
1269 : }
1270 : else
1271 : {
1272 230 : standbyState = STANDBY_SNAPSHOT_READY;
1273 :
1274 230 : standbySnapshotPendingXmin = InvalidTransactionId;
1275 :
1276 : /*
1277 : * If the 'xids' array didn't include all subtransactions, we have to
1278 : * mark any snapshots taken as overflowed.
1279 : */
1280 230 : if (running->subxid_status == SUBXIDS_IN_SUBTRANS)
1281 52 : procArray->lastOverflowedXid = latestObservedXid;
1282 : else
1283 : {
1284 : Assert(running->subxid_status == SUBXIDS_IN_ARRAY);
1285 178 : procArray->lastOverflowedXid = InvalidTransactionId;
1286 : }
1287 : }
1288 :
1289 : /*
1290 : * If a transaction wrote a commit record in the gap between taking and
1291 : * logging the snapshot then latestCompletedXid may already be higher than
1292 : * the value from the snapshot, so check before we use the incoming value.
1293 : * It also might not yet be set at all.
1294 : */
1295 230 : MaintainLatestCompletedXidRecovery(running->latestCompletedXid);
1296 :
1297 : /*
1298 : * NB: No need to increment TransamVariables->xactCompletionCount here,
1299 : * nobody can see it yet.
1300 : */
1301 :
1302 230 : LWLockRelease(ProcArrayLock);
1303 :
1304 230 : KnownAssignedXidsDisplay(DEBUG3);
1305 230 : if (standbyState == STANDBY_SNAPSHOT_READY)
1306 230 : elog(DEBUG1, "recovery snapshots are now enabled");
1307 : else
1308 0 : elog(DEBUG1,
1309 : "recovery snapshot waiting for non-overflowed snapshot or "
1310 : "until oldest active xid on standby is at least %u (now %u)",
1311 : standbySnapshotPendingXmin,
1312 : running->oldestRunningXid);
1313 : }
1314 :
1315 : /*
1316 : * ProcArrayApplyXidAssignment
1317 : * Process an XLOG_XACT_ASSIGNMENT WAL record
1318 : */
1319 : void
1320 42 : ProcArrayApplyXidAssignment(TransactionId topxid,
1321 : int nsubxids, TransactionId *subxids)
1322 : {
1323 : TransactionId max_xid;
1324 : int i;
1325 :
1326 : Assert(standbyState >= STANDBY_INITIALIZED);
1327 :
1328 42 : max_xid = TransactionIdLatest(topxid, nsubxids, subxids);
1329 :
1330 : /*
1331 : * Mark all the subtransactions as observed.
1332 : *
1333 : * NOTE: This will fail if the subxid contains too many previously
1334 : * unobserved xids to fit into known-assigned-xids. That shouldn't happen
1335 : * as the code stands, because xid-assignment records should never contain
1336 : * more than PGPROC_MAX_CACHED_SUBXIDS entries.
1337 : */
1338 42 : RecordKnownAssignedTransactionIds(max_xid);
1339 :
1340 : /*
1341 : * Notice that we update pg_subtrans with the top-level xid, rather than
1342 : * the parent xid. This is a difference between normal processing and
1343 : * recovery, yet is still correct in all cases. The reason is that
1344 : * subtransaction commit is not marked in clog until commit processing, so
1345 : * all aborted subtransactions have already been clearly marked in clog.
1346 : * As a result we are able to refer directly to the top-level
1347 : * transaction's state rather than skipping through all the intermediate
1348 : * states in the subtransaction tree. This should be the first time we
1349 : * have attempted to SubTransSetParent().
1350 : */
1351 2730 : for (i = 0; i < nsubxids; i++)
1352 2688 : SubTransSetParent(subxids[i], topxid);
1353 :
1354 : /* KnownAssignedXids isn't maintained yet, so we're done for now */
1355 42 : if (standbyState == STANDBY_INITIALIZED)
1356 0 : return;
1357 :
1358 : /*
1359 : * Uses same locking as transaction commit
1360 : */
1361 42 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
1362 :
1363 : /*
1364 : * Remove subxids from known-assigned-xacts.
1365 : */
1366 42 : KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids);
1367 :
1368 : /*
1369 : * Advance lastOverflowedXid to be at least the last of these subxids.
1370 : */
1371 42 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid))
1372 42 : procArray->lastOverflowedXid = max_xid;
1373 :
1374 42 : LWLockRelease(ProcArrayLock);
1375 : }
1376 :
1377 : /*
1378 : * TransactionIdIsInProgress -- is given transaction running in some backend
1379 : *
1380 : * Aside from some shortcuts such as checking RecentXmin and our own Xid,
1381 : * there are four possibilities for finding a running transaction:
1382 : *
1383 : * 1. The given Xid is a main transaction Id. We will find this out cheaply
1384 : * by looking at ProcGlobal->xids.
1385 : *
1386 : * 2. The given Xid is one of the cached subxact Xids in the PGPROC array.
1387 : * We can find this out cheaply too.
1388 : *
1389 : * 3. In Hot Standby mode, we must search the KnownAssignedXids list to see
1390 : * if the Xid is running on the primary.
1391 : *
1392 : * 4. Search the SubTrans tree to find the Xid's topmost parent, and then see
1393 : * if that is running according to ProcGlobal->xids[] or KnownAssignedXids.
1394 : * This is the slowest way, but sadly it has to be done always if the others
1395 : * failed, unless we see that the cached subxact sets are complete (none have
1396 : * overflowed).
1397 : *
1398 : * ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids
1399 : * while doing 1 and 3, we can release the ProcArrayLock while we do 4.
1400 : * This buys back some concurrency (and we can't retrieve the main Xids from
1401 : * ProcGlobal->xids[] again anyway; see GetNewTransactionId).
1402 : */
1403 : bool
1404 22539484 : TransactionIdIsInProgress(TransactionId xid)
1405 : {
1406 : static TransactionId *xids = NULL;
1407 : static TransactionId *other_xids;
1408 : XidCacheStatus *other_subxidstates;
1409 22539484 : int nxids = 0;
1410 22539484 : ProcArrayStruct *arrayP = procArray;
1411 : TransactionId topxid;
1412 : TransactionId latestCompletedXid;
1413 : int mypgxactoff;
1414 : int numProcs;
1415 : int j;
1416 :
1417 : /*
1418 : * Don't bother checking a transaction older than RecentXmin; it could not
1419 : * possibly still be running. (Note: in particular, this guarantees that
1420 : * we reject InvalidTransactionId, FrozenTransactionId, etc as not
1421 : * running.)
1422 : */
1423 22539484 : if (TransactionIdPrecedes(xid, RecentXmin))
1424 : {
1425 : xc_by_recent_xmin_inc();
1426 9379608 : return false;
1427 : }
1428 :
1429 : /*
1430 : * We may have just checked the status of this transaction, so if it is
1431 : * already known to be completed, we can fall out without any access to
1432 : * shared memory.
1433 : */
1434 13159876 : if (TransactionIdEquals(cachedXidIsNotInProgress, xid))
1435 : {
1436 : xc_by_known_xact_inc();
1437 1781574 : return false;
1438 : }
1439 :
1440 : /*
1441 : * Also, we can handle our own transaction (and subtransactions) without
1442 : * any access to shared memory.
1443 : */
1444 11378302 : if (TransactionIdIsCurrentTransactionId(xid))
1445 : {
1446 : xc_by_my_xact_inc();
1447 400152 : return true;
1448 : }
1449 :
1450 : /*
1451 : * If first time through, get workspace to remember main XIDs in. We
1452 : * malloc it permanently to avoid repeated palloc/pfree overhead.
1453 : */
1454 10978150 : if (xids == NULL)
1455 : {
1456 : /*
1457 : * In hot standby mode, reserve enough space to hold all xids in the
1458 : * known-assigned list. If we later finish recovery, we no longer need
1459 : * the bigger array, but we don't bother to shrink it.
1460 : */
1461 2436 : int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs;
1462 :
1463 2436 : xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId));
1464 2436 : if (xids == NULL)
1465 0 : ereport(ERROR,
1466 : (errcode(ERRCODE_OUT_OF_MEMORY),
1467 : errmsg("out of memory")));
1468 : }
1469 :
1470 10978150 : other_xids = ProcGlobal->xids;
1471 10978150 : other_subxidstates = ProcGlobal->subxidStates;
1472 :
1473 10978150 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1474 :
1475 : /*
1476 : * Now that we have the lock, we can check latestCompletedXid; if the
1477 : * target Xid is after that, it's surely still running.
1478 : */
1479 10978150 : latestCompletedXid =
1480 10978150 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
1481 10978150 : if (TransactionIdPrecedes(latestCompletedXid, xid))
1482 : {
1483 3077534 : LWLockRelease(ProcArrayLock);
1484 : xc_by_latest_xid_inc();
1485 3077534 : return true;
1486 : }
1487 :
1488 : /* No shortcuts, gotta grovel through the array */
1489 7900616 : mypgxactoff = MyProc->pgxactoff;
1490 7900616 : numProcs = arrayP->numProcs;
1491 8265570 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
1492 : {
1493 : int pgprocno;
1494 : PGPROC *proc;
1495 : TransactionId pxid;
1496 : int pxids;
1497 :
1498 : /* Ignore ourselves --- dealt with it above */
1499 8240752 : if (pgxactoff == mypgxactoff)
1500 27780 : continue;
1501 :
1502 : /* Fetch xid just once - see GetNewTransactionId */
1503 8212972 : pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
1504 :
1505 8212972 : if (!TransactionIdIsValid(pxid))
1506 207210 : continue;
1507 :
1508 : /*
1509 : * Step 1: check the main Xid
1510 : */
1511 8005762 : if (TransactionIdEquals(pxid, xid))
1512 : {
1513 7875382 : LWLockRelease(ProcArrayLock);
1514 : xc_by_main_xid_inc();
1515 7875382 : return true;
1516 : }
1517 :
1518 : /*
1519 : * We can ignore main Xids that are younger than the target Xid, since
1520 : * the target could not possibly be their child.
1521 : */
1522 130380 : if (TransactionIdPrecedes(xid, pxid))
1523 66692 : continue;
1524 :
1525 : /*
1526 : * Step 2: check the cached child-Xids arrays
1527 : */
1528 63688 : pxids = other_subxidstates[pgxactoff].count;
1529 63688 : pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */
1530 63688 : pgprocno = arrayP->pgprocnos[pgxactoff];
1531 63688 : proc = &allProcs[pgprocno];
1532 120218 : for (j = pxids - 1; j >= 0; j--)
1533 : {
1534 : /* Fetch xid just once - see GetNewTransactionId */
1535 56946 : TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]);
1536 :
1537 56946 : if (TransactionIdEquals(cxid, xid))
1538 : {
1539 416 : LWLockRelease(ProcArrayLock);
1540 : xc_by_child_xid_inc();
1541 416 : return true;
1542 : }
1543 : }
1544 :
1545 : /*
1546 : * Save the main Xid for step 4. We only need to remember main Xids
1547 : * that have uncached children. (Note: there is no race condition
1548 : * here because the overflowed flag cannot be cleared, only set, while
1549 : * we hold ProcArrayLock. So we can't miss an Xid that we need to
1550 : * worry about.)
1551 : */
1552 63272 : if (other_subxidstates[pgxactoff].overflowed)
1553 496 : xids[nxids++] = pxid;
1554 : }
1555 :
1556 : /*
1557 : * Step 3: in hot standby mode, check the known-assigned-xids list. XIDs
1558 : * in the list must be treated as running.
1559 : */
1560 24818 : if (RecoveryInProgress())
1561 : {
1562 : /* none of the PGPROC entries should have XIDs in hot standby mode */
1563 : Assert(nxids == 0);
1564 :
1565 2 : if (KnownAssignedXidExists(xid))
1566 : {
1567 0 : LWLockRelease(ProcArrayLock);
1568 : xc_by_known_assigned_inc();
1569 0 : return true;
1570 : }
1571 :
1572 : /*
1573 : * If the KnownAssignedXids overflowed, we have to check pg_subtrans
1574 : * too. Fetch all xids from KnownAssignedXids that are lower than
1575 : * xid, since if xid is a subtransaction its parent will always have a
1576 : * lower value. Note we will collect both main and subXIDs here, but
1577 : * there's no help for it.
1578 : */
1579 2 : if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid))
1580 0 : nxids = KnownAssignedXidsGet(xids, xid);
1581 : }
1582 :
1583 24818 : LWLockRelease(ProcArrayLock);
1584 :
1585 : /*
1586 : * If none of the relevant caches overflowed, we know the Xid is not
1587 : * running without even looking at pg_subtrans.
1588 : */
1589 24818 : if (nxids == 0)
1590 : {
1591 : xc_no_overflow_inc();
1592 24322 : cachedXidIsNotInProgress = xid;
1593 24322 : return false;
1594 : }
1595 :
1596 : /*
1597 : * Step 4: have to check pg_subtrans.
1598 : *
1599 : * At this point, we know it's either a subtransaction of one of the Xids
1600 : * in xids[], or it's not running. If it's an already-failed
1601 : * subtransaction, we want to say "not running" even though its parent may
1602 : * still be running. So first, check pg_xact to see if it's been aborted.
1603 : */
1604 : xc_slow_answer_inc();
1605 :
1606 496 : if (TransactionIdDidAbort(xid))
1607 : {
1608 0 : cachedXidIsNotInProgress = xid;
1609 0 : return false;
1610 : }
1611 :
1612 : /*
1613 : * It isn't aborted, so check whether the transaction tree it belongs to
1614 : * is still running (or, more precisely, whether it was running when we
1615 : * held ProcArrayLock).
1616 : */
1617 496 : topxid = SubTransGetTopmostTransaction(xid);
1618 : Assert(TransactionIdIsValid(topxid));
1619 810 : if (!TransactionIdEquals(topxid, xid) &&
1620 314 : pg_lfind32(topxid, xids, nxids))
1621 314 : return true;
1622 :
1623 182 : cachedXidIsNotInProgress = xid;
1624 182 : return false;
1625 : }
1626 :
1627 :
1628 : /*
1629 : * Determine XID horizons.
1630 : *
1631 : * This is used by wrapper functions like GetOldestNonRemovableTransactionId()
1632 : * (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as
1633 : * well as "internally" by GlobalVisUpdate() (see comment above struct
1634 : * GlobalVisState).
1635 : *
1636 : * See the definition of ComputeXidHorizonsResult for the various computed
1637 : * horizons.
1638 : *
1639 : * For VACUUM separate horizons (used to decide which deleted tuples must
1640 : * be preserved), for shared and non-shared tables are computed. For shared
1641 : * relations backends in all databases must be considered, but for non-shared
1642 : * relations that's not required, since only backends in my own database could
1643 : * ever see the tuples in them. Also, we can ignore concurrently running lazy
1644 : * VACUUMs because (a) they must be working on other tables, and (b) they
1645 : * don't need to do snapshot-based lookups.
1646 : *
1647 : * This also computes a horizon used to truncate pg_subtrans. For that
1648 : * backends in all databases have to be considered, and concurrently running
1649 : * lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans
1650 : * accesses.
1651 : *
1652 : * Note: we include all currently running xids in the set of considered xids.
1653 : * This ensures that if a just-started xact has not yet set its snapshot,
1654 : * when it does set the snapshot it cannot set xmin less than what we compute.
1655 : * See notes in src/backend/access/transam/README.
1656 : *
1657 : * Note: despite the above, it's possible for the calculated values to move
1658 : * backwards on repeated calls. The calculated values are conservative, so
1659 : * that anything older is definitely not considered as running by anyone
1660 : * anymore, but the exact values calculated depend on a number of things. For
1661 : * example, if there are no transactions running in the current database, the
1662 : * horizon for normal tables will be latestCompletedXid. If a transaction
1663 : * begins after that, its xmin will include in-progress transactions in other
1664 : * databases that started earlier, so another call will return a lower value.
1665 : * Nonetheless it is safe to vacuum a table in the current database with the
1666 : * first result. There are also replication-related effects: a walsender
1667 : * process can set its xmin based on transactions that are no longer running
1668 : * on the primary but are still being replayed on the standby, thus possibly
1669 : * making the values go backwards. In this case there is a possibility that
1670 : * we lose data that the standby would like to have, but unless the standby
1671 : * uses a replication slot to make its xmin persistent there is little we can
1672 : * do about that --- data is only protected if the walsender runs continuously
1673 : * while queries are executed on the standby. (The Hot Standby code deals
1674 : * with such cases by failing standby queries that needed to access
1675 : * already-removed data, so there's no integrity bug.)
1676 : *
1677 : * Note: the approximate horizons (see definition of GlobalVisState) are
1678 : * updated by the computations done here. That's currently required for
1679 : * correctness and a small optimization. Without doing so it's possible that
1680 : * heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative
1681 : * horizon than later when deciding which tuples can be removed - which the
1682 : * code doesn't expect (breaking HOT).
1683 : */
1684 : static void
1685 341476 : ComputeXidHorizons(ComputeXidHorizonsResult *h)
1686 : {
1687 341476 : ProcArrayStruct *arrayP = procArray;
1688 : TransactionId kaxmin;
1689 341476 : bool in_recovery = RecoveryInProgress();
1690 341476 : TransactionId *other_xids = ProcGlobal->xids;
1691 :
1692 : /* inferred after ProcArrayLock is released */
1693 341476 : h->catalog_oldest_nonremovable = InvalidTransactionId;
1694 :
1695 341476 : LWLockAcquire(ProcArrayLock, LW_SHARED);
1696 :
1697 341476 : h->latest_completed = TransamVariables->latestCompletedXid;
1698 :
1699 : /*
1700 : * We initialize the MIN() calculation with latestCompletedXid + 1. This
1701 : * is a lower bound for the XIDs that might appear in the ProcArray later,
1702 : * and so protects us against overestimating the result due to future
1703 : * additions.
1704 : */
1705 : {
1706 : TransactionId initial;
1707 :
1708 341476 : initial = XidFromFullTransactionId(h->latest_completed);
1709 : Assert(TransactionIdIsValid(initial));
1710 341476 : TransactionIdAdvance(initial);
1711 :
1712 341476 : h->oldest_considered_running = initial;
1713 341476 : h->shared_oldest_nonremovable = initial;
1714 341476 : h->data_oldest_nonremovable = initial;
1715 :
1716 : /*
1717 : * Only modifications made by this backend affect the horizon for
1718 : * temporary relations. Instead of a check in each iteration of the
1719 : * loop over all PGPROCs it is cheaper to just initialize to the
1720 : * current top-level xid any.
1721 : *
1722 : * Without an assigned xid we could use a horizon as aggressive as
1723 : * GetNewTransactionId(), but we can get away with the much cheaper
1724 : * latestCompletedXid + 1: If this backend has no xid there, by
1725 : * definition, can't be any newer changes in the temp table than
1726 : * latestCompletedXid.
1727 : */
1728 341476 : if (TransactionIdIsValid(MyProc->xid))
1729 67316 : h->temp_oldest_nonremovable = MyProc->xid;
1730 : else
1731 274160 : h->temp_oldest_nonremovable = initial;
1732 : }
1733 :
1734 : /*
1735 : * Fetch slot horizons while ProcArrayLock is held - the
1736 : * LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside
1737 : * the lock.
1738 : */
1739 341476 : h->slot_xmin = procArray->replication_slot_xmin;
1740 341476 : h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
1741 :
1742 2056256 : for (int index = 0; index < arrayP->numProcs; index++)
1743 : {
1744 1714780 : int pgprocno = arrayP->pgprocnos[index];
1745 1714780 : PGPROC *proc = &allProcs[pgprocno];
1746 1714780 : int8 statusFlags = ProcGlobal->statusFlags[index];
1747 : TransactionId xid;
1748 : TransactionId xmin;
1749 :
1750 : /* Fetch xid just once - see GetNewTransactionId */
1751 1714780 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
1752 1714780 : xmin = UINT32_ACCESS_ONCE(proc->xmin);
1753 :
1754 : /*
1755 : * Consider both the transaction's Xmin, and its Xid.
1756 : *
1757 : * We must check both because a transaction might have an Xmin but not
1758 : * (yet) an Xid; conversely, if it has an Xid, that could determine
1759 : * some not-yet-set Xmin.
1760 : */
1761 1714780 : xmin = TransactionIdOlder(xmin, xid);
1762 :
1763 : /* if neither is set, this proc doesn't influence the horizon */
1764 1714780 : if (!TransactionIdIsValid(xmin))
1765 761976 : continue;
1766 :
1767 : /*
1768 : * Don't ignore any procs when determining which transactions might be
1769 : * considered running. While slots should ensure logical decoding
1770 : * backends are protected even without this check, it can't hurt to
1771 : * include them here as well..
1772 : */
1773 952804 : h->oldest_considered_running =
1774 952804 : TransactionIdOlder(h->oldest_considered_running, xmin);
1775 :
1776 : /*
1777 : * Skip over backends either vacuuming (which is ok with rows being
1778 : * removed, as long as pg_subtrans is not truncated) or doing logical
1779 : * decoding (which manages xmin separately, check below).
1780 : */
1781 952804 : if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING))
1782 282794 : continue;
1783 :
1784 : /* shared tables need to take backends in all databases into account */
1785 670010 : h->shared_oldest_nonremovable =
1786 670010 : TransactionIdOlder(h->shared_oldest_nonremovable, xmin);
1787 :
1788 : /*
1789 : * Normally sessions in other databases are ignored for anything but
1790 : * the shared horizon.
1791 : *
1792 : * However, include them when MyDatabaseId is not (yet) set. A
1793 : * backend in the process of starting up must not compute a "too
1794 : * aggressive" horizon, otherwise we could end up using it to prune
1795 : * still-needed data away. If the current backend never connects to a
1796 : * database this is harmless, because data_oldest_nonremovable will
1797 : * never be utilized.
1798 : *
1799 : * Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always
1800 : * be included. (This flag is used for hot standby feedback, which
1801 : * can't be tied to a specific database.)
1802 : *
1803 : * Also, while in recovery we cannot compute an accurate per-database
1804 : * horizon, as all xids are managed via the KnownAssignedXids
1805 : * machinery.
1806 : */
1807 670010 : if (proc->databaseId == MyDatabaseId ||
1808 36986 : MyDatabaseId == InvalidOid ||
1809 22432 : (statusFlags & PROC_AFFECTS_ALL_HORIZONS) ||
1810 : in_recovery)
1811 : {
1812 647582 : h->data_oldest_nonremovable =
1813 647582 : TransactionIdOlder(h->data_oldest_nonremovable, xmin);
1814 : }
1815 : }
1816 :
1817 : /*
1818 : * If in recovery fetch oldest xid in KnownAssignedXids, will be applied
1819 : * after lock is released.
1820 : */
1821 341476 : if (in_recovery)
1822 684 : kaxmin = KnownAssignedXidsGetOldestXmin();
1823 :
1824 : /*
1825 : * No other information from shared state is needed, release the lock
1826 : * immediately. The rest of the computations can be done without a lock.
1827 : */
1828 341476 : LWLockRelease(ProcArrayLock);
1829 :
1830 341476 : if (in_recovery)
1831 : {
1832 684 : h->oldest_considered_running =
1833 684 : TransactionIdOlder(h->oldest_considered_running, kaxmin);
1834 684 : h->shared_oldest_nonremovable =
1835 684 : TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin);
1836 684 : h->data_oldest_nonremovable =
1837 684 : TransactionIdOlder(h->data_oldest_nonremovable, kaxmin);
1838 : /* temp relations cannot be accessed in recovery */
1839 : }
1840 :
1841 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1842 : h->shared_oldest_nonremovable));
1843 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1844 : h->data_oldest_nonremovable));
1845 :
1846 : /*
1847 : * Check whether there are replication slots requiring an older xmin.
1848 : */
1849 341476 : h->shared_oldest_nonremovable =
1850 341476 : TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin);
1851 341476 : h->data_oldest_nonremovable =
1852 341476 : TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin);
1853 :
1854 : /*
1855 : * The only difference between catalog / data horizons is that the slot's
1856 : * catalog xmin is applied to the catalog one (so catalogs can be accessed
1857 : * for logical decoding). Initialize with data horizon, and then back up
1858 : * further if necessary. Have to back up the shared horizon as well, since
1859 : * that also can contain catalogs.
1860 : */
1861 341476 : h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable;
1862 341476 : h->shared_oldest_nonremovable =
1863 341476 : TransactionIdOlder(h->shared_oldest_nonremovable,
1864 : h->slot_catalog_xmin);
1865 341476 : h->catalog_oldest_nonremovable = h->data_oldest_nonremovable;
1866 341476 : h->catalog_oldest_nonremovable =
1867 341476 : TransactionIdOlder(h->catalog_oldest_nonremovable,
1868 : h->slot_catalog_xmin);
1869 :
1870 : /*
1871 : * It's possible that slots backed up the horizons further than
1872 : * oldest_considered_running. Fix.
1873 : */
1874 341476 : h->oldest_considered_running =
1875 341476 : TransactionIdOlder(h->oldest_considered_running,
1876 : h->shared_oldest_nonremovable);
1877 341476 : h->oldest_considered_running =
1878 341476 : TransactionIdOlder(h->oldest_considered_running,
1879 : h->catalog_oldest_nonremovable);
1880 341476 : h->oldest_considered_running =
1881 341476 : TransactionIdOlder(h->oldest_considered_running,
1882 : h->data_oldest_nonremovable);
1883 :
1884 : /*
1885 : * shared horizons have to be at least as old as the oldest visible in
1886 : * current db
1887 : */
1888 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1889 : h->data_oldest_nonremovable));
1890 : Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable,
1891 : h->catalog_oldest_nonremovable));
1892 :
1893 : /*
1894 : * Horizons need to ensure that pg_subtrans access is still possible for
1895 : * the relevant backends.
1896 : */
1897 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1898 : h->shared_oldest_nonremovable));
1899 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1900 : h->catalog_oldest_nonremovable));
1901 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1902 : h->data_oldest_nonremovable));
1903 : Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1904 : h->temp_oldest_nonremovable));
1905 : Assert(!TransactionIdIsValid(h->slot_xmin) ||
1906 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1907 : h->slot_xmin));
1908 : Assert(!TransactionIdIsValid(h->slot_catalog_xmin) ||
1909 : TransactionIdPrecedesOrEquals(h->oldest_considered_running,
1910 : h->slot_catalog_xmin));
1911 :
1912 : /* update approximate horizons with the computed horizons */
1913 341476 : GlobalVisUpdateApply(h);
1914 341476 : }
1915 :
1916 : /*
1917 : * Determine what kind of visibility horizon needs to be used for a
1918 : * relation. If rel is NULL, the most conservative horizon is used.
1919 : */
1920 : static inline GlobalVisHorizonKind
1921 29197366 : GlobalVisHorizonKindForRel(Relation rel)
1922 : {
1923 : /*
1924 : * Other relkinds currently don't contain xids, nor always the necessary
1925 : * logical decoding markers.
1926 : */
1927 : Assert(!rel ||
1928 : rel->rd_rel->relkind == RELKIND_RELATION ||
1929 : rel->rd_rel->relkind == RELKIND_MATVIEW ||
1930 : rel->rd_rel->relkind == RELKIND_TOASTVALUE);
1931 :
1932 29197366 : if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress())
1933 202838 : return VISHORIZON_SHARED;
1934 28994528 : else if (IsCatalogRelation(rel) ||
1935 22755434 : RelationIsAccessibleInLogicalDecoding(rel))
1936 6239102 : return VISHORIZON_CATALOG;
1937 22755426 : else if (!RELATION_IS_LOCAL(rel))
1938 22646288 : return VISHORIZON_DATA;
1939 : else
1940 109138 : return VISHORIZON_TEMP;
1941 : }
1942 :
1943 : /*
1944 : * Return the oldest XID for which deleted tuples must be preserved in the
1945 : * passed table.
1946 : *
1947 : * If rel is not NULL the horizon may be considerably more recent than
1948 : * otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon
1949 : * that is correct (but not optimal) for all relations will be returned.
1950 : *
1951 : * This is used by VACUUM to decide which deleted tuples must be preserved in
1952 : * the passed in table.
1953 : */
1954 : TransactionId
1955 230890 : GetOldestNonRemovableTransactionId(Relation rel)
1956 : {
1957 : ComputeXidHorizonsResult horizons;
1958 :
1959 230890 : ComputeXidHorizons(&horizons);
1960 :
1961 230890 : switch (GlobalVisHorizonKindForRel(rel))
1962 : {
1963 34458 : case VISHORIZON_SHARED:
1964 34458 : return horizons.shared_oldest_nonremovable;
1965 133810 : case VISHORIZON_CATALOG:
1966 133810 : return horizons.catalog_oldest_nonremovable;
1967 37402 : case VISHORIZON_DATA:
1968 37402 : return horizons.data_oldest_nonremovable;
1969 25220 : case VISHORIZON_TEMP:
1970 25220 : return horizons.temp_oldest_nonremovable;
1971 : }
1972 :
1973 : /* just to prevent compiler warnings */
1974 0 : return InvalidTransactionId;
1975 : }
1976 :
1977 : /*
1978 : * Return the oldest transaction id any currently running backend might still
1979 : * consider running. This should not be used for visibility / pruning
1980 : * determinations (see GetOldestNonRemovableTransactionId()), but for
1981 : * decisions like up to where pg_subtrans can be truncated.
1982 : */
1983 : TransactionId
1984 3496 : GetOldestTransactionIdConsideredRunning(void)
1985 : {
1986 : ComputeXidHorizonsResult horizons;
1987 :
1988 3496 : ComputeXidHorizons(&horizons);
1989 :
1990 3496 : return horizons.oldest_considered_running;
1991 : }
1992 :
1993 : /*
1994 : * Return the visibility horizons for a hot standby feedback message.
1995 : */
1996 : void
1997 100 : GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin)
1998 : {
1999 : ComputeXidHorizonsResult horizons;
2000 :
2001 100 : ComputeXidHorizons(&horizons);
2002 :
2003 : /*
2004 : * Don't want to use shared_oldest_nonremovable here, as that contains the
2005 : * effect of replication slot's catalog_xmin. We want to send a separate
2006 : * feedback for the catalog horizon, so the primary can remove data table
2007 : * contents more aggressively.
2008 : */
2009 100 : *xmin = horizons.shared_oldest_nonremovable_raw;
2010 100 : *catalog_xmin = horizons.slot_catalog_xmin;
2011 100 : }
2012 :
2013 : /*
2014 : * GetMaxSnapshotXidCount -- get max size for snapshot XID array
2015 : *
2016 : * We have to export this for use by snapmgr.c.
2017 : */
2018 : int
2019 68668 : GetMaxSnapshotXidCount(void)
2020 : {
2021 68668 : return procArray->maxProcs;
2022 : }
2023 :
2024 : /*
2025 : * GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array
2026 : *
2027 : * We have to export this for use by snapmgr.c.
2028 : */
2029 : int
2030 68286 : GetMaxSnapshotSubxidCount(void)
2031 : {
2032 68286 : return TOTAL_MAX_CACHED_SUBXIDS;
2033 : }
2034 :
2035 : /*
2036 : * Helper function for GetSnapshotData() that checks if the bulk of the
2037 : * visibility information in the snapshot is still valid. If so, it updates
2038 : * the fields that need to change and returns true. Otherwise it returns
2039 : * false.
2040 : *
2041 : * This very likely can be evolved to not need ProcArrayLock held (at very
2042 : * least in the case we already hold a snapshot), but that's for another day.
2043 : */
2044 : static bool
2045 4216878 : GetSnapshotDataReuse(Snapshot snapshot)
2046 : {
2047 : uint64 curXactCompletionCount;
2048 :
2049 : Assert(LWLockHeldByMe(ProcArrayLock));
2050 :
2051 4216878 : if (unlikely(snapshot->snapXactCompletionCount == 0))
2052 68248 : return false;
2053 :
2054 4148630 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2055 4148630 : if (curXactCompletionCount != snapshot->snapXactCompletionCount)
2056 735732 : return false;
2057 :
2058 : /*
2059 : * If the current xactCompletionCount is still the same as it was at the
2060 : * time the snapshot was built, we can be sure that rebuilding the
2061 : * contents of the snapshot the hard way would result in the same snapshot
2062 : * contents:
2063 : *
2064 : * As explained in transam/README, the set of xids considered running by
2065 : * GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot
2066 : * contents only depend on transactions with xids and xactCompletionCount
2067 : * is incremented whenever a transaction with an xid finishes (while
2068 : * holding ProcArrayLock exclusively). Thus the xactCompletionCount check
2069 : * ensures we would detect if the snapshot would have changed.
2070 : *
2071 : * As the snapshot contents are the same as it was before, it is safe to
2072 : * re-enter the snapshot's xmin into the PGPROC array. None of the rows
2073 : * visible under the snapshot could already have been removed (that'd
2074 : * require the set of running transactions to change) and it fulfills the
2075 : * requirement that concurrent GetSnapshotData() calls yield the same
2076 : * xmin.
2077 : */
2078 3412898 : if (!TransactionIdIsValid(MyProc->xmin))
2079 1266246 : MyProc->xmin = TransactionXmin = snapshot->xmin;
2080 :
2081 3412898 : RecentXmin = snapshot->xmin;
2082 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2083 :
2084 3412898 : snapshot->curcid = GetCurrentCommandId(false);
2085 3412898 : snapshot->active_count = 0;
2086 3412898 : snapshot->regd_count = 0;
2087 3412898 : snapshot->copied = false;
2088 :
2089 3412898 : return true;
2090 : }
2091 :
2092 : /*
2093 : * GetSnapshotData -- returns information about running transactions.
2094 : *
2095 : * The returned snapshot includes xmin (lowest still-running xact ID),
2096 : * xmax (highest completed xact ID + 1), and a list of running xact IDs
2097 : * in the range xmin <= xid < xmax. It is used as follows:
2098 : * All xact IDs < xmin are considered finished.
2099 : * All xact IDs >= xmax are considered still running.
2100 : * For an xact ID xmin <= xid < xmax, consult list to see whether
2101 : * it is considered running or not.
2102 : * This ensures that the set of transactions seen as "running" by the
2103 : * current xact will not change after it takes the snapshot.
2104 : *
2105 : * All running top-level XIDs are included in the snapshot, except for lazy
2106 : * VACUUM processes. We also try to include running subtransaction XIDs,
2107 : * but since PGPROC has only a limited cache area for subxact XIDs, full
2108 : * information may not be available. If we find any overflowed subxid arrays,
2109 : * we have to mark the snapshot's subxid data as overflowed, and extra work
2110 : * *may* need to be done to determine what's running (see XidInMVCCSnapshot()).
2111 : *
2112 : * We also update the following backend-global variables:
2113 : * TransactionXmin: the oldest xmin of any snapshot in use in the
2114 : * current transaction (this is the same as MyProc->xmin).
2115 : * RecentXmin: the xmin computed for the most recent snapshot. XIDs
2116 : * older than this are known not running any more.
2117 : *
2118 : * And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels
2119 : * for the benefit of the GlobalVisTest* family of functions.
2120 : *
2121 : * Note: this function should probably not be called with an argument that's
2122 : * not statically allocated (see xip allocation below).
2123 : */
2124 : Snapshot
2125 4216878 : GetSnapshotData(Snapshot snapshot)
2126 : {
2127 4216878 : ProcArrayStruct *arrayP = procArray;
2128 4216878 : TransactionId *other_xids = ProcGlobal->xids;
2129 : TransactionId xmin;
2130 : TransactionId xmax;
2131 4216878 : int count = 0;
2132 4216878 : int subcount = 0;
2133 4216878 : bool suboverflowed = false;
2134 : FullTransactionId latest_completed;
2135 : TransactionId oldestxid;
2136 : int mypgxactoff;
2137 : TransactionId myxid;
2138 : uint64 curXactCompletionCount;
2139 :
2140 4216878 : TransactionId replication_slot_xmin = InvalidTransactionId;
2141 4216878 : TransactionId replication_slot_catalog_xmin = InvalidTransactionId;
2142 :
2143 : Assert(snapshot != NULL);
2144 :
2145 : /*
2146 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2147 : * be sufficient. But it seems better to do the malloc while not holding
2148 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2149 : * more subxip storage than is probably needed.
2150 : *
2151 : * This does open a possibility for avoiding repeated malloc/free: since
2152 : * maxProcs does not change at runtime, we can simply reuse the previous
2153 : * xip arrays if any. (This relies on the fact that all callers pass
2154 : * static SnapshotData structs.)
2155 : */
2156 4216878 : if (snapshot->xip == NULL)
2157 : {
2158 : /*
2159 : * First call for this snapshot. Snapshot is same size whether or not
2160 : * we are in recovery, see later comments.
2161 : */
2162 68236 : snapshot->xip = (TransactionId *)
2163 68236 : malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId));
2164 68236 : if (snapshot->xip == NULL)
2165 0 : ereport(ERROR,
2166 : (errcode(ERRCODE_OUT_OF_MEMORY),
2167 : errmsg("out of memory")));
2168 : Assert(snapshot->subxip == NULL);
2169 68236 : snapshot->subxip = (TransactionId *)
2170 68236 : malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId));
2171 68236 : if (snapshot->subxip == NULL)
2172 0 : ereport(ERROR,
2173 : (errcode(ERRCODE_OUT_OF_MEMORY),
2174 : errmsg("out of memory")));
2175 : }
2176 :
2177 : /*
2178 : * It is sufficient to get shared lock on ProcArrayLock, even if we are
2179 : * going to set MyProc->xmin.
2180 : */
2181 4216878 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2182 :
2183 4216878 : if (GetSnapshotDataReuse(snapshot))
2184 : {
2185 3412898 : LWLockRelease(ProcArrayLock);
2186 3412898 : return snapshot;
2187 : }
2188 :
2189 803980 : latest_completed = TransamVariables->latestCompletedXid;
2190 803980 : mypgxactoff = MyProc->pgxactoff;
2191 803980 : myxid = other_xids[mypgxactoff];
2192 : Assert(myxid == MyProc->xid);
2193 :
2194 803980 : oldestxid = TransamVariables->oldestXid;
2195 803980 : curXactCompletionCount = TransamVariables->xactCompletionCount;
2196 :
2197 : /* xmax is always latestCompletedXid + 1 */
2198 803980 : xmax = XidFromFullTransactionId(latest_completed);
2199 803980 : TransactionIdAdvance(xmax);
2200 : Assert(TransactionIdIsNormal(xmax));
2201 :
2202 : /* initialize xmin calculation with xmax */
2203 803980 : xmin = xmax;
2204 :
2205 : /* take own xid into account, saves a check inside the loop */
2206 803980 : if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin))
2207 58864 : xmin = myxid;
2208 :
2209 803980 : snapshot->takenDuringRecovery = RecoveryInProgress();
2210 :
2211 803980 : if (!snapshot->takenDuringRecovery)
2212 : {
2213 801354 : int numProcs = arrayP->numProcs;
2214 801354 : TransactionId *xip = snapshot->xip;
2215 801354 : int *pgprocnos = arrayP->pgprocnos;
2216 801354 : XidCacheStatus *subxidStates = ProcGlobal->subxidStates;
2217 801354 : uint8 *allStatusFlags = ProcGlobal->statusFlags;
2218 :
2219 : /*
2220 : * First collect set of pgxactoff/xids that need to be included in the
2221 : * snapshot.
2222 : */
2223 7024194 : for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++)
2224 : {
2225 : /* Fetch xid just once - see GetNewTransactionId */
2226 6222840 : TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]);
2227 : uint8 statusFlags;
2228 :
2229 : Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff);
2230 :
2231 : /*
2232 : * If the transaction has no XID assigned, we can skip it; it
2233 : * won't have sub-XIDs either.
2234 : */
2235 6222840 : if (likely(xid == InvalidTransactionId))
2236 4595584 : continue;
2237 :
2238 : /*
2239 : * We don't include our own XIDs (if any) in the snapshot. It
2240 : * needs to be included in the xmin computation, but we did so
2241 : * outside the loop.
2242 : */
2243 1627256 : if (pgxactoff == mypgxactoff)
2244 113716 : continue;
2245 :
2246 : /*
2247 : * The only way we are able to get here with a non-normal xid is
2248 : * during bootstrap - with this backend using
2249 : * BootstrapTransactionId. But the above test should filter that
2250 : * out.
2251 : */
2252 : Assert(TransactionIdIsNormal(xid));
2253 :
2254 : /*
2255 : * If the XID is >= xmax, we can skip it; such transactions will
2256 : * be treated as running anyway (and any sub-XIDs will also be >=
2257 : * xmax).
2258 : */
2259 1513540 : if (!NormalTransactionIdPrecedes(xid, xmax))
2260 416184 : continue;
2261 :
2262 : /*
2263 : * Skip over backends doing logical decoding which manages xmin
2264 : * separately (check below) and ones running LAZY VACUUM.
2265 : */
2266 1097356 : statusFlags = allStatusFlags[pgxactoff];
2267 1097356 : if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM))
2268 186 : continue;
2269 :
2270 1097170 : if (NormalTransactionIdPrecedes(xid, xmin))
2271 586744 : xmin = xid;
2272 :
2273 : /* Add XID to snapshot. */
2274 1097170 : xip[count++] = xid;
2275 :
2276 : /*
2277 : * Save subtransaction XIDs if possible (if we've already
2278 : * overflowed, there's no point). Note that the subxact XIDs must
2279 : * be later than their parent, so no need to check them against
2280 : * xmin. We could filter against xmax, but it seems better not to
2281 : * do that much work while holding the ProcArrayLock.
2282 : *
2283 : * The other backend can add more subxids concurrently, but cannot
2284 : * remove any. Hence it's important to fetch nxids just once.
2285 : * Should be safe to use memcpy, though. (We needn't worry about
2286 : * missing any xids added concurrently, because they must postdate
2287 : * xmax.)
2288 : *
2289 : * Again, our own XIDs are not included in the snapshot.
2290 : */
2291 1097170 : if (!suboverflowed)
2292 : {
2293 :
2294 1097162 : if (subxidStates[pgxactoff].overflowed)
2295 1080 : suboverflowed = true;
2296 : else
2297 : {
2298 1096082 : int nsubxids = subxidStates[pgxactoff].count;
2299 :
2300 1096082 : if (nsubxids > 0)
2301 : {
2302 10484 : int pgprocno = pgprocnos[pgxactoff];
2303 10484 : PGPROC *proc = &allProcs[pgprocno];
2304 :
2305 10484 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2306 :
2307 10484 : memcpy(snapshot->subxip + subcount,
2308 10484 : proc->subxids.xids,
2309 : nsubxids * sizeof(TransactionId));
2310 10484 : subcount += nsubxids;
2311 : }
2312 : }
2313 : }
2314 : }
2315 : }
2316 : else
2317 : {
2318 : /*
2319 : * We're in hot standby, so get XIDs from KnownAssignedXids.
2320 : *
2321 : * We store all xids directly into subxip[]. Here's why:
2322 : *
2323 : * In recovery we don't know which xids are top-level and which are
2324 : * subxacts, a design choice that greatly simplifies xid processing.
2325 : *
2326 : * It seems like we would want to try to put xids into xip[] only, but
2327 : * that is fairly small. We would either need to make that bigger or
2328 : * to increase the rate at which we WAL-log xid assignment; neither is
2329 : * an appealing choice.
2330 : *
2331 : * We could try to store xids into xip[] first and then into subxip[]
2332 : * if there are too many xids. That only works if the snapshot doesn't
2333 : * overflow because we do not search subxip[] in that case. A simpler
2334 : * way is to just store all xids in the subxip array because this is
2335 : * by far the bigger array. We just leave the xip array empty.
2336 : *
2337 : * Either way we need to change the way XidInMVCCSnapshot() works
2338 : * depending upon when the snapshot was taken, or change normal
2339 : * snapshot processing so it matches.
2340 : *
2341 : * Note: It is possible for recovery to end before we finish taking
2342 : * the snapshot, and for newly assigned transaction ids to be added to
2343 : * the ProcArray. xmax cannot change while we hold ProcArrayLock, so
2344 : * those newly added transaction ids would be filtered away, so we
2345 : * need not be concerned about them.
2346 : */
2347 2626 : subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin,
2348 : xmax);
2349 :
2350 2626 : if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid))
2351 12 : suboverflowed = true;
2352 : }
2353 :
2354 :
2355 : /*
2356 : * Fetch into local variable while ProcArrayLock is held - the
2357 : * LWLockRelease below is a barrier, ensuring this happens inside the
2358 : * lock.
2359 : */
2360 803980 : replication_slot_xmin = procArray->replication_slot_xmin;
2361 803980 : replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin;
2362 :
2363 803980 : if (!TransactionIdIsValid(MyProc->xmin))
2364 435052 : MyProc->xmin = TransactionXmin = xmin;
2365 :
2366 803980 : LWLockRelease(ProcArrayLock);
2367 :
2368 : /* maintain state for GlobalVis* */
2369 : {
2370 : TransactionId def_vis_xid;
2371 : TransactionId def_vis_xid_data;
2372 : FullTransactionId def_vis_fxid;
2373 : FullTransactionId def_vis_fxid_data;
2374 : FullTransactionId oldestfxid;
2375 :
2376 : /*
2377 : * Converting oldestXid is only safe when xid horizon cannot advance,
2378 : * i.e. holding locks. While we don't hold the lock anymore, all the
2379 : * necessary data has been gathered with lock held.
2380 : */
2381 803980 : oldestfxid = FullXidRelativeTo(latest_completed, oldestxid);
2382 :
2383 : /* Check whether there's a replication slot requiring an older xmin. */
2384 : def_vis_xid_data =
2385 803980 : TransactionIdOlder(xmin, replication_slot_xmin);
2386 :
2387 : /*
2388 : * Rows in non-shared, non-catalog tables possibly could be vacuumed
2389 : * if older than this xid.
2390 : */
2391 803980 : def_vis_xid = def_vis_xid_data;
2392 :
2393 : /*
2394 : * Check whether there's a replication slot requiring an older catalog
2395 : * xmin.
2396 : */
2397 : def_vis_xid =
2398 803980 : TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid);
2399 :
2400 803980 : def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid);
2401 803980 : def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data);
2402 :
2403 : /*
2404 : * Check if we can increase upper bound. As a previous
2405 : * GlobalVisUpdate() might have computed more aggressive values, don't
2406 : * overwrite them if so.
2407 : */
2408 : GlobalVisSharedRels.definitely_needed =
2409 803980 : FullTransactionIdNewer(def_vis_fxid,
2410 : GlobalVisSharedRels.definitely_needed);
2411 : GlobalVisCatalogRels.definitely_needed =
2412 803980 : FullTransactionIdNewer(def_vis_fxid,
2413 : GlobalVisCatalogRels.definitely_needed);
2414 : GlobalVisDataRels.definitely_needed =
2415 803980 : FullTransactionIdNewer(def_vis_fxid_data,
2416 : GlobalVisDataRels.definitely_needed);
2417 : /* See temp_oldest_nonremovable computation in ComputeXidHorizons() */
2418 803980 : if (TransactionIdIsNormal(myxid))
2419 : GlobalVisTempRels.definitely_needed =
2420 113512 : FullXidRelativeTo(latest_completed, myxid);
2421 : else
2422 : {
2423 690468 : GlobalVisTempRels.definitely_needed = latest_completed;
2424 690468 : FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed);
2425 : }
2426 :
2427 : /*
2428 : * Check if we know that we can initialize or increase the lower
2429 : * bound. Currently the only cheap way to do so is to use
2430 : * TransamVariables->oldestXid as input.
2431 : *
2432 : * We should definitely be able to do better. We could e.g. put a
2433 : * global lower bound value into TransamVariables.
2434 : */
2435 : GlobalVisSharedRels.maybe_needed =
2436 803980 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
2437 : oldestfxid);
2438 : GlobalVisCatalogRels.maybe_needed =
2439 803980 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
2440 : oldestfxid);
2441 : GlobalVisDataRels.maybe_needed =
2442 803980 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
2443 : oldestfxid);
2444 : /* accurate value known */
2445 803980 : GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed;
2446 : }
2447 :
2448 803980 : RecentXmin = xmin;
2449 : Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin));
2450 :
2451 803980 : snapshot->xmin = xmin;
2452 803980 : snapshot->xmax = xmax;
2453 803980 : snapshot->xcnt = count;
2454 803980 : snapshot->subxcnt = subcount;
2455 803980 : snapshot->suboverflowed = suboverflowed;
2456 803980 : snapshot->snapXactCompletionCount = curXactCompletionCount;
2457 :
2458 803980 : snapshot->curcid = GetCurrentCommandId(false);
2459 :
2460 : /*
2461 : * This is a new snapshot, so set both refcounts are zero, and mark it as
2462 : * not copied in persistent memory.
2463 : */
2464 803980 : snapshot->active_count = 0;
2465 803980 : snapshot->regd_count = 0;
2466 803980 : snapshot->copied = false;
2467 :
2468 803980 : return snapshot;
2469 : }
2470 :
2471 : /*
2472 : * ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin
2473 : *
2474 : * This is called when installing a snapshot imported from another
2475 : * transaction. To ensure that OldestXmin doesn't go backwards, we must
2476 : * check that the source transaction is still running, and we'd better do
2477 : * that atomically with installing the new xmin.
2478 : *
2479 : * Returns true if successful, false if source xact is no longer running.
2480 : */
2481 : bool
2482 32 : ProcArrayInstallImportedXmin(TransactionId xmin,
2483 : VirtualTransactionId *sourcevxid)
2484 : {
2485 32 : bool result = false;
2486 32 : ProcArrayStruct *arrayP = procArray;
2487 : int index;
2488 :
2489 : Assert(TransactionIdIsNormal(xmin));
2490 32 : if (!sourcevxid)
2491 0 : return false;
2492 :
2493 : /* Get lock so source xact can't end while we're doing this */
2494 32 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2495 :
2496 : /*
2497 : * Find the PGPROC entry of the source transaction. (This could use
2498 : * GetPGProcByNumber(), unless it's a prepared xact. But this isn't
2499 : * performance critical.)
2500 : */
2501 32 : for (index = 0; index < arrayP->numProcs; index++)
2502 : {
2503 32 : int pgprocno = arrayP->pgprocnos[index];
2504 32 : PGPROC *proc = &allProcs[pgprocno];
2505 32 : int statusFlags = ProcGlobal->statusFlags[index];
2506 : TransactionId xid;
2507 :
2508 : /* Ignore procs running LAZY VACUUM */
2509 32 : if (statusFlags & PROC_IN_VACUUM)
2510 0 : continue;
2511 :
2512 : /* We are only interested in the specific virtual transaction. */
2513 32 : if (proc->vxid.procNumber != sourcevxid->procNumber)
2514 0 : continue;
2515 32 : if (proc->vxid.lxid != sourcevxid->localTransactionId)
2516 0 : continue;
2517 :
2518 : /*
2519 : * We check the transaction's database ID for paranoia's sake: if it's
2520 : * in another DB then its xmin does not cover us. Caller should have
2521 : * detected this already, so we just treat any funny cases as
2522 : * "transaction not found".
2523 : */
2524 32 : if (proc->databaseId != MyDatabaseId)
2525 0 : continue;
2526 :
2527 : /*
2528 : * Likewise, let's just make real sure its xmin does cover us.
2529 : */
2530 32 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2531 32 : if (!TransactionIdIsNormal(xid) ||
2532 32 : !TransactionIdPrecedesOrEquals(xid, xmin))
2533 0 : continue;
2534 :
2535 : /*
2536 : * We're good. Install the new xmin. As in GetSnapshotData, set
2537 : * TransactionXmin too. (Note that because snapmgr.c called
2538 : * GetSnapshotData first, we'll be overwriting a valid xmin here, so
2539 : * we don't check that.)
2540 : */
2541 32 : MyProc->xmin = TransactionXmin = xmin;
2542 :
2543 32 : result = true;
2544 32 : break;
2545 : }
2546 :
2547 32 : LWLockRelease(ProcArrayLock);
2548 :
2549 32 : return result;
2550 : }
2551 :
2552 : /*
2553 : * ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin
2554 : *
2555 : * This is like ProcArrayInstallImportedXmin, but we have a pointer to the
2556 : * PGPROC of the transaction from which we imported the snapshot, rather than
2557 : * an XID.
2558 : *
2559 : * Note that this function also copies statusFlags from the source `proc` in
2560 : * order to avoid the case where MyProc's xmin needs to be skipped for
2561 : * computing xid horizon.
2562 : *
2563 : * Returns true if successful, false if source xact is no longer running.
2564 : */
2565 : bool
2566 3268 : ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc)
2567 : {
2568 3268 : bool result = false;
2569 : TransactionId xid;
2570 :
2571 : Assert(TransactionIdIsNormal(xmin));
2572 : Assert(proc != NULL);
2573 :
2574 : /*
2575 : * Get an exclusive lock so that we can copy statusFlags from source proc.
2576 : */
2577 3268 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
2578 :
2579 : /*
2580 : * Be certain that the referenced PGPROC has an advertised xmin which is
2581 : * no later than the one we're installing, so that the system-wide xmin
2582 : * can't go backwards. Also, make sure it's running in the same database,
2583 : * so that the per-database xmin cannot go backwards.
2584 : */
2585 3268 : xid = UINT32_ACCESS_ONCE(proc->xmin);
2586 3268 : if (proc->databaseId == MyDatabaseId &&
2587 3268 : TransactionIdIsNormal(xid) &&
2588 3268 : TransactionIdPrecedesOrEquals(xid, xmin))
2589 : {
2590 : /*
2591 : * Install xmin and propagate the statusFlags that affect how the
2592 : * value is interpreted by vacuum.
2593 : */
2594 3268 : MyProc->xmin = TransactionXmin = xmin;
2595 3268 : MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) |
2596 3268 : (proc->statusFlags & PROC_XMIN_FLAGS);
2597 3268 : ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags;
2598 :
2599 3268 : result = true;
2600 : }
2601 :
2602 3268 : LWLockRelease(ProcArrayLock);
2603 :
2604 3268 : return result;
2605 : }
2606 :
2607 : /*
2608 : * GetRunningTransactionData -- returns information about running transactions.
2609 : *
2610 : * Similar to GetSnapshotData but returns more information. We include
2611 : * all PGPROCs with an assigned TransactionId, even VACUUM processes and
2612 : * prepared transactions.
2613 : *
2614 : * We acquire XidGenLock and ProcArrayLock, but the caller is responsible for
2615 : * releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc
2616 : * array until the caller has WAL-logged this snapshot, and releases the
2617 : * lock. Acquiring ProcArrayLock ensures that no transactions commit until the
2618 : * lock is released.
2619 : *
2620 : * The returned data structure is statically allocated; caller should not
2621 : * modify it, and must not assume it is valid past the next call.
2622 : *
2623 : * This is never executed during recovery so there is no need to look at
2624 : * KnownAssignedXids.
2625 : *
2626 : * Dummy PGPROCs from prepared transaction are included, meaning that this
2627 : * may return entries with duplicated TransactionId values coming from
2628 : * transaction finishing to prepare. Nothing is done about duplicated
2629 : * entries here to not hold on ProcArrayLock more than necessary.
2630 : *
2631 : * We don't worry about updating other counters, we want to keep this as
2632 : * simple as possible and leave GetSnapshotData() as the primary code for
2633 : * that bookkeeping.
2634 : *
2635 : * Note that if any transaction has overflowed its cached subtransactions
2636 : * then there is no real need include any subtransactions.
2637 : */
2638 : RunningTransactions
2639 2894 : GetRunningTransactionData(void)
2640 : {
2641 : /* result workspace */
2642 : static RunningTransactionsData CurrentRunningXactsData;
2643 :
2644 2894 : ProcArrayStruct *arrayP = procArray;
2645 2894 : TransactionId *other_xids = ProcGlobal->xids;
2646 2894 : RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData;
2647 : TransactionId latestCompletedXid;
2648 : TransactionId oldestRunningXid;
2649 : TransactionId oldestDatabaseRunningXid;
2650 : TransactionId *xids;
2651 : int index;
2652 : int count;
2653 : int subcount;
2654 : bool suboverflowed;
2655 :
2656 : Assert(!RecoveryInProgress());
2657 :
2658 : /*
2659 : * Allocating space for maxProcs xids is usually overkill; numProcs would
2660 : * be sufficient. But it seems better to do the malloc while not holding
2661 : * the lock, so we can't look at numProcs. Likewise, we allocate much
2662 : * more subxip storage than is probably needed.
2663 : *
2664 : * Should only be allocated in bgwriter, since only ever executed during
2665 : * checkpoints.
2666 : */
2667 2894 : if (CurrentRunningXacts->xids == NULL)
2668 : {
2669 : /*
2670 : * First call
2671 : */
2672 1112 : CurrentRunningXacts->xids = (TransactionId *)
2673 1112 : malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId));
2674 1112 : if (CurrentRunningXacts->xids == NULL)
2675 0 : ereport(ERROR,
2676 : (errcode(ERRCODE_OUT_OF_MEMORY),
2677 : errmsg("out of memory")));
2678 : }
2679 :
2680 2894 : xids = CurrentRunningXacts->xids;
2681 :
2682 2894 : count = subcount = 0;
2683 2894 : suboverflowed = false;
2684 :
2685 : /*
2686 : * Ensure that no xids enter or leave the procarray while we obtain
2687 : * snapshot.
2688 : */
2689 2894 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2690 2894 : LWLockAcquire(XidGenLock, LW_SHARED);
2691 :
2692 2894 : latestCompletedXid =
2693 2894 : XidFromFullTransactionId(TransamVariables->latestCompletedXid);
2694 2894 : oldestDatabaseRunningXid = oldestRunningXid =
2695 2894 : XidFromFullTransactionId(TransamVariables->nextXid);
2696 :
2697 : /*
2698 : * Spin over procArray collecting all xids
2699 : */
2700 14302 : for (index = 0; index < arrayP->numProcs; index++)
2701 : {
2702 : TransactionId xid;
2703 :
2704 : /* Fetch xid just once - see GetNewTransactionId */
2705 11408 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2706 :
2707 : /*
2708 : * We don't need to store transactions that don't have a TransactionId
2709 : * yet because they will not show as running on a standby server.
2710 : */
2711 11408 : if (!TransactionIdIsValid(xid))
2712 9402 : continue;
2713 :
2714 : /*
2715 : * Be careful not to exclude any xids before calculating the values of
2716 : * oldestRunningXid and suboverflowed, since these are used to clean
2717 : * up transaction information held on standbys.
2718 : */
2719 2006 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2720 1300 : oldestRunningXid = xid;
2721 :
2722 : /*
2723 : * Also, update the oldest running xid within the current database. As
2724 : * fetching pgprocno and PGPROC could cause cache misses, we do cheap
2725 : * TransactionId comparison first.
2726 : */
2727 2006 : if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid))
2728 : {
2729 2006 : int pgprocno = arrayP->pgprocnos[index];
2730 2006 : PGPROC *proc = &allProcs[pgprocno];
2731 :
2732 2006 : if (proc->databaseId == MyDatabaseId)
2733 426 : oldestDatabaseRunningXid = xid;
2734 : }
2735 :
2736 2006 : if (ProcGlobal->subxidStates[index].overflowed)
2737 4 : suboverflowed = true;
2738 :
2739 : /*
2740 : * If we wished to exclude xids this would be the right place for it.
2741 : * Procs with the PROC_IN_VACUUM flag set don't usually assign xids,
2742 : * but they do during truncation at the end when they get the lock and
2743 : * truncate, so it is not much of a problem to include them if they
2744 : * are seen and it is cleaner to include them.
2745 : */
2746 :
2747 2006 : xids[count++] = xid;
2748 : }
2749 :
2750 : /*
2751 : * Spin over procArray collecting all subxids, but only if there hasn't
2752 : * been a suboverflow.
2753 : */
2754 2894 : if (!suboverflowed)
2755 : {
2756 2890 : XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates;
2757 :
2758 14286 : for (index = 0; index < arrayP->numProcs; index++)
2759 : {
2760 11396 : int pgprocno = arrayP->pgprocnos[index];
2761 11396 : PGPROC *proc = &allProcs[pgprocno];
2762 : int nsubxids;
2763 :
2764 : /*
2765 : * Save subtransaction XIDs. Other backends can't add or remove
2766 : * entries while we're holding XidGenLock.
2767 : */
2768 11396 : nsubxids = other_subxidstates[index].count;
2769 11396 : if (nsubxids > 0)
2770 : {
2771 : /* barrier not really required, as XidGenLock is held, but ... */
2772 42 : pg_read_barrier(); /* pairs with GetNewTransactionId */
2773 :
2774 42 : memcpy(&xids[count], proc->subxids.xids,
2775 : nsubxids * sizeof(TransactionId));
2776 42 : count += nsubxids;
2777 42 : subcount += nsubxids;
2778 :
2779 : /*
2780 : * Top-level XID of a transaction is always less than any of
2781 : * its subxids, so we don't need to check if any of the
2782 : * subxids are smaller than oldestRunningXid
2783 : */
2784 : }
2785 : }
2786 : }
2787 :
2788 : /*
2789 : * It's important *not* to include the limits set by slots here because
2790 : * snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those
2791 : * were to be included here the initial value could never increase because
2792 : * of a circular dependency where slots only increase their limits when
2793 : * running xacts increases oldestRunningXid and running xacts only
2794 : * increases if slots do.
2795 : */
2796 :
2797 2894 : CurrentRunningXacts->xcnt = count - subcount;
2798 2894 : CurrentRunningXacts->subxcnt = subcount;
2799 2894 : CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY;
2800 2894 : CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid);
2801 2894 : CurrentRunningXacts->oldestRunningXid = oldestRunningXid;
2802 2894 : CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid;
2803 2894 : CurrentRunningXacts->latestCompletedXid = latestCompletedXid;
2804 :
2805 : Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid));
2806 : Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid));
2807 : Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid));
2808 :
2809 : /* We don't release the locks here, the caller is responsible for that */
2810 :
2811 2894 : return CurrentRunningXacts;
2812 : }
2813 :
2814 : /*
2815 : * GetOldestActiveTransactionId()
2816 : *
2817 : * Similar to GetSnapshotData but returns just oldestActiveXid. We include
2818 : * all PGPROCs with an assigned TransactionId, even VACUUM processes.
2819 : *
2820 : * If allDbs is true, we look at all databases, though there is no need to
2821 : * include WALSender since this has no effect on hot standby conflicts. If
2822 : * allDbs is false, skip processes attached to other databases.
2823 : *
2824 : * This is never executed during recovery so there is no need to look at
2825 : * KnownAssignedXids.
2826 : *
2827 : * We don't worry about updating other counters, we want to keep this as
2828 : * simple as possible and leave GetSnapshotData() as the primary code for
2829 : * that bookkeeping.
2830 : *
2831 : * inCommitOnly indicates getting the oldestActiveXid among the transactions
2832 : * in the commit critical section.
2833 : */
2834 : TransactionId
2835 9088 : GetOldestActiveTransactionId(bool inCommitOnly, bool allDbs)
2836 : {
2837 9088 : ProcArrayStruct *arrayP = procArray;
2838 9088 : TransactionId *other_xids = ProcGlobal->xids;
2839 : TransactionId oldestRunningXid;
2840 : int index;
2841 :
2842 : Assert(!RecoveryInProgress());
2843 :
2844 : /*
2845 : * Read nextXid, as the upper bound of what's still active.
2846 : *
2847 : * Reading a TransactionId is atomic, but we must grab the lock to make
2848 : * sure that all XIDs < nextXid are already present in the proc array (or
2849 : * have already completed), when we spin over it.
2850 : */
2851 9088 : LWLockAcquire(XidGenLock, LW_SHARED);
2852 9088 : oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid);
2853 9088 : LWLockRelease(XidGenLock);
2854 :
2855 : /*
2856 : * Spin over procArray collecting all xids and subxids.
2857 : */
2858 9088 : LWLockAcquire(ProcArrayLock, LW_SHARED);
2859 50224 : for (index = 0; index < arrayP->numProcs; index++)
2860 : {
2861 : TransactionId xid;
2862 41136 : int pgprocno = arrayP->pgprocnos[index];
2863 41136 : PGPROC *proc = &allProcs[pgprocno];
2864 :
2865 : /* Fetch xid just once - see GetNewTransactionId */
2866 41136 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2867 :
2868 41136 : if (!TransactionIdIsNormal(xid))
2869 32394 : continue;
2870 :
2871 8742 : if (inCommitOnly &&
2872 6894 : (proc->delayChkptFlags & DELAY_CHKPT_IN_COMMIT) == 0)
2873 5110 : continue;
2874 :
2875 3632 : if (!allDbs && proc->databaseId != MyDatabaseId)
2876 0 : continue;
2877 :
2878 3632 : if (TransactionIdPrecedes(xid, oldestRunningXid))
2879 2960 : oldestRunningXid = xid;
2880 :
2881 : /*
2882 : * Top-level XID of a transaction is always less than any of its
2883 : * subxids, so we don't need to check if any of the subxids are
2884 : * smaller than oldestRunningXid
2885 : */
2886 : }
2887 9088 : LWLockRelease(ProcArrayLock);
2888 :
2889 9088 : return oldestRunningXid;
2890 : }
2891 :
2892 : /*
2893 : * GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum
2894 : *
2895 : * Returns the oldest xid that we can guarantee not to have been affected by
2896 : * vacuum, i.e. no rows >= that xid have been vacuumed away unless the
2897 : * transaction aborted. Note that the value can (and most of the time will) be
2898 : * much more conservative than what really has been affected by vacuum, but we
2899 : * currently don't have better data available.
2900 : *
2901 : * This is useful to initialize the cutoff xid after which a new changeset
2902 : * extraction replication slot can start decoding changes.
2903 : *
2904 : * Must be called with ProcArrayLock held either shared or exclusively,
2905 : * although most callers will want to use exclusive mode since it is expected
2906 : * that the caller will immediately use the xid to peg the xmin horizon.
2907 : */
2908 : TransactionId
2909 1366 : GetOldestSafeDecodingTransactionId(bool catalogOnly)
2910 : {
2911 1366 : ProcArrayStruct *arrayP = procArray;
2912 : TransactionId oldestSafeXid;
2913 : int index;
2914 1366 : bool recovery_in_progress = RecoveryInProgress();
2915 :
2916 : Assert(LWLockHeldByMe(ProcArrayLock));
2917 :
2918 : /*
2919 : * Acquire XidGenLock, so no transactions can acquire an xid while we're
2920 : * running. If no transaction with xid were running concurrently a new xid
2921 : * could influence the RecentXmin et al.
2922 : *
2923 : * We initialize the computation to nextXid since that's guaranteed to be
2924 : * a safe, albeit pessimal, value.
2925 : */
2926 1366 : LWLockAcquire(XidGenLock, LW_SHARED);
2927 1366 : oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid);
2928 :
2929 : /*
2930 : * If there's already a slot pegging the xmin horizon, we can start with
2931 : * that value, it's guaranteed to be safe since it's computed by this
2932 : * routine initially and has been enforced since. We can always use the
2933 : * slot's general xmin horizon, but the catalog horizon is only usable
2934 : * when only catalog data is going to be looked at.
2935 : */
2936 1824 : if (TransactionIdIsValid(procArray->replication_slot_xmin) &&
2937 458 : TransactionIdPrecedes(procArray->replication_slot_xmin,
2938 : oldestSafeXid))
2939 24 : oldestSafeXid = procArray->replication_slot_xmin;
2940 :
2941 1366 : if (catalogOnly &&
2942 698 : TransactionIdIsValid(procArray->replication_slot_catalog_xmin) &&
2943 140 : TransactionIdPrecedes(procArray->replication_slot_catalog_xmin,
2944 : oldestSafeXid))
2945 56 : oldestSafeXid = procArray->replication_slot_catalog_xmin;
2946 :
2947 : /*
2948 : * If we're not in recovery, we walk over the procarray and collect the
2949 : * lowest xid. Since we're called with ProcArrayLock held and have
2950 : * acquired XidGenLock, no entries can vanish concurrently, since
2951 : * ProcGlobal->xids[i] is only set with XidGenLock held and only cleared
2952 : * with ProcArrayLock held.
2953 : *
2954 : * In recovery we can't lower the safe value besides what we've computed
2955 : * above, so we'll have to wait a bit longer there. We unfortunately can
2956 : * *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids
2957 : * machinery can miss values and return an older value than is safe.
2958 : */
2959 1366 : if (!recovery_in_progress)
2960 : {
2961 1302 : TransactionId *other_xids = ProcGlobal->xids;
2962 :
2963 : /*
2964 : * Spin over procArray collecting min(ProcGlobal->xids[i])
2965 : */
2966 6680 : for (index = 0; index < arrayP->numProcs; index++)
2967 : {
2968 : TransactionId xid;
2969 :
2970 : /* Fetch xid just once - see GetNewTransactionId */
2971 5378 : xid = UINT32_ACCESS_ONCE(other_xids[index]);
2972 :
2973 5378 : if (!TransactionIdIsNormal(xid))
2974 5360 : continue;
2975 :
2976 18 : if (TransactionIdPrecedes(xid, oldestSafeXid))
2977 16 : oldestSafeXid = xid;
2978 : }
2979 : }
2980 :
2981 1366 : LWLockRelease(XidGenLock);
2982 :
2983 1366 : return oldestSafeXid;
2984 : }
2985 :
2986 : /*
2987 : * GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are
2988 : * delaying checkpoint because they have critical actions in progress.
2989 : *
2990 : * Constructs an array of VXIDs of transactions that are currently in commit
2991 : * critical sections, as shown by having specified delayChkptFlags bits set
2992 : * in their PGPROC.
2993 : *
2994 : * Returns a palloc'd array that should be freed by the caller.
2995 : * *nvxids is the number of valid entries.
2996 : *
2997 : * Note that because backends set or clear delayChkptFlags without holding any
2998 : * lock, the result is somewhat indeterminate, but we don't really care. Even
2999 : * in a multiprocessor with delayed writes to shared memory, it should be
3000 : * certain that setting of delayChkptFlags will propagate to shared memory
3001 : * when the backend takes a lock, so we cannot fail to see a virtual xact as
3002 : * delayChkptFlags if it's already inserted its commit record. Whether it
3003 : * takes a little while for clearing of delayChkptFlags to propagate is
3004 : * unimportant for correctness.
3005 : */
3006 : VirtualTransactionId *
3007 6336 : GetVirtualXIDsDelayingChkpt(int *nvxids, int type)
3008 : {
3009 : VirtualTransactionId *vxids;
3010 6336 : ProcArrayStruct *arrayP = procArray;
3011 6336 : int count = 0;
3012 : int index;
3013 :
3014 : Assert(type != 0);
3015 :
3016 : /* allocate what's certainly enough result space */
3017 6336 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3018 :
3019 6336 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3020 :
3021 21072 : for (index = 0; index < arrayP->numProcs; index++)
3022 : {
3023 14736 : int pgprocno = arrayP->pgprocnos[index];
3024 14736 : PGPROC *proc = &allProcs[pgprocno];
3025 :
3026 14736 : if ((proc->delayChkptFlags & type) != 0)
3027 : {
3028 : VirtualTransactionId vxid;
3029 :
3030 72 : GET_VXID_FROM_PGPROC(vxid, *proc);
3031 72 : if (VirtualTransactionIdIsValid(vxid))
3032 72 : vxids[count++] = vxid;
3033 : }
3034 : }
3035 :
3036 6336 : LWLockRelease(ProcArrayLock);
3037 :
3038 6336 : *nvxids = count;
3039 6336 : return vxids;
3040 : }
3041 :
3042 : /*
3043 : * HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying?
3044 : *
3045 : * This is used with the results of GetVirtualXIDsDelayingChkpt to see if any
3046 : * of the specified VXIDs are still in critical sections of code.
3047 : *
3048 : * Note: this is O(N^2) in the number of vxacts that are/were delaying, but
3049 : * those numbers should be small enough for it not to be a problem.
3050 : */
3051 : bool
3052 94 : HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type)
3053 : {
3054 94 : bool result = false;
3055 94 : ProcArrayStruct *arrayP = procArray;
3056 : int index;
3057 :
3058 : Assert(type != 0);
3059 :
3060 94 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3061 :
3062 776 : for (index = 0; index < arrayP->numProcs; index++)
3063 : {
3064 716 : int pgprocno = arrayP->pgprocnos[index];
3065 716 : PGPROC *proc = &allProcs[pgprocno];
3066 : VirtualTransactionId vxid;
3067 :
3068 716 : GET_VXID_FROM_PGPROC(vxid, *proc);
3069 :
3070 716 : if ((proc->delayChkptFlags & type) != 0 &&
3071 64 : VirtualTransactionIdIsValid(vxid))
3072 : {
3073 : int i;
3074 :
3075 110 : for (i = 0; i < nvxids; i++)
3076 : {
3077 80 : if (VirtualTransactionIdEquals(vxid, vxids[i]))
3078 : {
3079 34 : result = true;
3080 34 : break;
3081 : }
3082 : }
3083 64 : if (result)
3084 34 : break;
3085 : }
3086 : }
3087 :
3088 94 : LWLockRelease(ProcArrayLock);
3089 :
3090 94 : return result;
3091 : }
3092 :
3093 : /*
3094 : * ProcNumberGetProc -- get a backend's PGPROC given its proc number
3095 : *
3096 : * The result may be out of date arbitrarily quickly, so the caller
3097 : * must be careful about how this information is used. NULL is
3098 : * returned if the backend is not active.
3099 : */
3100 : PGPROC *
3101 1150 : ProcNumberGetProc(ProcNumber procNumber)
3102 : {
3103 : PGPROC *result;
3104 :
3105 1150 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3106 4 : return NULL;
3107 1146 : result = GetPGProcByNumber(procNumber);
3108 :
3109 1146 : if (result->pid == 0)
3110 6 : return NULL;
3111 :
3112 1140 : return result;
3113 : }
3114 :
3115 : /*
3116 : * ProcNumberGetTransactionIds -- get a backend's transaction status
3117 : *
3118 : * Get the xid, xmin, nsubxid and overflow status of the backend. The
3119 : * result may be out of date arbitrarily quickly, so the caller must be
3120 : * careful about how this information is used.
3121 : */
3122 : void
3123 25522 : ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid,
3124 : TransactionId *xmin, int *nsubxid, bool *overflowed)
3125 : {
3126 : PGPROC *proc;
3127 :
3128 25522 : *xid = InvalidTransactionId;
3129 25522 : *xmin = InvalidTransactionId;
3130 25522 : *nsubxid = 0;
3131 25522 : *overflowed = false;
3132 :
3133 25522 : if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount)
3134 0 : return;
3135 25522 : proc = GetPGProcByNumber(procNumber);
3136 :
3137 : /* Need to lock out additions/removals of backends */
3138 25522 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3139 :
3140 25522 : if (proc->pid != 0)
3141 : {
3142 25522 : *xid = proc->xid;
3143 25522 : *xmin = proc->xmin;
3144 25522 : *nsubxid = proc->subxidStatus.count;
3145 25522 : *overflowed = proc->subxidStatus.overflowed;
3146 : }
3147 :
3148 25522 : LWLockRelease(ProcArrayLock);
3149 : }
3150 :
3151 : /*
3152 : * BackendPidGetProc -- get a backend's PGPROC given its PID
3153 : *
3154 : * Returns NULL if not found. Note that it is up to the caller to be
3155 : * sure that the question remains meaningful for long enough for the
3156 : * answer to be used ...
3157 : */
3158 : PGPROC *
3159 28188 : BackendPidGetProc(int pid)
3160 : {
3161 : PGPROC *result;
3162 :
3163 28188 : if (pid == 0) /* never match dummy PGPROCs */
3164 6 : return NULL;
3165 :
3166 28182 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3167 :
3168 28182 : result = BackendPidGetProcWithLock(pid);
3169 :
3170 28182 : LWLockRelease(ProcArrayLock);
3171 :
3172 28182 : return result;
3173 : }
3174 :
3175 : /*
3176 : * BackendPidGetProcWithLock -- get a backend's PGPROC given its PID
3177 : *
3178 : * Same as above, except caller must be holding ProcArrayLock. The found
3179 : * entry, if any, can be assumed to be valid as long as the lock remains held.
3180 : */
3181 : PGPROC *
3182 32336 : BackendPidGetProcWithLock(int pid)
3183 : {
3184 32336 : PGPROC *result = NULL;
3185 32336 : ProcArrayStruct *arrayP = procArray;
3186 : int index;
3187 :
3188 32336 : if (pid == 0) /* never match dummy PGPROCs */
3189 0 : return NULL;
3190 :
3191 127062 : for (index = 0; index < arrayP->numProcs; index++)
3192 : {
3193 112844 : PGPROC *proc = &allProcs[arrayP->pgprocnos[index]];
3194 :
3195 112844 : if (proc->pid == pid)
3196 : {
3197 18118 : result = proc;
3198 18118 : break;
3199 : }
3200 : }
3201 :
3202 32336 : return result;
3203 : }
3204 :
3205 : /*
3206 : * BackendXidGetPid -- get a backend's pid given its XID
3207 : *
3208 : * Returns 0 if not found or it's a prepared transaction. Note that
3209 : * it is up to the caller to be sure that the question remains
3210 : * meaningful for long enough for the answer to be used ...
3211 : *
3212 : * Only main transaction Ids are considered. This function is mainly
3213 : * useful for determining what backend owns a lock.
3214 : *
3215 : * Beware that not every xact has an XID assigned. However, as long as you
3216 : * only call this using an XID found on disk, you're safe.
3217 : */
3218 : int
3219 60 : BackendXidGetPid(TransactionId xid)
3220 : {
3221 60 : int result = 0;
3222 60 : ProcArrayStruct *arrayP = procArray;
3223 60 : TransactionId *other_xids = ProcGlobal->xids;
3224 : int index;
3225 :
3226 60 : if (xid == InvalidTransactionId) /* never match invalid xid */
3227 0 : return 0;
3228 :
3229 60 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3230 :
3231 184 : for (index = 0; index < arrayP->numProcs; index++)
3232 : {
3233 168 : if (other_xids[index] == xid)
3234 : {
3235 44 : int pgprocno = arrayP->pgprocnos[index];
3236 44 : PGPROC *proc = &allProcs[pgprocno];
3237 :
3238 44 : result = proc->pid;
3239 44 : break;
3240 : }
3241 : }
3242 :
3243 60 : LWLockRelease(ProcArrayLock);
3244 :
3245 60 : return result;
3246 : }
3247 :
3248 : /*
3249 : * IsBackendPid -- is a given pid a running backend
3250 : *
3251 : * This is not called by the backend, but is called by external modules.
3252 : */
3253 : bool
3254 4 : IsBackendPid(int pid)
3255 : {
3256 4 : return (BackendPidGetProc(pid) != NULL);
3257 : }
3258 :
3259 :
3260 : /*
3261 : * GetCurrentVirtualXIDs -- returns an array of currently active VXIDs.
3262 : *
3263 : * The array is palloc'd. The number of valid entries is returned into *nvxids.
3264 : *
3265 : * The arguments allow filtering the set of VXIDs returned. Our own process
3266 : * is always skipped. In addition:
3267 : * If limitXmin is not InvalidTransactionId, skip processes with
3268 : * xmin > limitXmin.
3269 : * If excludeXmin0 is true, skip processes with xmin = 0.
3270 : * If allDbs is false, skip processes attached to other databases.
3271 : * If excludeVacuum isn't zero, skip processes for which
3272 : * (statusFlags & excludeVacuum) is not zero.
3273 : *
3274 : * Note: the purpose of the limitXmin and excludeXmin0 parameters is to
3275 : * allow skipping backends whose oldest live snapshot is no older than
3276 : * some snapshot we have. Since we examine the procarray with only shared
3277 : * lock, there are race conditions: a backend could set its xmin just after
3278 : * we look. Indeed, on multiprocessors with weak memory ordering, the
3279 : * other backend could have set its xmin *before* we look. We know however
3280 : * that such a backend must have held shared ProcArrayLock overlapping our
3281 : * own hold of ProcArrayLock, else we would see its xmin update. Therefore,
3282 : * any snapshot the other backend is taking concurrently with our scan cannot
3283 : * consider any transactions as still running that we think are committed
3284 : * (since backends must hold ProcArrayLock exclusive to commit).
3285 : */
3286 : VirtualTransactionId *
3287 868 : GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0,
3288 : bool allDbs, int excludeVacuum,
3289 : int *nvxids)
3290 : {
3291 : VirtualTransactionId *vxids;
3292 868 : ProcArrayStruct *arrayP = procArray;
3293 868 : int count = 0;
3294 : int index;
3295 :
3296 : /* allocate what's certainly enough result space */
3297 868 : vxids = palloc_array(VirtualTransactionId, arrayP->maxProcs);
3298 :
3299 868 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3300 :
3301 5204 : for (index = 0; index < arrayP->numProcs; index++)
3302 : {
3303 4336 : int pgprocno = arrayP->pgprocnos[index];
3304 4336 : PGPROC *proc = &allProcs[pgprocno];
3305 4336 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3306 :
3307 4336 : if (proc == MyProc)
3308 868 : continue;
3309 :
3310 3468 : if (excludeVacuum & statusFlags)
3311 36 : continue;
3312 :
3313 3432 : if (allDbs || proc->databaseId == MyDatabaseId)
3314 : {
3315 : /* Fetch xmin just once - might change on us */
3316 1540 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3317 :
3318 1540 : if (excludeXmin0 && !TransactionIdIsValid(pxmin))
3319 866 : continue;
3320 :
3321 : /*
3322 : * InvalidTransactionId precedes all other XIDs, so a proc that
3323 : * hasn't set xmin yet will not be rejected by this test.
3324 : */
3325 1348 : if (!TransactionIdIsValid(limitXmin) ||
3326 674 : TransactionIdPrecedesOrEquals(pxmin, limitXmin))
3327 : {
3328 : VirtualTransactionId vxid;
3329 :
3330 626 : GET_VXID_FROM_PGPROC(vxid, *proc);
3331 626 : if (VirtualTransactionIdIsValid(vxid))
3332 626 : vxids[count++] = vxid;
3333 : }
3334 : }
3335 : }
3336 :
3337 868 : LWLockRelease(ProcArrayLock);
3338 :
3339 868 : *nvxids = count;
3340 868 : return vxids;
3341 : }
3342 :
3343 : /*
3344 : * GetConflictingVirtualXIDs -- returns an array of currently active VXIDs.
3345 : *
3346 : * Usage is limited to conflict resolution during recovery on standby servers.
3347 : * limitXmin is supplied as either a cutoff with snapshotConflictHorizon
3348 : * semantics, or InvalidTransactionId in cases where caller cannot accurately
3349 : * determine a safe snapshotConflictHorizon value.
3350 : *
3351 : * If limitXmin is InvalidTransactionId then we want to kill everybody,
3352 : * so we're not worried if they have a snapshot or not, nor does it really
3353 : * matter what type of lock we hold. Caller must avoid calling here with
3354 : * snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId
3355 : * during original execution, since that actually indicates that there is
3356 : * definitely no need for a recovery conflict (the snapshotConflictHorizon
3357 : * convention for InvalidTransactionId values is the opposite of our own!).
3358 : *
3359 : * All callers that are checking xmins always now supply a valid and useful
3360 : * value for limitXmin. The limitXmin is always lower than the lowest
3361 : * numbered KnownAssignedXid that is not already a FATAL error. This is
3362 : * because we only care about cleanup records that are cleaning up tuple
3363 : * versions from committed transactions. In that case they will only occur
3364 : * at the point where the record is less than the lowest running xid. That
3365 : * allows us to say that if any backend takes a snapshot concurrently with
3366 : * us then the conflict assessment made here would never include the snapshot
3367 : * that is being derived. So we take LW_SHARED on the ProcArray and allow
3368 : * concurrent snapshots when limitXmin is valid. We might think about adding
3369 : * Assert(limitXmin < lowest(KnownAssignedXids))
3370 : * but that would not be true in the case of FATAL errors lagging in array,
3371 : * but we already know those are bogus anyway, so we skip that test.
3372 : *
3373 : * If dbOid is valid we skip backends attached to other databases.
3374 : *
3375 : * Be careful to *not* pfree the result from this function. We reuse
3376 : * this array sufficiently often that we use malloc for the result.
3377 : */
3378 : VirtualTransactionId *
3379 26432 : GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid)
3380 : {
3381 : static VirtualTransactionId *vxids;
3382 26432 : ProcArrayStruct *arrayP = procArray;
3383 26432 : int count = 0;
3384 : int index;
3385 :
3386 : /*
3387 : * If first time through, get workspace to remember main XIDs in. We
3388 : * malloc it permanently to avoid repeated palloc/pfree overhead. Allow
3389 : * result space, remembering room for a terminator.
3390 : */
3391 26432 : if (vxids == NULL)
3392 : {
3393 56 : vxids = (VirtualTransactionId *)
3394 56 : malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1));
3395 56 : if (vxids == NULL)
3396 0 : ereport(ERROR,
3397 : (errcode(ERRCODE_OUT_OF_MEMORY),
3398 : errmsg("out of memory")));
3399 : }
3400 :
3401 26432 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3402 :
3403 26994 : for (index = 0; index < arrayP->numProcs; index++)
3404 : {
3405 562 : int pgprocno = arrayP->pgprocnos[index];
3406 562 : PGPROC *proc = &allProcs[pgprocno];
3407 :
3408 : /* Exclude prepared transactions */
3409 562 : if (proc->pid == 0)
3410 0 : continue;
3411 :
3412 562 : if (!OidIsValid(dbOid) ||
3413 550 : proc->databaseId == dbOid)
3414 : {
3415 : /* Fetch xmin just once - can't change on us, but good coding */
3416 32 : TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin);
3417 :
3418 : /*
3419 : * We ignore an invalid pxmin because this means that backend has
3420 : * no snapshot currently. We hold a Share lock to avoid contention
3421 : * with users taking snapshots. That is not a problem because the
3422 : * current xmin is always at least one higher than the latest
3423 : * removed xid, so any new snapshot would never conflict with the
3424 : * test here.
3425 : */
3426 32 : if (!TransactionIdIsValid(limitXmin) ||
3427 6 : (TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin)))
3428 : {
3429 : VirtualTransactionId vxid;
3430 :
3431 4 : GET_VXID_FROM_PGPROC(vxid, *proc);
3432 4 : if (VirtualTransactionIdIsValid(vxid))
3433 4 : vxids[count++] = vxid;
3434 : }
3435 : }
3436 : }
3437 :
3438 26432 : LWLockRelease(ProcArrayLock);
3439 :
3440 : /* add the terminator */
3441 26432 : vxids[count].procNumber = INVALID_PROC_NUMBER;
3442 26432 : vxids[count].localTransactionId = InvalidLocalTransactionId;
3443 :
3444 26432 : return vxids;
3445 : }
3446 :
3447 : /*
3448 : * CancelVirtualTransaction - used in recovery conflict processing
3449 : *
3450 : * Returns pid of the process signaled, or 0 if not found.
3451 : */
3452 : pid_t
3453 6 : CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode)
3454 : {
3455 6 : return SignalVirtualTransaction(vxid, sigmode, true);
3456 : }
3457 :
3458 : pid_t
3459 10 : SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode,
3460 : bool conflictPending)
3461 : {
3462 10 : ProcArrayStruct *arrayP = procArray;
3463 : int index;
3464 10 : pid_t pid = 0;
3465 :
3466 10 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3467 :
3468 10 : for (index = 0; index < arrayP->numProcs; index++)
3469 : {
3470 10 : int pgprocno = arrayP->pgprocnos[index];
3471 10 : PGPROC *proc = &allProcs[pgprocno];
3472 : VirtualTransactionId procvxid;
3473 :
3474 10 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3475 :
3476 10 : if (procvxid.procNumber == vxid.procNumber &&
3477 10 : procvxid.localTransactionId == vxid.localTransactionId)
3478 : {
3479 10 : proc->recoveryConflictPending = conflictPending;
3480 10 : pid = proc->pid;
3481 10 : if (pid != 0)
3482 : {
3483 : /*
3484 : * Kill the pid if it's still here. If not, that's what we
3485 : * wanted so ignore any errors.
3486 : */
3487 10 : (void) SendProcSignal(pid, sigmode, vxid.procNumber);
3488 : }
3489 10 : break;
3490 : }
3491 : }
3492 :
3493 10 : LWLockRelease(ProcArrayLock);
3494 :
3495 10 : return pid;
3496 : }
3497 :
3498 : /*
3499 : * MinimumActiveBackends --- count backends (other than myself) that are
3500 : * in active transactions. Return true if the count exceeds the
3501 : * minimum threshold passed. This is used as a heuristic to decide if
3502 : * a pre-XLOG-flush delay is worthwhile during commit.
3503 : *
3504 : * Do not count backends that are blocked waiting for locks, since they are
3505 : * not going to get to run until someone else commits.
3506 : */
3507 : bool
3508 0 : MinimumActiveBackends(int min)
3509 : {
3510 0 : ProcArrayStruct *arrayP = procArray;
3511 0 : int count = 0;
3512 : int index;
3513 :
3514 : /* Quick short-circuit if no minimum is specified */
3515 0 : if (min == 0)
3516 0 : return true;
3517 :
3518 : /*
3519 : * Note: for speed, we don't acquire ProcArrayLock. This is a little bit
3520 : * bogus, but since we are only testing fields for zero or nonzero, it
3521 : * should be OK. The result is only used for heuristic purposes anyway...
3522 : */
3523 0 : for (index = 0; index < arrayP->numProcs; index++)
3524 : {
3525 0 : int pgprocno = arrayP->pgprocnos[index];
3526 0 : PGPROC *proc = &allProcs[pgprocno];
3527 :
3528 : /*
3529 : * Since we're not holding a lock, need to be prepared to deal with
3530 : * garbage, as someone could have incremented numProcs but not yet
3531 : * filled the structure.
3532 : *
3533 : * If someone just decremented numProcs, 'proc' could also point to a
3534 : * PGPROC entry that's no longer in the array. It still points to a
3535 : * PGPROC struct, though, because freed PGPROC entries just go to the
3536 : * free list and are recycled. Its contents are nonsense in that case,
3537 : * but that's acceptable for this function.
3538 : */
3539 0 : if (pgprocno == -1)
3540 0 : continue; /* do not count deleted entries */
3541 0 : if (proc == MyProc)
3542 0 : continue; /* do not count myself */
3543 0 : if (proc->xid == InvalidTransactionId)
3544 0 : continue; /* do not count if no XID assigned */
3545 0 : if (proc->pid == 0)
3546 0 : continue; /* do not count prepared xacts */
3547 0 : if (proc->waitLock != NULL)
3548 0 : continue; /* do not count if blocked on a lock */
3549 0 : count++;
3550 0 : if (count >= min)
3551 0 : break;
3552 : }
3553 :
3554 0 : return count >= min;
3555 : }
3556 :
3557 : /*
3558 : * CountDBBackends --- count backends that are using specified database
3559 : */
3560 : int
3561 32 : CountDBBackends(Oid databaseid)
3562 : {
3563 32 : ProcArrayStruct *arrayP = procArray;
3564 32 : int count = 0;
3565 : int index;
3566 :
3567 32 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3568 :
3569 48 : for (index = 0; index < arrayP->numProcs; index++)
3570 : {
3571 16 : int pgprocno = arrayP->pgprocnos[index];
3572 16 : PGPROC *proc = &allProcs[pgprocno];
3573 :
3574 16 : if (proc->pid == 0)
3575 0 : continue; /* do not count prepared xacts */
3576 16 : if (!OidIsValid(databaseid) ||
3577 16 : proc->databaseId == databaseid)
3578 4 : count++;
3579 : }
3580 :
3581 32 : LWLockRelease(ProcArrayLock);
3582 :
3583 32 : return count;
3584 : }
3585 :
3586 : /*
3587 : * CountDBConnections --- counts database backends (only regular backends)
3588 : */
3589 : int
3590 0 : CountDBConnections(Oid databaseid)
3591 : {
3592 0 : ProcArrayStruct *arrayP = procArray;
3593 0 : int count = 0;
3594 : int index;
3595 :
3596 0 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3597 :
3598 0 : for (index = 0; index < arrayP->numProcs; index++)
3599 : {
3600 0 : int pgprocno = arrayP->pgprocnos[index];
3601 0 : PGPROC *proc = &allProcs[pgprocno];
3602 :
3603 0 : if (proc->pid == 0)
3604 0 : continue; /* do not count prepared xacts */
3605 0 : if (!proc->isRegularBackend)
3606 0 : continue; /* count only regular backend processes */
3607 0 : if (!OidIsValid(databaseid) ||
3608 0 : proc->databaseId == databaseid)
3609 0 : count++;
3610 : }
3611 :
3612 0 : LWLockRelease(ProcArrayLock);
3613 :
3614 0 : return count;
3615 : }
3616 :
3617 : /*
3618 : * CancelDBBackends --- cancel backends that are using specified database
3619 : */
3620 : void
3621 20 : CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending)
3622 : {
3623 20 : ProcArrayStruct *arrayP = procArray;
3624 : int index;
3625 :
3626 : /* tell all backends to die */
3627 20 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3628 :
3629 40 : for (index = 0; index < arrayP->numProcs; index++)
3630 : {
3631 20 : int pgprocno = arrayP->pgprocnos[index];
3632 20 : PGPROC *proc = &allProcs[pgprocno];
3633 :
3634 20 : if (databaseid == InvalidOid || proc->databaseId == databaseid)
3635 : {
3636 : VirtualTransactionId procvxid;
3637 : pid_t pid;
3638 :
3639 20 : GET_VXID_FROM_PGPROC(procvxid, *proc);
3640 :
3641 20 : proc->recoveryConflictPending = conflictPending;
3642 20 : pid = proc->pid;
3643 20 : if (pid != 0)
3644 : {
3645 : /*
3646 : * Kill the pid if it's still here. If not, that's what we
3647 : * wanted so ignore any errors.
3648 : */
3649 20 : (void) SendProcSignal(pid, sigmode, procvxid.procNumber);
3650 : }
3651 : }
3652 : }
3653 :
3654 20 : LWLockRelease(ProcArrayLock);
3655 20 : }
3656 :
3657 : /*
3658 : * CountUserBackends --- count backends that are used by specified user
3659 : * (only regular backends, not any type of background worker)
3660 : */
3661 : int
3662 0 : CountUserBackends(Oid roleid)
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->isRegularBackend)
3678 0 : continue; /* count only regular backend processes */
3679 0 : if (proc->roleId == roleid)
3680 0 : count++;
3681 : }
3682 :
3683 0 : LWLockRelease(ProcArrayLock);
3684 :
3685 0 : return count;
3686 : }
3687 :
3688 : /*
3689 : * CountOtherDBBackends -- check for other backends running in the given DB
3690 : *
3691 : * If there are other backends in the DB, we will wait a maximum of 5 seconds
3692 : * for them to exit (or 0.3s for testing purposes). Autovacuum backends are
3693 : * encouraged to exit early by sending them SIGTERM, but normal user backends
3694 : * are just waited for. If background workers connected to this database are
3695 : * marked as interruptible, they are terminated.
3696 : *
3697 : * The current backend is always ignored; it is caller's responsibility to
3698 : * check whether the current backend uses the given DB, if it's important.
3699 : *
3700 : * Returns true if there are (still) other backends in the DB, false if not.
3701 : * Also, *nbackends and *nprepared are set to the number of other backends
3702 : * and prepared transactions in the DB, respectively.
3703 : *
3704 : * This function is used to interlock DROP DATABASE and related commands
3705 : * against there being any active backends in the target DB --- dropping the
3706 : * DB while active backends remain would be a Bad Thing. Note that we cannot
3707 : * detect here the possibility of a newly-started backend that is trying to
3708 : * connect to the doomed database, so additional interlocking is needed during
3709 : * backend startup. The caller should normally hold an exclusive lock on the
3710 : * target DB before calling this, which is one reason we mustn't wait
3711 : * indefinitely.
3712 : */
3713 : bool
3714 918 : CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared)
3715 : {
3716 918 : ProcArrayStruct *arrayP = procArray;
3717 :
3718 : #define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */
3719 : int autovac_pids[MAXAUTOVACPIDS];
3720 :
3721 : /*
3722 : * Retry up to 50 times with 100ms between attempts (max 5s total). Can be
3723 : * reduced to 3 attempts (max 0.3s total) to speed up tests.
3724 : */
3725 918 : int ntries = 50;
3726 :
3727 : #ifdef USE_INJECTION_POINTS
3728 918 : if (IS_INJECTION_POINT_ATTACHED("procarray-reduce-count"))
3729 2 : ntries = 3;
3730 : #endif
3731 :
3732 934 : for (int tries = 0; tries < ntries; tries++)
3733 : {
3734 932 : int nautovacs = 0;
3735 932 : bool found = false;
3736 : int index;
3737 :
3738 932 : CHECK_FOR_INTERRUPTS();
3739 :
3740 932 : *nbackends = *nprepared = 0;
3741 :
3742 932 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3743 :
3744 3454 : for (index = 0; index < arrayP->numProcs; index++)
3745 : {
3746 2522 : int pgprocno = arrayP->pgprocnos[index];
3747 2522 : PGPROC *proc = &allProcs[pgprocno];
3748 2522 : uint8 statusFlags = ProcGlobal->statusFlags[index];
3749 :
3750 2522 : if (proc->databaseId != databaseId)
3751 2300 : continue;
3752 222 : if (proc == MyProc)
3753 206 : continue;
3754 :
3755 16 : found = true;
3756 :
3757 16 : if (proc->pid == 0)
3758 0 : (*nprepared)++;
3759 : else
3760 : {
3761 16 : (*nbackends)++;
3762 16 : if ((statusFlags & PROC_IS_AUTOVACUUM) &&
3763 : nautovacs < MAXAUTOVACPIDS)
3764 0 : autovac_pids[nautovacs++] = proc->pid;
3765 : }
3766 : }
3767 :
3768 932 : LWLockRelease(ProcArrayLock);
3769 :
3770 932 : if (!found)
3771 916 : return false; /* no conflicting backends, so done */
3772 :
3773 : /*
3774 : * Send SIGTERM to any conflicting autovacuums before sleeping. We
3775 : * postpone this step until after the loop because we don't want to
3776 : * hold ProcArrayLock while issuing kill(). We have no idea what might
3777 : * block kill() inside the kernel...
3778 : */
3779 16 : for (index = 0; index < nautovacs; index++)
3780 0 : (void) kill(autovac_pids[index], SIGTERM); /* ignore any error */
3781 :
3782 : /*
3783 : * Terminate all background workers for this database, if they have
3784 : * requested it (BGWORKER_INTERRUPTIBLE).
3785 : */
3786 16 : TerminateBackgroundWorkersForDatabase(databaseId);
3787 :
3788 : /* sleep, then try again */
3789 16 : pg_usleep(100 * 1000L); /* 100ms */
3790 : }
3791 :
3792 2 : return true; /* timed out, still conflicts */
3793 : }
3794 :
3795 : /*
3796 : * Terminate existing connections to the specified database. This routine
3797 : * is used by the DROP DATABASE command when user has asked to forcefully
3798 : * drop the database.
3799 : *
3800 : * The current backend is always ignored; it is caller's responsibility to
3801 : * check whether the current backend uses the given DB, if it's important.
3802 : *
3803 : * If the target database has a prepared transaction or permissions checks
3804 : * fail for a connection, this fails without terminating anything.
3805 : */
3806 : void
3807 2 : TerminateOtherDBBackends(Oid databaseId)
3808 : {
3809 2 : ProcArrayStruct *arrayP = procArray;
3810 2 : List *pids = NIL;
3811 2 : int nprepared = 0;
3812 : int i;
3813 :
3814 2 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3815 :
3816 8 : for (i = 0; i < procArray->numProcs; i++)
3817 : {
3818 6 : int pgprocno = arrayP->pgprocnos[i];
3819 6 : PGPROC *proc = &allProcs[pgprocno];
3820 :
3821 6 : if (proc->databaseId != databaseId)
3822 6 : continue;
3823 0 : if (proc == MyProc)
3824 0 : continue;
3825 :
3826 0 : if (proc->pid != 0)
3827 0 : pids = lappend_int(pids, proc->pid);
3828 : else
3829 0 : nprepared++;
3830 : }
3831 :
3832 2 : LWLockRelease(ProcArrayLock);
3833 :
3834 2 : if (nprepared > 0)
3835 0 : ereport(ERROR,
3836 : (errcode(ERRCODE_OBJECT_IN_USE),
3837 : errmsg("database \"%s\" is being used by prepared transactions",
3838 : get_database_name(databaseId)),
3839 : errdetail_plural("There is %d prepared transaction using the database.",
3840 : "There are %d prepared transactions using the database.",
3841 : nprepared,
3842 : nprepared)));
3843 :
3844 2 : if (pids)
3845 : {
3846 : ListCell *lc;
3847 :
3848 : /*
3849 : * Permissions checks relax the pg_terminate_backend checks in two
3850 : * ways, both by omitting the !OidIsValid(proc->roleId) check:
3851 : *
3852 : * - Accept terminating autovacuum workers, since DROP DATABASE
3853 : * without FORCE terminates them.
3854 : *
3855 : * - Accept terminating bgworkers. For bgworker authors, it's
3856 : * convenient to be able to recommend FORCE if a worker is blocking
3857 : * DROP DATABASE unexpectedly.
3858 : *
3859 : * Unlike pg_terminate_backend, we don't raise some warnings - like
3860 : * "PID %d is not a PostgreSQL server process", because for us already
3861 : * finished session is not a problem.
3862 : */
3863 0 : foreach(lc, pids)
3864 : {
3865 0 : int pid = lfirst_int(lc);
3866 0 : PGPROC *proc = BackendPidGetProc(pid);
3867 :
3868 0 : if (proc != NULL)
3869 : {
3870 0 : if (superuser_arg(proc->roleId) && !superuser())
3871 0 : ereport(ERROR,
3872 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3873 : errmsg("permission denied to terminate process"),
3874 : errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.",
3875 : "SUPERUSER", "SUPERUSER")));
3876 :
3877 0 : if (!has_privs_of_role(GetUserId(), proc->roleId) &&
3878 0 : !has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND))
3879 0 : ereport(ERROR,
3880 : (errcode(ERRCODE_INSUFFICIENT_PRIVILEGE),
3881 : errmsg("permission denied to terminate process"),
3882 : errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.",
3883 : "pg_signal_backend")));
3884 : }
3885 : }
3886 :
3887 : /*
3888 : * There's a race condition here: once we release the ProcArrayLock,
3889 : * it's possible for the session to exit before we issue kill. That
3890 : * race condition possibility seems too unlikely to worry about. See
3891 : * pg_signal_backend.
3892 : */
3893 0 : foreach(lc, pids)
3894 : {
3895 0 : int pid = lfirst_int(lc);
3896 0 : PGPROC *proc = BackendPidGetProc(pid);
3897 :
3898 0 : if (proc != NULL)
3899 : {
3900 : /*
3901 : * If we have setsid(), signal the backend's whole process
3902 : * group
3903 : */
3904 : #ifdef HAVE_SETSID
3905 0 : (void) kill(-pid, SIGTERM);
3906 : #else
3907 : (void) kill(pid, SIGTERM);
3908 : #endif
3909 : }
3910 : }
3911 : }
3912 2 : }
3913 :
3914 : /*
3915 : * ProcArraySetReplicationSlotXmin
3916 : *
3917 : * Install limits to future computations of the xmin horizon to prevent vacuum
3918 : * and HOT pruning from removing affected rows still needed by clients with
3919 : * replication slots.
3920 : */
3921 : void
3922 4862 : ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin,
3923 : bool already_locked)
3924 : {
3925 : Assert(!already_locked || LWLockHeldByMe(ProcArrayLock));
3926 :
3927 4862 : if (!already_locked)
3928 3896 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3929 :
3930 4862 : procArray->replication_slot_xmin = xmin;
3931 4862 : procArray->replication_slot_catalog_xmin = catalog_xmin;
3932 :
3933 4862 : if (!already_locked)
3934 3896 : LWLockRelease(ProcArrayLock);
3935 :
3936 4862 : elog(DEBUG1, "xmin required by slots: data %u, catalog %u",
3937 : xmin, catalog_xmin);
3938 4862 : }
3939 :
3940 : /*
3941 : * ProcArrayGetReplicationSlotXmin
3942 : *
3943 : * Return the current slot xmin limits. That's useful to be able to remove
3944 : * data that's older than those limits.
3945 : */
3946 : void
3947 44 : ProcArrayGetReplicationSlotXmin(TransactionId *xmin,
3948 : TransactionId *catalog_xmin)
3949 : {
3950 44 : LWLockAcquire(ProcArrayLock, LW_SHARED);
3951 :
3952 44 : if (xmin != NULL)
3953 0 : *xmin = procArray->replication_slot_xmin;
3954 :
3955 44 : if (catalog_xmin != NULL)
3956 44 : *catalog_xmin = procArray->replication_slot_catalog_xmin;
3957 :
3958 44 : LWLockRelease(ProcArrayLock);
3959 44 : }
3960 :
3961 : /*
3962 : * XidCacheRemoveRunningXids
3963 : *
3964 : * Remove a bunch of TransactionIds from the list of known-running
3965 : * subtransactions for my backend. Both the specified xid and those in
3966 : * the xids[] array (of length nxids) are removed from the subxids cache.
3967 : * latestXid must be the latest XID among the group.
3968 : */
3969 : void
3970 1336 : XidCacheRemoveRunningXids(TransactionId xid,
3971 : int nxids, const TransactionId *xids,
3972 : TransactionId latestXid)
3973 : {
3974 : int i,
3975 : j;
3976 : XidCacheStatus *mysubxidstat;
3977 :
3978 : Assert(TransactionIdIsValid(xid));
3979 :
3980 : /*
3981 : * We must hold ProcArrayLock exclusively in order to remove transactions
3982 : * from the PGPROC array. (See src/backend/access/transam/README.) It's
3983 : * possible this could be relaxed since we know this routine is only used
3984 : * to abort subtransactions, but pending closer analysis we'd best be
3985 : * conservative.
3986 : *
3987 : * Note that we do not have to be careful about memory ordering of our own
3988 : * reads wrt. GetNewTransactionId() here - only this process can modify
3989 : * relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be
3990 : * careful about our own writes being well ordered.
3991 : */
3992 1336 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
3993 :
3994 1336 : mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff];
3995 :
3996 : /*
3997 : * Under normal circumstances xid and xids[] will be in increasing order,
3998 : * as will be the entries in subxids. Scan backwards to avoid O(N^2)
3999 : * behavior when removing a lot of xids.
4000 : */
4001 1396 : for (i = nxids - 1; i >= 0; i--)
4002 : {
4003 60 : TransactionId anxid = xids[i];
4004 :
4005 60 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4006 : {
4007 60 : if (TransactionIdEquals(MyProc->subxids.xids[j], anxid))
4008 : {
4009 60 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4010 60 : pg_write_barrier();
4011 60 : mysubxidstat->count--;
4012 60 : MyProc->subxidStatus.count--;
4013 60 : break;
4014 : }
4015 : }
4016 :
4017 : /*
4018 : * Ordinarily we should have found it, unless the cache has
4019 : * overflowed. However it's also possible for this routine to be
4020 : * invoked multiple times for the same subtransaction, in case of an
4021 : * error during AbortSubTransaction. So instead of Assert, emit a
4022 : * debug warning.
4023 : */
4024 60 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4025 0 : elog(WARNING, "did not find subXID %u in MyProc", anxid);
4026 : }
4027 :
4028 1464 : for (j = MyProc->subxidStatus.count - 1; j >= 0; j--)
4029 : {
4030 1462 : if (TransactionIdEquals(MyProc->subxids.xids[j], xid))
4031 : {
4032 1334 : MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1];
4033 1334 : pg_write_barrier();
4034 1334 : mysubxidstat->count--;
4035 1334 : MyProc->subxidStatus.count--;
4036 1334 : break;
4037 : }
4038 : }
4039 : /* Ordinarily we should have found it, unless the cache has overflowed */
4040 1336 : if (j < 0 && !MyProc->subxidStatus.overflowed)
4041 0 : elog(WARNING, "did not find subXID %u in MyProc", xid);
4042 :
4043 : /* Also advance global latestCompletedXid while holding the lock */
4044 1336 : MaintainLatestCompletedXid(latestXid);
4045 :
4046 : /* ... and xactCompletionCount */
4047 1336 : TransamVariables->xactCompletionCount++;
4048 :
4049 1336 : LWLockRelease(ProcArrayLock);
4050 1336 : }
4051 :
4052 : #ifdef XIDCACHE_DEBUG
4053 :
4054 : /*
4055 : * Print stats about effectiveness of XID cache
4056 : */
4057 : static void
4058 : DisplayXidCache(void)
4059 : {
4060 : fprintf(stderr,
4061 : "XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n",
4062 : xc_by_recent_xmin,
4063 : xc_by_known_xact,
4064 : xc_by_my_xact,
4065 : xc_by_latest_xid,
4066 : xc_by_main_xid,
4067 : xc_by_child_xid,
4068 : xc_by_known_assigned,
4069 : xc_no_overflow,
4070 : xc_slow_answer);
4071 : }
4072 : #endif /* XIDCACHE_DEBUG */
4073 :
4074 : /*
4075 : * If rel != NULL, return test state appropriate for relation, otherwise
4076 : * return state usable for all relations. The latter may consider XIDs as
4077 : * not-yet-visible-to-everyone that a state for a specific relation would
4078 : * already consider visible-to-everyone.
4079 : *
4080 : * This needs to be called while a snapshot is active or registered, otherwise
4081 : * there are wraparound and other dangers.
4082 : *
4083 : * See comment for GlobalVisState for details.
4084 : */
4085 : GlobalVisState *
4086 28966476 : GlobalVisTestFor(Relation rel)
4087 : {
4088 28966476 : GlobalVisState *state = NULL;
4089 :
4090 : /* XXX: we should assert that a snapshot is pushed or registered */
4091 : Assert(RecentXmin);
4092 :
4093 28966476 : switch (GlobalVisHorizonKindForRel(rel))
4094 : {
4095 168380 : case VISHORIZON_SHARED:
4096 168380 : state = &GlobalVisSharedRels;
4097 168380 : break;
4098 6105292 : case VISHORIZON_CATALOG:
4099 6105292 : state = &GlobalVisCatalogRels;
4100 6105292 : break;
4101 22608886 : case VISHORIZON_DATA:
4102 22608886 : state = &GlobalVisDataRels;
4103 22608886 : break;
4104 83918 : case VISHORIZON_TEMP:
4105 83918 : state = &GlobalVisTempRels;
4106 83918 : break;
4107 : }
4108 :
4109 : Assert(FullTransactionIdIsValid(state->definitely_needed) &&
4110 : FullTransactionIdIsValid(state->maybe_needed));
4111 :
4112 28966476 : return state;
4113 : }
4114 :
4115 : /*
4116 : * Return true if it's worth updating the accurate maybe_needed boundary.
4117 : *
4118 : * As it is somewhat expensive to determine xmin horizons, we don't want to
4119 : * repeatedly do so when there is a low likelihood of it being beneficial.
4120 : *
4121 : * The current heuristic is that we update only if RecentXmin has changed
4122 : * since the last update. If the oldest currently running transaction has not
4123 : * finished, it is unlikely that recomputing the horizon would be useful.
4124 : */
4125 : static bool
4126 901586 : GlobalVisTestShouldUpdate(GlobalVisState *state)
4127 : {
4128 : /* hasn't been updated yet */
4129 901586 : if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin))
4130 17850 : return true;
4131 :
4132 : /*
4133 : * If the maybe_needed/definitely_needed boundaries are the same, it's
4134 : * unlikely to be beneficial to refresh boundaries.
4135 : */
4136 883736 : if (FullTransactionIdFollowsOrEquals(state->maybe_needed,
4137 : state->definitely_needed))
4138 0 : return false;
4139 :
4140 : /* does the last snapshot built have a different xmin? */
4141 883736 : return RecentXmin != ComputeXidHorizonsResultLastXmin;
4142 : }
4143 :
4144 : static void
4145 341476 : GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons)
4146 : {
4147 : GlobalVisSharedRels.maybe_needed =
4148 341476 : FullXidRelativeTo(horizons->latest_completed,
4149 : horizons->shared_oldest_nonremovable);
4150 : GlobalVisCatalogRels.maybe_needed =
4151 341476 : FullXidRelativeTo(horizons->latest_completed,
4152 : horizons->catalog_oldest_nonremovable);
4153 : GlobalVisDataRels.maybe_needed =
4154 341476 : FullXidRelativeTo(horizons->latest_completed,
4155 : horizons->data_oldest_nonremovable);
4156 : GlobalVisTempRels.maybe_needed =
4157 341476 : FullXidRelativeTo(horizons->latest_completed,
4158 : horizons->temp_oldest_nonremovable);
4159 :
4160 : /*
4161 : * In longer running transactions it's possible that transactions we
4162 : * previously needed to treat as running aren't around anymore. So update
4163 : * definitely_needed to not be earlier than maybe_needed.
4164 : */
4165 : GlobalVisSharedRels.definitely_needed =
4166 341476 : FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed,
4167 : GlobalVisSharedRels.definitely_needed);
4168 : GlobalVisCatalogRels.definitely_needed =
4169 341476 : FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed,
4170 : GlobalVisCatalogRels.definitely_needed);
4171 : GlobalVisDataRels.definitely_needed =
4172 341476 : FullTransactionIdNewer(GlobalVisDataRels.maybe_needed,
4173 : GlobalVisDataRels.definitely_needed);
4174 341476 : GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed;
4175 :
4176 341476 : ComputeXidHorizonsResultLastXmin = RecentXmin;
4177 341476 : }
4178 :
4179 : /*
4180 : * Update boundaries in GlobalVis{Shared,Catalog, Data}Rels
4181 : * using ComputeXidHorizons().
4182 : */
4183 : static void
4184 106990 : GlobalVisUpdate(void)
4185 : {
4186 : ComputeXidHorizonsResult horizons;
4187 :
4188 : /* updates the horizons as a side-effect */
4189 106990 : ComputeXidHorizons(&horizons);
4190 106990 : }
4191 :
4192 : /*
4193 : * Return true if no snapshot still considers fxid to be running.
4194 : *
4195 : * The state passed needs to have been initialized for the relation fxid is
4196 : * from (NULL is also OK), otherwise the result may not be correct.
4197 : *
4198 : * See comment for GlobalVisState for details.
4199 : */
4200 : bool
4201 19486982 : GlobalVisTestIsRemovableFullXid(GlobalVisState *state,
4202 : FullTransactionId fxid)
4203 : {
4204 : /*
4205 : * If fxid is older than maybe_needed bound, it definitely is visible to
4206 : * everyone.
4207 : */
4208 19486982 : if (FullTransactionIdPrecedes(fxid, state->maybe_needed))
4209 5026614 : return true;
4210 :
4211 : /*
4212 : * If fxid is >= definitely_needed bound, it is very likely to still be
4213 : * considered running.
4214 : */
4215 14460368 : if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed))
4216 13558782 : return false;
4217 :
4218 : /*
4219 : * fxid is between maybe_needed and definitely_needed, i.e. there might or
4220 : * might not exist a snapshot considering fxid running. If it makes sense,
4221 : * update boundaries and recheck.
4222 : */
4223 901586 : if (GlobalVisTestShouldUpdate(state))
4224 : {
4225 106990 : GlobalVisUpdate();
4226 :
4227 : Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed));
4228 :
4229 106990 : return FullTransactionIdPrecedes(fxid, state->maybe_needed);
4230 : }
4231 : else
4232 794596 : return false;
4233 : }
4234 :
4235 : /*
4236 : * Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids.
4237 : *
4238 : * It is crucial that this only gets called for xids from a source that
4239 : * protects against xid wraparounds (e.g. from a table and thus protected by
4240 : * relfrozenxid).
4241 : */
4242 : bool
4243 19486302 : GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid)
4244 : {
4245 : FullTransactionId fxid;
4246 :
4247 : /*
4248 : * Convert 32 bit argument to FullTransactionId. We can do so safely
4249 : * because we know the xid has to, at the very least, be between
4250 : * [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a
4251 : * lock to determine either, we can just compare with
4252 : * state->definitely_needed, which was based on those value at the time
4253 : * the current snapshot was built.
4254 : */
4255 19486302 : fxid = FullXidRelativeTo(state->definitely_needed, xid);
4256 :
4257 19486302 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4258 : }
4259 :
4260 : /*
4261 : * Convenience wrapper around GlobalVisTestFor() and
4262 : * GlobalVisTestIsRemovableFullXid(), see their comments.
4263 : */
4264 : bool
4265 680 : GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid)
4266 : {
4267 : GlobalVisState *state;
4268 :
4269 680 : state = GlobalVisTestFor(rel);
4270 :
4271 680 : return GlobalVisTestIsRemovableFullXid(state, fxid);
4272 : }
4273 :
4274 : /*
4275 : * Convenience wrapper around GlobalVisTestFor() and
4276 : * GlobalVisTestIsRemovableXid(), see their comments.
4277 : */
4278 : bool
4279 16 : GlobalVisCheckRemovableXid(Relation rel, TransactionId xid)
4280 : {
4281 : GlobalVisState *state;
4282 :
4283 16 : state = GlobalVisTestFor(rel);
4284 :
4285 16 : return GlobalVisTestIsRemovableXid(state, xid);
4286 : }
4287 :
4288 : /*
4289 : * Convert a 32 bit transaction id into 64 bit transaction id, by assuming it
4290 : * is within MaxTransactionId / 2 of XidFromFullTransactionId(rel).
4291 : *
4292 : * Be very careful about when to use this function. It can only safely be used
4293 : * when there is a guarantee that xid is within MaxTransactionId / 2 xids of
4294 : * rel. That e.g. can be guaranteed if the caller assures a snapshot is
4295 : * held by the backend and xid is from a table (where vacuum/freezing ensures
4296 : * the xid has to be within that range), or if xid is from the procarray and
4297 : * prevents xid wraparound that way.
4298 : */
4299 : static inline FullTransactionId
4300 23660066 : FullXidRelativeTo(FullTransactionId rel, TransactionId xid)
4301 : {
4302 23660066 : TransactionId rel_xid = XidFromFullTransactionId(rel);
4303 :
4304 : Assert(TransactionIdIsValid(xid));
4305 : Assert(TransactionIdIsValid(rel_xid));
4306 :
4307 : /* not guaranteed to find issues, but likely to catch mistakes */
4308 : AssertTransactionIdInAllowableRange(xid);
4309 :
4310 47320132 : return FullTransactionIdFromU64(U64FromFullTransactionId(rel)
4311 23660066 : + (int32) (xid - rel_xid));
4312 : }
4313 :
4314 :
4315 : /* ----------------------------------------------
4316 : * KnownAssignedTransactionIds sub-module
4317 : * ----------------------------------------------
4318 : */
4319 :
4320 : /*
4321 : * In Hot Standby mode, we maintain a list of transactions that are (or were)
4322 : * running on the primary at the current point in WAL. These XIDs must be
4323 : * treated as running by standby transactions, even though they are not in
4324 : * the standby server's PGPROC array.
4325 : *
4326 : * We record all XIDs that we know have been assigned. That includes all the
4327 : * XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have
4328 : * been assigned. We can deduce the existence of unobserved XIDs because we
4329 : * know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids
4330 : * list expands as new XIDs are observed or inferred, and contracts when
4331 : * transaction completion records arrive.
4332 : *
4333 : * During hot standby we do not fret too much about the distinction between
4334 : * top-level XIDs and subtransaction XIDs. We store both together in the
4335 : * KnownAssignedXids list. In backends, this is copied into snapshots in
4336 : * GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot()
4337 : * doesn't care about the distinction either. Subtransaction XIDs are
4338 : * effectively treated as top-level XIDs and in the typical case pg_subtrans
4339 : * links are *not* maintained (which does not affect visibility).
4340 : *
4341 : * We have room in KnownAssignedXids and in snapshots to hold maxProcs *
4342 : * (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must
4343 : * report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at
4344 : * least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these
4345 : * records, we mark the subXIDs as children of the top XID in pg_subtrans,
4346 : * and then remove them from KnownAssignedXids. This prevents overflow of
4347 : * KnownAssignedXids and snapshots, at the cost that status checks for these
4348 : * subXIDs will take a slower path through TransactionIdIsInProgress().
4349 : * This means that KnownAssignedXids is not necessarily complete for subXIDs,
4350 : * though it should be complete for top-level XIDs; this is the same situation
4351 : * that holds with respect to the PGPROC entries in normal running.
4352 : *
4353 : * When we throw away subXIDs from KnownAssignedXids, we need to keep track of
4354 : * that, similarly to tracking overflow of a PGPROC's subxids array. We do
4355 : * that by remembering the lastOverflowedXid, ie the last thrown-away subXID.
4356 : * As long as that is within the range of interesting XIDs, we have to assume
4357 : * that subXIDs are missing from snapshots. (Note that subXID overflow occurs
4358 : * on primary when 65th subXID arrives, whereas on standby it occurs when 64th
4359 : * subXID arrives - that is not an error.)
4360 : *
4361 : * Should a backend on primary somehow disappear before it can write an abort
4362 : * record, then we just leave those XIDs in KnownAssignedXids. They actually
4363 : * aborted but we think they were running; the distinction is irrelevant
4364 : * because either way any changes done by the transaction are not visible to
4365 : * backends in the standby. We prune KnownAssignedXids when
4366 : * XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the
4367 : * array due to such dead XIDs.
4368 : */
4369 :
4370 : /*
4371 : * RecordKnownAssignedTransactionIds
4372 : * Record the given XID in KnownAssignedXids, as well as any preceding
4373 : * unobserved XIDs.
4374 : *
4375 : * RecordKnownAssignedTransactionIds() should be run for *every* WAL record
4376 : * associated with a transaction. Must be called for each record after we
4377 : * have executed StartupCLOG() et al, since we must ExtendCLOG() etc..
4378 : *
4379 : * Called during recovery in analogy with and in place of GetNewTransactionId()
4380 : */
4381 : void
4382 5035540 : RecordKnownAssignedTransactionIds(TransactionId xid)
4383 : {
4384 : Assert(standbyState >= STANDBY_INITIALIZED);
4385 : Assert(TransactionIdIsValid(xid));
4386 : Assert(TransactionIdIsValid(latestObservedXid));
4387 :
4388 5035540 : elog(DEBUG4, "record known xact %u latestObservedXid %u",
4389 : xid, latestObservedXid);
4390 :
4391 : /*
4392 : * When a newly observed xid arrives, it is frequently the case that it is
4393 : * *not* the next xid in sequence. When this occurs, we must treat the
4394 : * intervening xids as running also.
4395 : */
4396 5035540 : if (TransactionIdFollows(xid, latestObservedXid))
4397 : {
4398 : TransactionId next_expected_xid;
4399 :
4400 : /*
4401 : * Extend subtrans like we do in GetNewTransactionId() during normal
4402 : * operation using individual extend steps. Note that we do not need
4403 : * to extend clog since its extensions are WAL logged.
4404 : *
4405 : * This part has to be done regardless of standbyState since we
4406 : * immediately start assigning subtransactions to their toplevel
4407 : * transactions.
4408 : */
4409 45838 : next_expected_xid = latestObservedXid;
4410 93328 : while (TransactionIdPrecedes(next_expected_xid, xid))
4411 : {
4412 47490 : TransactionIdAdvance(next_expected_xid);
4413 47490 : ExtendSUBTRANS(next_expected_xid);
4414 : }
4415 : Assert(next_expected_xid == xid);
4416 :
4417 : /*
4418 : * If the KnownAssignedXids machinery isn't up yet, there's nothing
4419 : * more to do since we don't track assigned xids yet.
4420 : */
4421 45838 : if (standbyState <= STANDBY_INITIALIZED)
4422 : {
4423 0 : latestObservedXid = xid;
4424 0 : return;
4425 : }
4426 :
4427 : /*
4428 : * Add (latestObservedXid, xid] onto the KnownAssignedXids array.
4429 : */
4430 45838 : next_expected_xid = latestObservedXid;
4431 45838 : TransactionIdAdvance(next_expected_xid);
4432 45838 : KnownAssignedXidsAdd(next_expected_xid, xid, false);
4433 :
4434 : /*
4435 : * Now we can advance latestObservedXid
4436 : */
4437 45838 : latestObservedXid = xid;
4438 :
4439 : /* TransamVariables->nextXid must be beyond any observed xid */
4440 45838 : AdvanceNextFullTransactionIdPastXid(latestObservedXid);
4441 : }
4442 : }
4443 :
4444 : /*
4445 : * ExpireTreeKnownAssignedTransactionIds
4446 : * Remove the given XIDs from KnownAssignedXids.
4447 : *
4448 : * Called during recovery in analogy with and in place of ProcArrayEndTransaction()
4449 : */
4450 : void
4451 44422 : ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids,
4452 : TransactionId *subxids, TransactionId max_xid)
4453 : {
4454 : Assert(standbyState >= STANDBY_INITIALIZED);
4455 :
4456 : /*
4457 : * Uses same locking as transaction commit
4458 : */
4459 44422 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4460 :
4461 44422 : KnownAssignedXidsRemoveTree(xid, nsubxids, subxids);
4462 :
4463 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4464 44422 : MaintainLatestCompletedXidRecovery(max_xid);
4465 :
4466 : /* ... and xactCompletionCount */
4467 44422 : TransamVariables->xactCompletionCount++;
4468 :
4469 44422 : LWLockRelease(ProcArrayLock);
4470 44422 : }
4471 :
4472 : /*
4473 : * ExpireAllKnownAssignedTransactionIds
4474 : * Remove all entries in KnownAssignedXids and reset lastOverflowedXid.
4475 : */
4476 : void
4477 230 : ExpireAllKnownAssignedTransactionIds(void)
4478 : {
4479 : FullTransactionId latestXid;
4480 :
4481 230 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4482 230 : KnownAssignedXidsRemovePreceding(InvalidTransactionId);
4483 :
4484 : /* Reset latestCompletedXid to nextXid - 1 */
4485 : Assert(FullTransactionIdIsValid(TransamVariables->nextXid));
4486 230 : latestXid = TransamVariables->nextXid;
4487 230 : FullTransactionIdRetreat(&latestXid);
4488 230 : TransamVariables->latestCompletedXid = latestXid;
4489 :
4490 : /*
4491 : * Any transactions that were in-progress were effectively aborted, so
4492 : * advance xactCompletionCount.
4493 : */
4494 230 : TransamVariables->xactCompletionCount++;
4495 :
4496 : /*
4497 : * Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after
4498 : * the call of this function. But do this for unification with what
4499 : * ExpireOldKnownAssignedTransactionIds() do.
4500 : */
4501 230 : procArray->lastOverflowedXid = InvalidTransactionId;
4502 230 : LWLockRelease(ProcArrayLock);
4503 230 : }
4504 :
4505 : /*
4506 : * ExpireOldKnownAssignedTransactionIds
4507 : * Remove KnownAssignedXids entries preceding the given XID and
4508 : * potentially reset lastOverflowedXid.
4509 : */
4510 : void
4511 1684 : ExpireOldKnownAssignedTransactionIds(TransactionId xid)
4512 : {
4513 : TransactionId latestXid;
4514 :
4515 1684 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4516 :
4517 : /* As in ProcArrayEndTransaction, advance latestCompletedXid */
4518 1684 : latestXid = xid;
4519 1684 : TransactionIdRetreat(latestXid);
4520 1684 : MaintainLatestCompletedXidRecovery(latestXid);
4521 :
4522 : /* ... and xactCompletionCount */
4523 1684 : TransamVariables->xactCompletionCount++;
4524 :
4525 : /*
4526 : * Reset lastOverflowedXid if we know all transactions that have been
4527 : * possibly running are being gone. Not doing so could cause an incorrect
4528 : * lastOverflowedXid value, which makes extra snapshots be marked as
4529 : * suboverflowed.
4530 : */
4531 1684 : if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid))
4532 1672 : procArray->lastOverflowedXid = InvalidTransactionId;
4533 1684 : KnownAssignedXidsRemovePreceding(xid);
4534 1684 : LWLockRelease(ProcArrayLock);
4535 1684 : }
4536 :
4537 : /*
4538 : * KnownAssignedTransactionIdsIdleMaintenance
4539 : * Opportunistically do maintenance work when the startup process
4540 : * is about to go idle.
4541 : */
4542 : void
4543 12848 : KnownAssignedTransactionIdsIdleMaintenance(void)
4544 : {
4545 12848 : KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false);
4546 12848 : }
4547 :
4548 :
4549 : /*
4550 : * Private module functions to manipulate KnownAssignedXids
4551 : *
4552 : * There are 5 main uses of the KnownAssignedXids data structure:
4553 : *
4554 : * * backends taking snapshots - all valid XIDs need to be copied out
4555 : * * backends seeking to determine presence of a specific XID
4556 : * * startup process adding new known-assigned XIDs
4557 : * * startup process removing specific XIDs as transactions end
4558 : * * startup process pruning array when special WAL records arrive
4559 : *
4560 : * This data structure is known to be a hot spot during Hot Standby, so we
4561 : * go to some lengths to make these operations as efficient and as concurrent
4562 : * as possible.
4563 : *
4564 : * The XIDs are stored in an array in sorted order --- TransactionIdPrecedes
4565 : * order, to be exact --- to allow binary search for specific XIDs. Note:
4566 : * in general TransactionIdPrecedes would not provide a total order, but
4567 : * we know that the entries present at any instant should not extend across
4568 : * a large enough fraction of XID space to wrap around (the primary would
4569 : * shut down for fear of XID wrap long before that happens). So it's OK to
4570 : * use TransactionIdPrecedes as a binary-search comparator.
4571 : *
4572 : * It's cheap to maintain the sortedness during insertions, since new known
4573 : * XIDs are always reported in XID order; we just append them at the right.
4574 : *
4575 : * To keep individual deletions cheap, we need to allow gaps in the array.
4576 : * This is implemented by marking array elements as valid or invalid using
4577 : * the parallel boolean array KnownAssignedXidsValid[]. A deletion is done
4578 : * by setting KnownAssignedXidsValid[i] to false, *without* clearing the
4579 : * XID entry itself. This preserves the property that the XID entries are
4580 : * sorted, so we can do binary searches easily. Periodically we compress
4581 : * out the unused entries; that's much cheaper than having to compress the
4582 : * array immediately on every deletion.
4583 : *
4584 : * The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[]
4585 : * are those with indexes tail <= i < head; items outside this subscript range
4586 : * have unspecified contents. When head reaches the end of the array, we
4587 : * force compression of unused entries rather than wrapping around, since
4588 : * allowing wraparound would greatly complicate the search logic. We maintain
4589 : * an explicit tail pointer so that pruning of old XIDs can be done without
4590 : * immediately moving the array contents. In most cases only a small fraction
4591 : * of the array contains valid entries at any instant.
4592 : *
4593 : * Although only the startup process can ever change the KnownAssignedXids
4594 : * data structure, we still need interlocking so that standby backends will
4595 : * not observe invalid intermediate states. The convention is that backends
4596 : * must hold shared ProcArrayLock to examine the array. To remove XIDs from
4597 : * the array, the startup process must hold ProcArrayLock exclusively, for
4598 : * the usual transactional reasons (compare commit/abort of a transaction
4599 : * during normal running). Compressing unused entries out of the array
4600 : * likewise requires exclusive lock. To add XIDs to the array, we just insert
4601 : * them into slots to the right of the head pointer and then advance the head
4602 : * pointer. This doesn't require any lock at all, but on machines with weak
4603 : * memory ordering, we need to be careful that other processors see the array
4604 : * element changes before they see the head pointer change. We handle this by
4605 : * using memory barriers when reading or writing the head/tail pointers (unless
4606 : * the caller holds ProcArrayLock exclusively).
4607 : *
4608 : * Algorithmic analysis:
4609 : *
4610 : * If we have a maximum of M slots, with N XIDs currently spread across
4611 : * S elements then we have N <= S <= M always.
4612 : *
4613 : * * Adding a new XID is O(1) and needs no lock (unless compression must
4614 : * happen)
4615 : * * Compressing the array is O(S) and requires exclusive lock
4616 : * * Removing an XID is O(logS) and requires exclusive lock
4617 : * * Taking a snapshot is O(S) and requires shared lock
4618 : * * Checking for an XID is O(logS) and requires shared lock
4619 : *
4620 : * In comparison, using a hash table for KnownAssignedXids would mean that
4621 : * taking snapshots would be O(M). If we can maintain S << M then the
4622 : * sorted array technique will deliver significantly faster snapshots.
4623 : * If we try to keep S too small then we will spend too much time compressing,
4624 : * so there is an optimal point for any workload mix. We use a heuristic to
4625 : * decide when to compress the array, though trimming also helps reduce
4626 : * frequency of compressing. The heuristic requires us to track the number of
4627 : * currently valid XIDs in the array (N). Except in special cases, we'll
4628 : * compress when S >= 2N. Bounding S at 2N in turn bounds the time for
4629 : * taking a snapshot to be O(N), which it would have to be anyway.
4630 : */
4631 :
4632 :
4633 : /*
4634 : * Compress KnownAssignedXids by shifting valid data down to the start of the
4635 : * array, removing any gaps.
4636 : *
4637 : * A compression step is forced if "reason" is KAX_NO_SPACE, otherwise
4638 : * we do it only if a heuristic indicates it's a good time to do it.
4639 : *
4640 : * Compression requires holding ProcArrayLock in exclusive mode.
4641 : * Caller must pass haveLock = true if it already holds the lock.
4642 : */
4643 : static void
4644 58996 : KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock)
4645 : {
4646 58996 : ProcArrayStruct *pArray = procArray;
4647 : int head,
4648 : tail,
4649 : nelements;
4650 : int compress_index;
4651 : int i;
4652 :
4653 : /* Counters for compression heuristics */
4654 : static unsigned int transactionEndsCounter;
4655 : static TimestampTz lastCompressTs;
4656 :
4657 : /* Tuning constants */
4658 : #define KAX_COMPRESS_FREQUENCY 128 /* in transactions */
4659 : #define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */
4660 :
4661 : /*
4662 : * Since only the startup process modifies the head/tail pointers, we
4663 : * don't need a lock to read them here.
4664 : */
4665 58996 : head = pArray->headKnownAssignedXids;
4666 58996 : tail = pArray->tailKnownAssignedXids;
4667 58996 : nelements = head - tail;
4668 :
4669 : /*
4670 : * If we can choose whether to compress, use a heuristic to avoid
4671 : * compressing too often or not often enough. "Compress" here simply
4672 : * means moving the values to the beginning of the array, so it is not as
4673 : * complex or costly as typical data compression algorithms.
4674 : */
4675 58996 : if (nelements == pArray->numKnownAssignedXids)
4676 : {
4677 : /*
4678 : * When there are no gaps between head and tail, don't bother to
4679 : * compress, except in the KAX_NO_SPACE case where we must compress to
4680 : * create some space after the head.
4681 : */
4682 25994 : if (reason != KAX_NO_SPACE)
4683 25994 : return;
4684 : }
4685 33002 : else if (reason == KAX_TRANSACTION_END)
4686 : {
4687 : /*
4688 : * Consider compressing only once every so many commits. Frequency
4689 : * determined by benchmarks.
4690 : */
4691 29198 : if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0)
4692 28952 : return;
4693 :
4694 : /*
4695 : * Furthermore, compress only if the used part of the array is less
4696 : * than 50% full (see comments above).
4697 : */
4698 246 : if (nelements < 2 * pArray->numKnownAssignedXids)
4699 22 : return;
4700 : }
4701 3804 : else if (reason == KAX_STARTUP_PROCESS_IDLE)
4702 : {
4703 : /*
4704 : * We're about to go idle for lack of new WAL, so we might as well
4705 : * compress. But not too often, to avoid ProcArray lock contention
4706 : * with readers.
4707 : */
4708 3574 : if (lastCompressTs != 0)
4709 : {
4710 : TimestampTz compress_after;
4711 :
4712 3572 : compress_after = TimestampTzPlusMilliseconds(lastCompressTs,
4713 : KAX_COMPRESS_IDLE_INTERVAL);
4714 3572 : if (GetCurrentTimestamp() < compress_after)
4715 3378 : return;
4716 : }
4717 : }
4718 :
4719 : /* Need to compress, so get the lock if we don't have it. */
4720 650 : if (!haveLock)
4721 196 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
4722 :
4723 : /*
4724 : * We compress the array by reading the valid values from tail to head,
4725 : * re-aligning data to 0th element.
4726 : */
4727 650 : compress_index = 0;
4728 20886 : for (i = tail; i < head; i++)
4729 : {
4730 20236 : if (KnownAssignedXidsValid[i])
4731 : {
4732 2804 : KnownAssignedXids[compress_index] = KnownAssignedXids[i];
4733 2804 : KnownAssignedXidsValid[compress_index] = true;
4734 2804 : compress_index++;
4735 : }
4736 : }
4737 : Assert(compress_index == pArray->numKnownAssignedXids);
4738 :
4739 650 : pArray->tailKnownAssignedXids = 0;
4740 650 : pArray->headKnownAssignedXids = compress_index;
4741 :
4742 650 : if (!haveLock)
4743 196 : LWLockRelease(ProcArrayLock);
4744 :
4745 : /* Update timestamp for maintenance. No need to hold lock for this. */
4746 650 : lastCompressTs = GetCurrentTimestamp();
4747 : }
4748 :
4749 : /*
4750 : * Add xids into KnownAssignedXids at the head of the array.
4751 : *
4752 : * xids from from_xid to to_xid, inclusive, are added to the array.
4753 : *
4754 : * If exclusive_lock is true then caller already holds ProcArrayLock in
4755 : * exclusive mode, so we need no extra locking here. Else caller holds no
4756 : * lock, so we need to be sure we maintain sufficient interlocks against
4757 : * concurrent readers. (Only the startup process ever calls this, so no need
4758 : * to worry about concurrent writers.)
4759 : */
4760 : static void
4761 45846 : KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid,
4762 : bool exclusive_lock)
4763 : {
4764 45846 : ProcArrayStruct *pArray = procArray;
4765 : TransactionId next_xid;
4766 : int head,
4767 : tail;
4768 : int nxids;
4769 : int i;
4770 :
4771 : Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid));
4772 :
4773 : /*
4774 : * Calculate how many array slots we'll need. Normally this is cheap; in
4775 : * the unusual case where the XIDs cross the wrap point, we do it the hard
4776 : * way.
4777 : */
4778 45846 : if (to_xid >= from_xid)
4779 45846 : nxids = to_xid - from_xid + 1;
4780 : else
4781 : {
4782 0 : nxids = 1;
4783 0 : next_xid = from_xid;
4784 0 : while (TransactionIdPrecedes(next_xid, to_xid))
4785 : {
4786 0 : nxids++;
4787 0 : TransactionIdAdvance(next_xid);
4788 : }
4789 : }
4790 :
4791 : /*
4792 : * Since only the startup process modifies the head/tail pointers, we
4793 : * don't need a lock to read them here.
4794 : */
4795 45846 : head = pArray->headKnownAssignedXids;
4796 45846 : tail = pArray->tailKnownAssignedXids;
4797 :
4798 : Assert(head >= 0 && head <= pArray->maxKnownAssignedXids);
4799 : Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids);
4800 :
4801 : /*
4802 : * Verify that insertions occur in TransactionId sequence. Note that even
4803 : * if the last existing element is marked invalid, it must still have a
4804 : * correctly sequenced XID value.
4805 : */
4806 77600 : if (head > tail &&
4807 31754 : TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid))
4808 : {
4809 0 : KnownAssignedXidsDisplay(LOG);
4810 0 : elog(ERROR, "out-of-order XID insertion in KnownAssignedXids");
4811 : }
4812 :
4813 : /*
4814 : * If our xids won't fit in the remaining space, compress out free space
4815 : */
4816 45846 : if (head + nxids > pArray->maxKnownAssignedXids)
4817 : {
4818 0 : KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock);
4819 :
4820 0 : head = pArray->headKnownAssignedXids;
4821 : /* note: we no longer care about the tail pointer */
4822 :
4823 : /*
4824 : * If it still won't fit then we're out of memory
4825 : */
4826 0 : if (head + nxids > pArray->maxKnownAssignedXids)
4827 0 : elog(ERROR, "too many KnownAssignedXids");
4828 : }
4829 :
4830 : /* Now we can insert the xids into the space starting at head */
4831 45846 : next_xid = from_xid;
4832 93344 : for (i = 0; i < nxids; i++)
4833 : {
4834 47498 : KnownAssignedXids[head] = next_xid;
4835 47498 : KnownAssignedXidsValid[head] = true;
4836 47498 : TransactionIdAdvance(next_xid);
4837 47498 : head++;
4838 : }
4839 :
4840 : /* Adjust count of number of valid entries */
4841 45846 : pArray->numKnownAssignedXids += nxids;
4842 :
4843 : /*
4844 : * Now update the head pointer. We use a write barrier to ensure that
4845 : * other processors see the above array updates before they see the head
4846 : * pointer change. The barrier isn't required if we're holding
4847 : * ProcArrayLock exclusively.
4848 : */
4849 45846 : if (!exclusive_lock)
4850 45838 : pg_write_barrier();
4851 :
4852 45846 : pArray->headKnownAssignedXids = head;
4853 45846 : }
4854 :
4855 : /*
4856 : * KnownAssignedXidsSearch
4857 : *
4858 : * Searches KnownAssignedXids for a specific xid and optionally removes it.
4859 : * Returns true if it was found, false if not.
4860 : *
4861 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4862 : * Exclusive lock must be held for remove = true.
4863 : */
4864 : static bool
4865 49766 : KnownAssignedXidsSearch(TransactionId xid, bool remove)
4866 : {
4867 49766 : ProcArrayStruct *pArray = procArray;
4868 : int first,
4869 : last;
4870 : int head;
4871 : int tail;
4872 49766 : int result_index = -1;
4873 :
4874 49766 : tail = pArray->tailKnownAssignedXids;
4875 49766 : head = pArray->headKnownAssignedXids;
4876 :
4877 : /*
4878 : * Only the startup process removes entries, so we don't need the read
4879 : * barrier in that case.
4880 : */
4881 49766 : if (!remove)
4882 2 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
4883 :
4884 : /*
4885 : * Standard binary search. Note we can ignore the KnownAssignedXidsValid
4886 : * array here, since even invalid entries will contain sorted XIDs.
4887 : */
4888 49766 : first = tail;
4889 49766 : last = head - 1;
4890 175464 : while (first <= last)
4891 : {
4892 : int mid_index;
4893 : TransactionId mid_xid;
4894 :
4895 173076 : mid_index = (first + last) / 2;
4896 173076 : mid_xid = KnownAssignedXids[mid_index];
4897 :
4898 173076 : if (xid == mid_xid)
4899 : {
4900 47378 : result_index = mid_index;
4901 47378 : break;
4902 : }
4903 125698 : else if (TransactionIdPrecedes(xid, mid_xid))
4904 26856 : last = mid_index - 1;
4905 : else
4906 98842 : first = mid_index + 1;
4907 : }
4908 :
4909 49766 : if (result_index < 0)
4910 2388 : return false; /* not in array */
4911 :
4912 47378 : if (!KnownAssignedXidsValid[result_index])
4913 52 : return false; /* in array, but invalid */
4914 :
4915 47326 : if (remove)
4916 : {
4917 47326 : KnownAssignedXidsValid[result_index] = false;
4918 :
4919 47326 : pArray->numKnownAssignedXids--;
4920 : Assert(pArray->numKnownAssignedXids >= 0);
4921 :
4922 : /*
4923 : * If we're removing the tail element then advance tail pointer over
4924 : * any invalid elements. This will speed future searches.
4925 : */
4926 47326 : if (result_index == tail)
4927 : {
4928 16378 : tail++;
4929 29894 : while (tail < head && !KnownAssignedXidsValid[tail])
4930 13516 : tail++;
4931 16378 : if (tail >= head)
4932 : {
4933 : /* Array is empty, so we can reset both pointers */
4934 14072 : pArray->headKnownAssignedXids = 0;
4935 14072 : pArray->tailKnownAssignedXids = 0;
4936 : }
4937 : else
4938 : {
4939 2306 : pArray->tailKnownAssignedXids = tail;
4940 : }
4941 : }
4942 : }
4943 :
4944 47326 : return true;
4945 : }
4946 :
4947 : /*
4948 : * Is the specified XID present in KnownAssignedXids[]?
4949 : *
4950 : * Caller must hold ProcArrayLock in shared or exclusive mode.
4951 : */
4952 : static bool
4953 2 : KnownAssignedXidExists(TransactionId xid)
4954 : {
4955 : Assert(TransactionIdIsValid(xid));
4956 :
4957 2 : return KnownAssignedXidsSearch(xid, false);
4958 : }
4959 :
4960 : /*
4961 : * Remove the specified XID from KnownAssignedXids[].
4962 : *
4963 : * Caller must hold ProcArrayLock in exclusive mode.
4964 : */
4965 : static void
4966 49764 : KnownAssignedXidsRemove(TransactionId xid)
4967 : {
4968 : Assert(TransactionIdIsValid(xid));
4969 :
4970 49764 : elog(DEBUG4, "remove KnownAssignedXid %u", xid);
4971 :
4972 : /*
4973 : * Note: we cannot consider it an error to remove an XID that's not
4974 : * present. We intentionally remove subxact IDs while processing
4975 : * XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be
4976 : * removed again when the top-level xact commits or aborts.
4977 : *
4978 : * It might be possible to track such XIDs to distinguish this case from
4979 : * actual errors, but it would be complicated and probably not worth it.
4980 : * So, just ignore the search result.
4981 : */
4982 49764 : (void) KnownAssignedXidsSearch(xid, true);
4983 49764 : }
4984 :
4985 : /*
4986 : * KnownAssignedXidsRemoveTree
4987 : * Remove xid (if it's not InvalidTransactionId) and all the subxids.
4988 : *
4989 : * Caller must hold ProcArrayLock in exclusive mode.
4990 : */
4991 : static void
4992 44464 : KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids,
4993 : TransactionId *subxids)
4994 : {
4995 : int i;
4996 :
4997 44464 : if (TransactionIdIsValid(xid))
4998 44422 : KnownAssignedXidsRemove(xid);
4999 :
5000 49806 : for (i = 0; i < nsubxids; i++)
5001 5342 : KnownAssignedXidsRemove(subxids[i]);
5002 :
5003 : /* Opportunistically compress the array */
5004 44464 : KnownAssignedXidsCompress(KAX_TRANSACTION_END, true);
5005 44464 : }
5006 :
5007 : /*
5008 : * Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid
5009 : * then clear the whole table.
5010 : *
5011 : * Caller must hold ProcArrayLock in exclusive mode.
5012 : */
5013 : static void
5014 1914 : KnownAssignedXidsRemovePreceding(TransactionId removeXid)
5015 : {
5016 1914 : ProcArrayStruct *pArray = procArray;
5017 1914 : int count = 0;
5018 : int head,
5019 : tail,
5020 : i;
5021 :
5022 1914 : if (!TransactionIdIsValid(removeXid))
5023 : {
5024 230 : elog(DEBUG4, "removing all KnownAssignedXids");
5025 230 : pArray->numKnownAssignedXids = 0;
5026 230 : pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0;
5027 230 : return;
5028 : }
5029 :
5030 1684 : elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid);
5031 :
5032 : /*
5033 : * Mark entries invalid starting at the tail. Since array is sorted, we
5034 : * can stop as soon as we reach an entry >= removeXid.
5035 : */
5036 1684 : tail = pArray->tailKnownAssignedXids;
5037 1684 : head = pArray->headKnownAssignedXids;
5038 :
5039 1684 : for (i = tail; i < head; i++)
5040 : {
5041 450 : if (KnownAssignedXidsValid[i])
5042 : {
5043 450 : TransactionId knownXid = KnownAssignedXids[i];
5044 :
5045 450 : if (TransactionIdFollowsOrEquals(knownXid, removeXid))
5046 450 : break;
5047 :
5048 0 : if (!StandbyTransactionIdIsPrepared(knownXid))
5049 : {
5050 0 : KnownAssignedXidsValid[i] = false;
5051 0 : count++;
5052 : }
5053 : }
5054 : }
5055 :
5056 1684 : pArray->numKnownAssignedXids -= count;
5057 : Assert(pArray->numKnownAssignedXids >= 0);
5058 :
5059 : /*
5060 : * Advance the tail pointer if we've marked the tail item invalid.
5061 : */
5062 1684 : for (i = tail; i < head; i++)
5063 : {
5064 450 : if (KnownAssignedXidsValid[i])
5065 450 : break;
5066 : }
5067 1684 : if (i >= head)
5068 : {
5069 : /* Array is empty, so we can reset both pointers */
5070 1234 : pArray->headKnownAssignedXids = 0;
5071 1234 : pArray->tailKnownAssignedXids = 0;
5072 : }
5073 : else
5074 : {
5075 450 : pArray->tailKnownAssignedXids = i;
5076 : }
5077 :
5078 : /* Opportunistically compress the array */
5079 1684 : KnownAssignedXidsCompress(KAX_PRUNE, true);
5080 : }
5081 :
5082 : /*
5083 : * KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids.
5084 : * We filter out anything >= xmax.
5085 : *
5086 : * Returns the number of XIDs stored into xarray[]. Caller is responsible
5087 : * that array is large enough.
5088 : *
5089 : * Caller must hold ProcArrayLock in (at least) shared mode.
5090 : */
5091 : static int
5092 0 : KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax)
5093 : {
5094 0 : TransactionId xtmp = InvalidTransactionId;
5095 :
5096 0 : return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax);
5097 : }
5098 :
5099 : /*
5100 : * KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus
5101 : * we reduce *xmin to the lowest xid value seen if not already lower.
5102 : *
5103 : * Caller must hold ProcArrayLock in (at least) shared mode.
5104 : */
5105 : static int
5106 2626 : KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin,
5107 : TransactionId xmax)
5108 : {
5109 2626 : int count = 0;
5110 : int head,
5111 : tail;
5112 : int i;
5113 :
5114 : /*
5115 : * Fetch head just once, since it may change while we loop. We can stop
5116 : * once we reach the initially seen head, since we are certain that an xid
5117 : * cannot enter and then leave the array while we hold ProcArrayLock. We
5118 : * might miss newly-added xids, but they should be >= xmax so irrelevant
5119 : * anyway.
5120 : */
5121 2626 : tail = procArray->tailKnownAssignedXids;
5122 2626 : head = procArray->headKnownAssignedXids;
5123 :
5124 2626 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5125 :
5126 2772 : for (i = tail; i < head; i++)
5127 : {
5128 : /* Skip any gaps in the array */
5129 404 : if (KnownAssignedXidsValid[i])
5130 : {
5131 296 : TransactionId knownXid = KnownAssignedXids[i];
5132 :
5133 : /*
5134 : * Update xmin if required. Only the first XID need be checked,
5135 : * since the array is sorted.
5136 : */
5137 582 : if (count == 0 &&
5138 286 : TransactionIdPrecedes(knownXid, *xmin))
5139 34 : *xmin = knownXid;
5140 :
5141 : /*
5142 : * Filter out anything >= xmax, again relying on sorted property
5143 : * of array.
5144 : */
5145 592 : if (TransactionIdIsValid(xmax) &&
5146 296 : TransactionIdFollowsOrEquals(knownXid, xmax))
5147 258 : break;
5148 :
5149 : /* Add knownXid into output array */
5150 38 : xarray[count++] = knownXid;
5151 : }
5152 : }
5153 :
5154 2626 : return count;
5155 : }
5156 :
5157 : /*
5158 : * Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId
5159 : * if nothing there.
5160 : */
5161 : static TransactionId
5162 684 : KnownAssignedXidsGetOldestXmin(void)
5163 : {
5164 : int head,
5165 : tail;
5166 : int i;
5167 :
5168 : /*
5169 : * Fetch head just once, since it may change while we loop.
5170 : */
5171 684 : tail = procArray->tailKnownAssignedXids;
5172 684 : head = procArray->headKnownAssignedXids;
5173 :
5174 684 : pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */
5175 :
5176 684 : for (i = tail; i < head; i++)
5177 : {
5178 : /* Skip any gaps in the array */
5179 280 : if (KnownAssignedXidsValid[i])
5180 280 : return KnownAssignedXids[i];
5181 : }
5182 :
5183 404 : return InvalidTransactionId;
5184 : }
5185 :
5186 : /*
5187 : * Display KnownAssignedXids to provide debug trail
5188 : *
5189 : * Currently this is only called within startup process, so we need no
5190 : * special locking.
5191 : *
5192 : * Note this is pretty expensive, and much of the expense will be incurred
5193 : * even if the elog message will get discarded. It's not currently called
5194 : * in any performance-critical places, however, so no need to be tenser.
5195 : */
5196 : static void
5197 238 : KnownAssignedXidsDisplay(int trace_level)
5198 : {
5199 238 : ProcArrayStruct *pArray = procArray;
5200 : StringInfoData buf;
5201 : int head,
5202 : tail,
5203 : i;
5204 238 : int nxids = 0;
5205 :
5206 238 : tail = pArray->tailKnownAssignedXids;
5207 238 : head = pArray->headKnownAssignedXids;
5208 :
5209 238 : initStringInfo(&buf);
5210 :
5211 254 : for (i = tail; i < head; i++)
5212 : {
5213 16 : if (KnownAssignedXidsValid[i])
5214 : {
5215 16 : nxids++;
5216 16 : appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]);
5217 : }
5218 : }
5219 :
5220 238 : elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s",
5221 : nxids,
5222 : pArray->numKnownAssignedXids,
5223 : pArray->tailKnownAssignedXids,
5224 : pArray->headKnownAssignedXids,
5225 : buf.data);
5226 :
5227 238 : pfree(buf.data);
5228 238 : }
5229 :
5230 : /*
5231 : * KnownAssignedXidsReset
5232 : * Resets KnownAssignedXids to be empty
5233 : */
5234 : static void
5235 0 : KnownAssignedXidsReset(void)
5236 : {
5237 0 : ProcArrayStruct *pArray = procArray;
5238 :
5239 0 : LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE);
5240 :
5241 0 : pArray->numKnownAssignedXids = 0;
5242 0 : pArray->tailKnownAssignedXids = 0;
5243 0 : pArray->headKnownAssignedXids = 0;
5244 :
5245 0 : LWLockRelease(ProcArrayLock);
5246 0 : }
|
